diff --git a/benchmarks/harbor/README.md b/benchmarks/harbor/README.md
index 78bbf4a..63ea2cf 100644
--- a/benchmarks/harbor/README.md
+++ b/benchmarks/harbor/README.md
@@ -285,6 +285,16 @@ restore a deleted container or missing task outputs from an old trial.
 Harbor's ordinary `--max-retries` setting reruns the entire trial, including
 the agent, and is independent of these verifier attempts.
 
+For a separately authorized infrastructure-recovery run, set the verifier
+kwarg `retry_dependency_errors=true` explicitly. This also retries recognized
+uv network download failures before pytest starts, using the same container,
+agent output, native time limit, and maximum of three total attempts. A CTRF
+report, pytest execution output, or a successful reward prevents this recovery;
+valid scored failures stop immediately. Exhausted dependency failures raise
+`VerifierDependencyError`, retain every attempt's evidence, and leave no
+canonical reward. The default remains timeout-only; record the authorization
+and selected policy in the experiment plan.
+
 ## Test the adapter
 
 ```bash
diff --git a/benchmarks/harbor/README.zh-CN.md b/benchmarks/harbor/README.zh-CN.md
index 3a551e2..74c797a 100644
--- a/benchmarks/harbor/README.zh-CN.md
+++ b/benchmarks/harbor/README.zh-CN.md
@@ -231,6 +231,13 @@ Harbor 原生的 provider 凭证和已经配置的 base URL 也会转换为 nano
 新运行，无法恢复旧 trial 已删除的容器或缺失的任务产物。Harbor 原生的
 `--max-retries` 会重新执行整道题（包括 agent），与这里的验证尝试次数相互独立。
 
+对单独授权的基础设施恢复运行，可显式设置验证器参数
+`retry_dependency_errors=true`。这也会重试 pytest 启动前识别到的 uv 网络下载
+错误，复用同一容器和 agent 产物，保留原生单次时限，总尝试次数仍最多为 3。
+只要已有 CTRF 报告、pytest 执行输出或成功评分，就不会进行这种恢复；有效的失败
+评分也会立即停止。依赖错误耗尽尝试后抛出 `VerifierDependencyError`，保留每次
+证据并清除当前评分。默认仍只重试超时；实验计划应记录授权和实际选用的策略。
+
 ## 测试 adapter
 
 ```bash
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/README.md b/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/README.md
new file mode 100644
index 0000000..db1b94a
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/README.md
@@ -0,0 +1,59 @@
+# Authorized retry of the two unscored Flash 1017 tasks
+
+**Completed:** sampler passed all nine official tests on the first verification;
+pipeline failed two of four cases on the second verification after its first
+attempt timed out during dependency preparation. No third attempt was started.
+See the [report](../../reports/deepseek-v41-flash-1017-recovery-20261018.md) and
+[audited results](../../results/deepseek-v41-flash-1017-recovery-20261018.json).
+
+The user requested another attempt on `adaptive-rejection-sampler` and
+`torch-pipeline-parallelism`, with at most three attempts. The initial
+[six-trial run](../../reports/deepseek-v41-flash-1017-rerun-20261018.md) left
+these two tasks unscored because verifier dependencies failed to download.
+The original containers were deleted and their artifact manifests contain no
+`/app` outputs, so this is a **new model trial per task**, not a regrade of the
+original submissions. All original results and evidence remain unchanged.
+
+The [registered plan](experiment.json) uses the standard Harbor entry point
+from the unchanged checkout `jobs/flash1017-recovery-source`, frozen at
+`35fe1a2d6a786619a2adace1cfa9d0acca08d533`. Agent Python source hashes exactly
+match the original agent. Only the adapter changes: its optional
+`retry_dependency_errors=true` verifier setting is enabled for this authorized
+recovery experiment. It remains disabled by default.
+
+Each task receives one fresh `deepseek/deepseek-flash` trial with the same
+prompt, pinned task revision, original image, resources, and budgets as before:
+100 replies, 65,536 output tokens per reply, a 720-second work budget, and
+separate 1800-second setup. No whole-trial retries are permitted.
+
+The verifier runs the original test script at most **three times total**,
+including its first attempt in this follow-up. Each attempt retains the task's
+original 900-second time limit. It may retry native verifier timeouts and
+recognized uv network download failures before pytest starts. Any official
+test result, including failure, stops the retries. A CTRF report, pytest
+execution output, or successful reward prevents download-error recovery.
+Unrelated errors are not retried. All attempts reuse the same fresh model
+output and container, preserve logs, and clear stale canonical rewards.
+Exhausted download errors remain unscored and raise `VerifierDependencyError`.
+Complete `/app` artifacts are collected for future inspection or verification.
+
+Both tasks receive fresh install-only preflight before model work. Their
+verifier dependency profiles remain explicitly `not_configured`; successful
+agent installation is not verifier dependency certification. The same verified
+APT and original uv/Python release caches are used. No task image, package
+source, official test script, or native per-attempt limit changes.
+
+The existing `benchmarks/harbor/scripts/run_comparison.py` runs
+`--phase preflight` before `--phase model` with this plan. Native jobs live under
+the frozen source checkout's `jobs/flash1017-recovery-*-20261018{,-input}/`
+directories. Orchestration records live at
+`jobs/flash1017-recovery-orchestration/`. Freeze both outputs before reviewing
+new trajectories or official test contents. Raw logs, trajectories, and `/app`
+artifacts remain local; publish derived findings and hashes only.
+
+The retry implementation passed all **178 Harbor workflow tests**, including
+real subprocess checks that dependency failures are opt-in, bounded, archived,
+and never confused with executed test failures or successful rewards. The
+bilingual adapter README documents the option. These tasks are already exposed
+development tasks; this follow-up does not establish checkpoint causality or a
+new score for the entire 34-task pool.
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/adaptive-rejection-sampler.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/adaptive-rejection-sampler.json
new file mode 100644
index 0000000..7417e2a
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/adaptive-rejection-sampler.json
@@ -0,0 +1,38 @@
+{
+  "n_concurrent_trials": 1,
+  "n_attempts": 1,
+  "agents": [
+    {
+      "name": "harbor_adapter:NanoPyCodeAgent",
+      "model_name": "deepseek/deepseek-flash",
+      "override_setup_timeout_sec": 1800,
+      "kwargs": {
+        "max_turns": 100,
+        "max_tokens": 65536,
+        "time_budget_seconds": 720
+      }
+    }
+  ],
+  "tasks": [
+    {
+      "name": "terminal-bench/adaptive-rejection-sampler",
+      "ref": "sha256:bcaa2399985cd57666018025846289ab25e193ae0dd8fb7f0ffab2410c24d4de"
+    }
+  ],
+  "environment": {
+    "type": "docker",
+    "extra_docker_compose": [
+      "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/dependency-cache.compose.json"
+    ]
+  },
+  "verifier": {
+    "import_path": "harbor_adapter:RetryingVerifier",
+    "kwargs": {
+      "max_attempts": 3,
+      "retry_dependency_errors": true
+    }
+  },
+  "artifacts": [
+    "/app"
+  ]
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/experiment.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/experiment.json
new file mode 100644
index 0000000..b9c62db
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/experiment.json
@@ -0,0 +1,159 @@
+{
+  "schema_version": 1,
+  "dataset_ref": "sha256:7d7bdc1cbedad549fc1140404bd4dc45e5fd0ea7c4186773687d177ad3a0699a",
+  "model": "deepseek/deepseek-flash",
+  "endpoint": "https://api.deepseek.com/anthropic",
+  "max_concurrent_trials": 2,
+  "whole_trial_retries": 0,
+  "provider_checkpoint": {
+    "api_model": "deepseek-flash",
+    "reported_checkpoint": "DeepSeek-V4.1-Flash-1017",
+    "verification": "Official DeepSeek documentation states that deepseek-flash now serves this checkpoint. The API alias is not an immutable checkpoint pin.",
+    "sources": [
+      "https://api-docs.deepseek.com/news/news261017/",
+      "https://api-docs.deepseek.com/quick_start/agent_integrations/"
+    ],
+    "checked_at": "2026-10-18T13:16:26.524186+00:00"
+  },
+  "registered_at": "2026-10-18T14:24:57.680454+00:00",
+  "experiment": "deepseek-v41-flash-1017-recovery-20261018",
+  "arms": {
+    "recovery": {
+      "source_root": "jobs/flash1017-recovery-source",
+      "source_ref": "35fe1a2d6a786619a2adace1cfa9d0acca08d533",
+      "prompt_file": "main-prompt.txt",
+      "prompt_sha256": "37c4d9d44640c0051dc5a20e5954a067b98e8c51be07a48f080aa059e725c55d",
+      "prompt_words": 385
+    }
+  },
+  "tasks": [
+    {
+      "name": "terminal-bench/adaptive-rejection-sampler",
+      "ref": "sha256:bcaa2399985cd57666018025846289ab25e193ae0dd8fb7f0ffab2410c24d4de",
+      "native_agent_timeout_seconds": 900,
+      "work_budget_seconds": 720,
+      "verifier_timeout_seconds": 900.0,
+      "docker_image": "alexgshaw/adaptive-rejection-sampler:20251031",
+      "resources": {
+        "cpus": 1,
+        "memory_mb": 2048,
+        "storage_mb": 10240,
+        "gpus": 0
+      },
+      "category": "scientific-computing",
+      "verifier_dependency_profile": "not_configured",
+      "task_file_sha256": {
+        "README.md": "95408217b9c484ffc18a02a91b7706323b6902919c72b24866a32e27a23543e2",
+        "environment/Dockerfile": "e644a3343e339303230f1441b7383f076d21c69a5a7aa6430b4f8a9d27f15c8b",
+        "environment/protected.tar.gz.enc": "7fc9bf1347bb97da8dede04b2c58638f225070913e222d76627068aebd1b1df2",
+        "instruction.md": "e6d28c28acd76bef61817427c74c2a29e3e8d75f535cbd708afcbfd02c54013f",
+        "solution/solve.sh": "fe443a8082b5183ba1623311ba0ee958888aefa633eff93766b9ae14e9bdd84d",
+        "task.toml": "6ed7b37f74beb46e9d47b0f5fad53025667c9548baff4e96d467b56d429fab1a",
+        "tests/test.sh": "38b43560d173cc2b952c3a3e17b8a480216d84e33450515e047bcb0d806b1e0a",
+        "tests/test_outputs.py": "547dc6e107f034f41703722aeceb6d0236e3fb69116fc3f2fbaba11884de352f"
+      }
+    },
+    {
+      "name": "terminal-bench/torch-pipeline-parallelism",
+      "ref": "sha256:db605337c749a872cea7b5b413429b3915bb4c3efe0f7875f0c46ce81bd8c4fb",
+      "native_agent_timeout_seconds": 900,
+      "work_budget_seconds": 720,
+      "verifier_timeout_seconds": 900.0,
+      "docker_image": "alexgshaw/torch-pipeline-parallelism:20251031",
+      "resources": {
+        "cpus": 1,
+        "memory_mb": 8192,
+        "storage_mb": 10240,
+        "gpus": 0
+      },
+      "category": "software-engineering",
+      "verifier_dependency_profile": "not_configured",
+      "task_file_sha256": {
+        "README.md": "297ddea07c230dfd246a8ab07337dd7abcd8e5e2a2b2e26590cb0a8af779839c",
+        "environment/Dockerfile": "c357ea41e4a4013c5c5980b8fef04af6a5ae250ccc6697cf3bad3127401a931f",
+        "instruction.md": "989bfcf0c54baf6c8cdc391743dbb2f9e858c5779dd85112536875d4d1b2dcd2",
+        "solution/solve.sh": "787198e4178d88b887697aefdd29819f29435ffdfe356545d92f58a30fb38b70",
+        "task.toml": "8b4f9f525c08afbbc4eca9168e46238a61ccb903486773e9ca4da64bb7ffe857",
+        "tests/test.sh": "cc9fd28b6590993fc37e4a99da771cbf04726c8c62fe308e802612241da69dea",
+        "tests/test_outputs.py": "23574246a407163a8e6581c84287cc2f1eb745efbaad2a317a5a28c39958c64f"
+      }
+    }
+  ],
+  "queue": [
+    {
+      "arm": "recovery",
+      "repeat": 1,
+      "task": "terminal-bench/adaptive-rejection-sampler",
+      "config": "benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/adaptive-rejection-sampler.json",
+      "config_sha256": "8d83f8c29d1e53b5a78b96f635881f4cb1a0f7e40953314acd48dc995d609a5c",
+      "job_name": "flash1017-recovery-adaptive-rejection-sampler-20261018"
+    },
+    {
+      "arm": "recovery",
+      "repeat": 1,
+      "task": "terminal-bench/torch-pipeline-parallelism",
+      "config": "benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/torch-pipeline-parallelism.json",
+      "config_sha256": "bc0b10b9e0ee7533fbaee316f84c9e63d7e043e53fc4921926f0e6da6e8991e6",
+      "job_name": "flash1017-recovery-torch-pipeline-parallelism-20261018"
+    }
+  ],
+  "total_model_runs": 2,
+  "preflight": [
+    {
+      "arm": "recovery",
+      "config": "benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/install-only.json",
+      "config_sha256": "a90e454b71e7059f742959ed6e820a13232749b2662b6a07242c13bd425b545f",
+      "job_name": "flash1017-recovery-install-only-20261018"
+    }
+  ],
+  "selection_file": "selection.json",
+  "selection_sha256": "c3a1ade78b2f42888f3e9391a51cdae86b9ddf707ab68c9a72ff9ca77809670c",
+  "image_digests_file": "image-digests.json",
+  "image_digests_sha256": "f7d43972f735f3500e2aa478689b362c00df05d805bb0a994fff4d686c2c7177",
+  "log_dir": "jobs/flash1017-recovery-orchestration",
+  "additional_file_sha256": {
+    "benchmarks/harbor/configs/dev34-main-baseline-20261012/dependency-release-manifest-v5.json": "bdf3d3ef8b7743700c5294c4a511c686a3928c178a77fd4abcba90ee4557ffcc",
+    "benchmarks/harbor/configs/dev34-main-baseline-20261012/transport-v3/apt-cache.conf": "d0e1ffb6dbd02c54978ba2df386326311a59f6d3774990180f22bcc6adfaedbc",
+    "benchmarks/harbor/configs/dev34-main-baseline-20261012/transport-v5/dependency-cache.compose.json": "60d778603976454bb407b8273626851357ee41489cdb0395b06dcc7aab9d151e",
+    "benchmarks/harbor/scripts/apt_index_proxy.py": "bc15f7adf1d7968d627ceca4683ab86742a7188a1b1955ba1fef37ba48e8a936",
+    "benchmarks/harbor/scripts/cache_image.py": "639709c03a7b0c3eb6aa78b4e87e6184b49b4b102853a31e33dd832d759180c0",
+    "benchmarks/harbor/scripts/run_comparison.py": "04f2106f0099a7b4d77e9206a36767f98a5e92c875ec78dd2de858c4936c12a8",
+    "benchmarks/harbor/scripts/uv_release_cache.py": "993cc35b1f3a9e8e171b662ab6d4524c5fe059f0d1f0e275a5bb1930111e0ff5",
+    "benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/dependency-cache.compose.json": "fc2c6a22fb28563171f21808cd103851c69ce02bd3aa8c8c770eea20af87f49f",
+    "benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/dependency-release-manifest.json": "77dfda40e45b870a8ce7f07a1e058d416d94a04ab3d27c2f21bc370f22151394",
+    "benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/python-agent-release-manifest.json": "78aecd7c397c469196f9d246a160bf1fc99ad2d51cb71320b7d26ebf839647ce",
+    "benchmarks/harbor/src/harbor_adapter/verifier.py": "44bf247282437a0092bd7e649fd7fa9c776b6d34a0c14ffedcc10b37a47608ef",
+    "benchmarks/harbor/src/harbor_adapter/_verifier_runner.py": "d718f7c5528c156df549d68ab2deff247a289d08a453c6d083b72d0e9c7013dc"
+  },
+  "authorization": {
+    "request": "Retry the two unscored tasks, at most three attempts.",
+    "scope": "This follow-up explicitly enables retrying recognized pre-test uv network download failures in addition to native verifier timeouts.",
+    "limit": "At most three total verifier attempts per fresh model output, each with the original 900-second limit. Stop immediately on an executed official test result, including failure.",
+    "original_workspace_availability": "The original containers were deleted and artifact manifests contain no /app outputs; exact original workspace regrading is unavailable.",
+    "new_model_policy": "One fresh model trial per selected task, no whole-trial retries. Within each new trial, reuse its output and container for all authorized verifier attempts.",
+    "artifact_policy": "Collect the complete /app directory for future verification without another model run."
+  },
+  "budget_policy": "Native task agent timeout minus 180 seconds, separate 1800-second setup. Native verifier per-attempt time limits; max three attempts, retry only timeouts against the same agent output.",
+  "preflight_policy": "Two fresh installation-only checks. Both verifier dependency profiles remain not_configured; no verifier dependency certification is claimed.",
+  "information_policy": "All selected tasks have prior test and trace exposure. Supply only the original task instruction and unchanged headless prompt; inject no previous diagnoses or fixes. Freeze both outputs before new trace or test-content review.",
+  "agent_source_sha256": {
+    "nanopycodeagent/__init__.py": "b922299f65b3db2fb1514df5aeda4413e599f595d2595f45a353b4dc551a1eb1",
+    "nanopycodeagent/agent.py": "aa58960803be77797c2addf65d7cc785a4be54f40ad58d53b9eea9fca1cf4b56",
+    "nanopycodeagent/atif.py": "ca4793b971d125d904888b858452db8db30e21f0d9b87818133062fb6cebe270",
+    "nanopycodeagent/bash_tool.py": "dcf0366fb9ebdca258d4c40b1a50b78b2ec79fc4b85cbb253ac9a9ce9534574f",
+    "nanopycodeagent/cli.py": "e335789b3f09440abd441547a2e3e2feb32b90c68e9ba5fc8237de8f6bac2932",
+    "nanopycodeagent/cost.py": "30a8df17942c424796e8532f45e1cb0784ed6541cea44284b863db3dae2f7b70",
+    "nanopycodeagent/deadline.py": "26a861721e6897a4e02e6ce364cb3ef1f1dc7d4ad9fb2f772404276f5c9e9d5f",
+    "nanopycodeagent/edit_tool.py": "78e2bec911279082b081366ab05cb4cbb37d8910b20b6245cccd4b0424489d0d",
+    "nanopycodeagent/event_journal.py": "8974919c2d460afb52bc57ec37e7e271851290c0845d01b4ad0f9f760cde0814",
+    "nanopycodeagent/read_tool.py": "e0075da19f2e1c6ccf34a3056a0aec92398e5afe67869937a096c6f1e04181c3",
+    "nanopycodeagent/settings.py": "491bf5c6eca50eb4c5fe04b7bfe62a8bfc7d2e1fad05b8eeac1680c2fcf3674e",
+    "nanopycodeagent/terminal.py": "8d27b40976b2b6d476e5106dfea982b055115435b4ba16b7f758412a05b75f04",
+    "nanopycodeagent/tool_validation.py": "3a1a85a7222dd1504bdf58f37e65a51d11de13d16c4e7df42c72b4cd58b5e5e5",
+    "nanopycodeagent/transport.py": "152d3ff78be0e3b2d92e0abbbfb652d73466aff57ad316c06dc0e072411167fe",
+    "nanopycodeagent/write_tool.py": "7671e83464fa1e942f25742fc573d0712e32053b5fc0291257884b027f7f6206"
+  },
+  "adapter_change": "Only the verifier adapter gains opt-in, bounded pre-test dependency-download recovery. The agent Python source is byte-identical to the original six-trial run.",
+  "prior_run": "deepseek-v41-flash-1017-rerun-20261018",
+  "scope": "Two new model trials after explicitly authorized infrastructure recovery, not a regrade of the old outputs and not an update to the full development-pool score."
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/image-digests.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/image-digests.json
new file mode 100644
index 0000000..2b3f772
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/image-digests.json
@@ -0,0 +1,25 @@
+{
+  "observed_at": "2026-10-18T13:16:26.429384+00:00",
+  "all_match_previous_baseline": true,
+  "images": [
+    {
+      "task": "terminal-bench/torch-pipeline-parallelism",
+      "image": "alexgshaw/torch-pipeline-parallelism:20251031",
+      "image_id": "sha256:3cb7b39d86f1704b20e668851da0b08b34b3270c01d583f136d65190984a8d1b",
+      "repo_digests": [
+        "alexgshaw/torch-pipeline-parallelism@sha256:3cb7b39d86f1704b20e668851da0b08b34b3270c01d583f136d65190984a8d1b"
+      ],
+      "size_bytes": 29725531
+    },
+    {
+      "task": "terminal-bench/adaptive-rejection-sampler",
+      "image": "alexgshaw/adaptive-rejection-sampler:20251031",
+      "image_id": "sha256:985bed39e8448bc95f138b5f766b59003a95c4be171c96af8e76448338f3dd8f",
+      "repo_digests": [
+        "alexgshaw/adaptive-rejection-sampler@sha256:985bed39e8448bc95f138b5f766b59003a95c4be171c96af8e76448338f3dd8f"
+      ],
+      "size_bytes": 29731756
+    }
+  ],
+  "source_observation": "Same original identities verified before and after the initial six-trial run."
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/install-only.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/install-only.json
new file mode 100644
index 0000000..4d18262
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/install-only.json
@@ -0,0 +1,43 @@
+{
+  "n_concurrent_trials": 2,
+  "n_attempts": 1,
+  "agents": [
+    {
+      "name": "harbor_adapter:NanoPyCodeAgent",
+      "model_name": "deepseek/deepseek-flash",
+      "override_setup_timeout_sec": 1800,
+      "kwargs": {
+        "max_turns": 100,
+        "max_tokens": 65536,
+        "time_budget_seconds": 720
+      }
+    }
+  ],
+  "tasks": [
+    {
+      "name": "terminal-bench/adaptive-rejection-sampler",
+      "ref": "sha256:bcaa2399985cd57666018025846289ab25e193ae0dd8fb7f0ffab2410c24d4de"
+    },
+    {
+      "name": "terminal-bench/torch-pipeline-parallelism",
+      "ref": "sha256:db605337c749a872cea7b5b413429b3915bb4c3efe0f7875f0c46ce81bd8c4fb"
+    }
+  ],
+  "environment": {
+    "type": "docker",
+    "extra_docker_compose": [
+      "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/dependency-cache.compose.json"
+    ]
+  },
+  "verifier": {
+    "import_path": "harbor_adapter:RetryingVerifier",
+    "kwargs": {
+      "max_attempts": 3,
+      "retry_dependency_errors": true
+    }
+  },
+  "artifacts": [
+    "/app"
+  ],
+  "install_only": true
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/main-prompt.txt b/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/main-prompt.txt
new file mode 100644
index 0000000..56116f5
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/main-prompt.txt
@@ -0,0 +1,9 @@
+You are nanoPyCodeAgent, a coding agent running non-interactively on a task handed to you up front. There is no user to reply to you: never ask a clarifying question, never stop to wait for confirmation, and never present a plan for approval — decide on your own and carry it out. Work the task through to the end, then check the result with the tools instead of assuming it worked. When it is done, answer with a short summary and no further tool calls: that reply is what ends the run. Runtime budget reminders report remaining time and model replies; when either is running low, prioritize saving the required deliverables, necessary verification, and a final summary. Do not start optional work after the requirements are satisfied.
+
+Identify the required deliverables, interfaces, and constraints. Treat examples as evidence, not as the full contract unless the task says so. Inspect relevant files, dependencies, and existing checks instead of guessing. Task-supplied workspace files, including source, executables, and tests, are available references unless the task restricts their use. Consult general documentation as needed. Do not seek or use task-specific reference solutions or hidden evaluation tests from outside the supplied materials unless the task explicitly requests them. Derive checks from the stated requirements.
+
+Verify the requested behavior through the real entry point when feasible. Check the content and meaning of outputs, not only command success or file existence. For logic or numerical work, choose a few representative and boundary cases justified by the task, using an independent reference, invariant, or consistency check when available. Do not derive every expectation from the implementation being tested or weaken requirements to make a check pass. Keep verification proportional to the task and remaining budget.
+
+When a check fails, inspect the error and relevant output, revise the underlying assumption or implementation, and rerun the affected check. Keep the task's constraints intact. Once appropriate checks support the required behavior, save the deliverables and finish. State what was verified and any remaining limitation without claiming unobserved success.
+
+Prefer the read tool for viewing files, the edit tool for changing part of an existing file, and the write tool for creating files or rewriting them whole. Use the bash tool to run commands, search with grep, and complete tasks that need real command output instead of guessing.
\ No newline at end of file
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/selection.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/selection.json
new file mode 100644
index 0000000..7c5eee0
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/selection.json
@@ -0,0 +1,21 @@
+{
+  "previous_results_path": "benchmarks/harbor/results/deepseek-v41-flash-1017-rerun-20261018.json",
+  "previous_results_sha256": "d6cce2cb8048fbb7d5ea97e5a886d3c2caa89d4883039ef1c52e4422b3b0c87d",
+  "previous_trials": [
+    {
+      "task": "terminal-bench/adaptive-rejection-sampler",
+      "trial": "adaptive-rejection-sampler__Bim6vqF",
+      "classification": "verifier_error",
+      "raw_reward": 0.0,
+      "score": null
+    },
+    {
+      "task": "terminal-bench/torch-pipeline-parallelism",
+      "trial": "torch-pipeline-parallelism__q96EdZw",
+      "classification": "verifier_error",
+      "raw_reward": 0.0,
+      "score": null
+    }
+  ],
+  "rule": "Only the two unscored verifier errors; all earlier trial records remain immutable."
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/torch-pipeline-parallelism.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/torch-pipeline-parallelism.json
new file mode 100644
index 0000000..f24a4b4
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/torch-pipeline-parallelism.json
@@ -0,0 +1,38 @@
+{
+  "n_concurrent_trials": 1,
+  "n_attempts": 1,
+  "agents": [
+    {
+      "name": "harbor_adapter:NanoPyCodeAgent",
+      "model_name": "deepseek/deepseek-flash",
+      "override_setup_timeout_sec": 1800,
+      "kwargs": {
+        "max_turns": 100,
+        "max_tokens": 65536,
+        "time_budget_seconds": 720
+      }
+    }
+  ],
+  "tasks": [
+    {
+      "name": "terminal-bench/torch-pipeline-parallelism",
+      "ref": "sha256:db605337c749a872cea7b5b413429b3915bb4c3efe0f7875f0c46ce81bd8c4fb"
+    }
+  ],
+  "environment": {
+    "type": "docker",
+    "extra_docker_compose": [
+      "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/dependency-cache.compose.json"
+    ]
+  },
+  "verifier": {
+    "import_path": "harbor_adapter:RetryingVerifier",
+    "kwargs": {
+      "max_attempts": 3,
+      "retry_dependency_errors": true
+    }
+  },
+  "artifacts": [
+    "/app"
+  ]
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/README.md b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/README.md
new file mode 100644
index 0000000..3ff894b
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/README.md
@@ -0,0 +1,91 @@
+# DeepSeek Flash 1017: rerun the six previous failures
+
+**Completed:** three passes, one scored failure, and two unscored verifier
+dependency-download errors. See the
+[final report](../../reports/deepseek-v41-flash-1017-rerun-20261018.md) and
+[audited results](../../results/deepseek-v41-flash-1017-rerun-20261018.json).
+The sampler and pipeline native zeros are retained as raw evidence; their
+analytical scores are null because official tests never started.
+
+This experiment selects all six failed tasks from the latest
+[34-task baseline](../../reports/dev34-main-baseline-20261012.md):
+`adaptive-rejection-sampler`, `db-wal-recovery`, `kv-store-grpc`,
+`query-optimize`, `torch-pipeline-parallelism`, and `regex-chess`.
+The [selection](selection.json) binds the previous result file and trial IDs.
+
+The [final registered plan](experiment-v2.json) runs each task once with
+`deepseek/deepseek-flash` against the official DeepSeek Anthropic-compatible
+endpoint. On October 18, 2026, DeepSeek's
+[API documentation](https://api-docs.deepseek.com/quick_start/agent_integrations/)
+states that `deepseek-flash` serves **DeepSeek-V4.1-Flash-1017**; see also the
+[October 17 announcement](https://api-docs.deepseek.com/news/news261017/).
+The API alias is not an immutable checkpoint pin.
+
+## Matched conditions
+
+- Agent and adapter: the original unchanged checkout at
+  `2eb99b608dfb39464d17d109a4a8f11afda607a9`, located at
+  `jobs/dev34-baseline-main-source`.
+- Prompt: the original [385-word headless prompt](main-prompt.txt).
+- Tasks: original pinned revisions, complete file hashes, resources, and
+  [image identities](image-digests.json), verified against the previous run.
+- Limits: 100 replies, 65,536 output tokens per request, and the native agent
+  timeout minus 180 seconds. That gives 720 seconds for five tasks and 3420
+  seconds for regex chess. Setup has a separate 1800-second limit.
+- Queue: ascending work budget, then task name; consecutive pairs complete
+  before the next pair starts, with concurrency two.
+- Transport: the previous v5 dependency-cache composition, including verified
+  original APT indexes and uv/Python release archives, plus the identical
+  original CPython archive described below. Images, package sources, official
+  tests, and native verifier limits remain unchanged.
+- Attempts: one fresh model trial per task, no whole-trial retries or
+  outcome-based replacements. Only verifier timeouts may retry, up to three
+  attempts with the same model output and original per-attempt limit.
+
+## Execution and evidence
+
+The existing `benchmarks/harbor/scripts/run_comparison.py` validates the plan and
+invokes the standard `uv run --project benchmarks/harbor python -m harbor_adapter
+run --config ...` entry point from the frozen source checkout. Run its
+`--phase preflight` before `--phase model` with this experiment's plan.
+Credentials come from the environment and must not be committed.
+
+Fresh install-only checks cover all six tasks before model work. All six
+verifier dependency profiles are **`not_configured`**: successful agent
+installation does not certify their verifier dependencies. Full trials repeat
+setup in fresh containers. Installation failures and verifier errors are
+reported separately from scored failures.
+
+The orchestration records live under
+`jobs/deepseek-v41-flash-1017-rerun-20261018-v2-orchestration/`; native inputs,
+source snapshots, logs, and results live under the frozen source checkout's
+`jobs/flash1017-*-20261018{,-input}/` directories. Raw logs and trajectories
+remain local because they can contain credentials. Only derived results and
+evidence hashes belong in version control.
+
+Freeze every model output before reviewing new trajectories or official test
+contents. Previous failure diagnoses and solutions are not supplied to the
+model. This is an exposed development pool selected for previous failure;
+one new trial per task cannot isolate a checkpoint effect, establish
+reliability, or provide a fresh score for the entire 34-task pool.
+
+## Pre-model transport amendment
+
+The [original plan](experiment.json) started two installation-only trials.
+Both downloaded CPython 3.13.9+20251120, which uv selected normally but which
+was absent from the previous cache registration. Transfers from the original
+GitHub URL progressed slowly. Before any model work, both installation trials
+were administratively interrupted; their native `CancelledError` results and
+the intervention record remain local. The other four original installation
+trials had not started.
+
+The existing release-cache utility fetched and verified the same original
+archive using public GitHub release metadata: **34,257,108 bytes**, SHA256
+`1bc9a94ab2425fb6ae58e45f69cab61647759b57895649701390822d976fa923`.
+The [additional manifest](python-agent-release-manifest.json),
+[combined manifest](dependency-release-manifest.json), and
+[v2 composition](transport-v2/dependency-cache.compose.json) register this cache
+hit. The final plan binds their hashes and repeats all six installation checks
+afresh. Python selection, archive contents, source code, tasks, resources,
+budgets, and model queue are unchanged. No scored model trial was interrupted
+or retried.
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/adaptive-rejection-sampler.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/adaptive-rejection-sampler.json
new file mode 100644
index 0000000..42caf2d
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/adaptive-rejection-sampler.json
@@ -0,0 +1,28 @@
+{
+  "n_concurrent_trials": 1,
+  "n_attempts": 1,
+  "agents": [
+    {
+      "name": "harbor_adapter:NanoPyCodeAgent",
+      "model_name": "deepseek/deepseek-flash",
+      "override_setup_timeout_sec": 1800,
+      "kwargs": {
+        "max_turns": 100,
+        "max_tokens": 65536,
+        "time_budget_seconds": 720
+      }
+    }
+  ],
+  "tasks": [
+    {
+      "name": "terminal-bench/adaptive-rejection-sampler",
+      "ref": "sha256:bcaa2399985cd57666018025846289ab25e193ae0dd8fb7f0ffab2410c24d4de"
+    }
+  ],
+  "environment": {
+    "type": "docker",
+    "extra_docker_compose": [
+      "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/configs/dev34-main-baseline-20261012/transport-v5/dependency-cache.compose.json"
+    ]
+  }
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/db-wal-recovery.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/db-wal-recovery.json
new file mode 100644
index 0000000..f9aa117
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/db-wal-recovery.json
@@ -0,0 +1,28 @@
+{
+  "n_concurrent_trials": 1,
+  "n_attempts": 1,
+  "agents": [
+    {
+      "name": "harbor_adapter:NanoPyCodeAgent",
+      "model_name": "deepseek/deepseek-flash",
+      "override_setup_timeout_sec": 1800,
+      "kwargs": {
+        "max_turns": 100,
+        "max_tokens": 65536,
+        "time_budget_seconds": 720
+      }
+    }
+  ],
+  "tasks": [
+    {
+      "name": "terminal-bench/db-wal-recovery",
+      "ref": "sha256:07e9193513bfce038b9f9d84962ecc85683eee0a642d14ee6fb7104b086ddab4"
+    }
+  ],
+  "environment": {
+    "type": "docker",
+    "extra_docker_compose": [
+      "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/configs/dev34-main-baseline-20261012/transport-v5/dependency-cache.compose.json"
+    ]
+  }
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/dependency-release-manifest.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/dependency-release-manifest.json
new file mode 100644
index 0000000..e762efb
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/dependency-release-manifest.json
@@ -0,0 +1,58 @@
+{
+  "registered_at": "2026-10-18T13:25:43.131474+00:00",
+  "metadata_source": "Original release metadata preserved by the source manifests; no versions or payloads substituted.",
+  "source_manifests": {
+    "benchmarks/harbor/configs/dev34-main-baseline-20261012/dependency-release-manifest-v5.json": "bdf3d3ef8b7743700c5294c4a511c686a3928c178a77fd4abcba90ee4557ffcc",
+    "python-agent-release-manifest.json": "78aecd7c397c469196f9d246a160bf1fc99ad2d51cb71320b7d26ebf839647ce"
+  },
+  "assets": [
+    {
+      "version": "0.9.11",
+      "filename": "uv-x86_64-unknown-linux-gnu.tar.gz",
+      "url": "https://github.com/astral-sh/uv/releases/download/0.9.11/uv-x86_64-unknown-linux-gnu.tar.gz",
+      "digest": "sha256:817c0722b437b4b45b9a7e0231616a09db76bab1b8d178ba7a9680c690db19f0",
+      "size": 21699161,
+      "asset_id": 318993196,
+      "release_metadata_sha256": "c1a1378ece65ce1aa8343e57bdbd305289a8fc6d3d9180e8bb6b0f69acd91413"
+    },
+    {
+      "version": "0.9.5",
+      "filename": "uv-x86_64-unknown-linux-gnu.tar.gz",
+      "url": "https://github.com/astral-sh/uv/releases/download/0.9.5/uv-x86_64-unknown-linux-gnu.tar.gz",
+      "digest": "sha256:2cf10babba653310606f8b49876cfb679928669e7ddaa1fb41fb00ce73e64f66",
+      "size": 21370871,
+      "asset_id": 306970260,
+      "release_metadata_sha256": "1921b17d3370aafb690b29d1d121ee8719ef3d3dbc0c7cf3b948c3059a927a5d"
+    },
+    {
+      "version": "20251014",
+      "filename": "cpython-3.13.9+20251014-x86_64-unknown-linux-gnu-install_only_stripped.tar.gz",
+      "url": "https://github.com/astral-sh/python-build-standalone/releases/download/20251014/cpython-3.13.9%2B20251014-x86_64-unknown-linux-gnu-install_only_stripped.tar.gz",
+      "digest": "sha256:c0ccda275948c79996e46993c2c5476ff5cca606dee530f1dea712179131b348",
+      "size": 33517986,
+      "asset_id": 304305340,
+      "release_metadata_sha256": "8e5db21cfcc2bcafcb2127f269f00cbfdd1f74842a685acd2a14e3c1ad6f8394",
+      "repository": "astral-sh/python-build-standalone"
+    },
+    {
+      "version": "20251120",
+      "filename": "cpython-3.14.0+20251120-x86_64-unknown-linux-gnu-install_only_stripped.tar.gz",
+      "url": "https://github.com/astral-sh/python-build-standalone/releases/download/20251120/cpython-3.14.0%2B20251120-x86_64-unknown-linux-gnu-install_only_stripped.tar.gz",
+      "digest": "sha256:66f385305ae0eefd6b65e2cf942bc91d943a61e26fb7581e029f760a2b04f393",
+      "size": 35376131,
+      "asset_id": 318881850,
+      "release_metadata_sha256": "e3a737862d2dddf11c73bfef6c2f2579e9d4ce5bd271cca53790c998590dee58",
+      "repository": "astral-sh/python-build-standalone"
+    },
+    {
+      "version": "20251120",
+      "filename": "cpython-3.13.9+20251120-x86_64-unknown-linux-gnu-install_only_stripped.tar.gz",
+      "url": "https://github.com/astral-sh/python-build-standalone/releases/download/20251120/cpython-3.13.9%2B20251120-x86_64-unknown-linux-gnu-install_only_stripped.tar.gz",
+      "digest": "sha256:1bc9a94ab2425fb6ae58e45f69cab61647759b57895649701390822d976fa923",
+      "size": 34257108,
+      "asset_id": 318881608,
+      "release_metadata_sha256": "705690ab2fd466e90b4c868635c529f3c7fd0cb597d399e350f399e3f3e35e6c",
+      "repository": "astral-sh/python-build-standalone"
+    }
+  ]
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/experiment-v2.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/experiment-v2.json
new file mode 100644
index 0000000..7ccd435
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/experiment-v2.json
@@ -0,0 +1,288 @@
+{
+  "schema_version": 1,
+  "dataset_ref": "sha256:7d7bdc1cbedad549fc1140404bd4dc45e5fd0ea7c4186773687d177ad3a0699a",
+  "model": "deepseek/deepseek-flash",
+  "endpoint": "https://api.deepseek.com/anthropic",
+  "arms": {
+    "main": {
+      "source_root": "jobs/dev34-baseline-main-source",
+      "source_ref": "2eb99b608dfb39464d17d109a4a8f11afda607a9",
+      "prompt_file": "main-prompt.txt",
+      "prompt_sha256": "37c4d9d44640c0051dc5a20e5954a067b98e8c51be07a48f080aa059e725c55d",
+      "prompt_words": 385
+    }
+  },
+  "max_concurrent_trials": 2,
+  "whole_trial_retries": 0,
+  "registered_at": "2026-10-18T13:25:43.132414+00:00",
+  "experiment": "deepseek-v41-flash-1017-rerun-20261018",
+  "selection_file": "selection.json",
+  "selection_sha256": "04c7a99606bc58a15ea6577548fdc81f12f8c6ef2770127743051ef2c4fd9bb7",
+  "total_model_runs": 6,
+  "tasks": [
+    {
+      "name": "terminal-bench/adaptive-rejection-sampler",
+      "ref": "sha256:bcaa2399985cd57666018025846289ab25e193ae0dd8fb7f0ffab2410c24d4de",
+      "native_agent_timeout_seconds": 900,
+      "work_budget_seconds": 720,
+      "verifier_timeout_seconds": 900.0,
+      "docker_image": "alexgshaw/adaptive-rejection-sampler:20251031",
+      "resources": {
+        "cpus": 1,
+        "memory_mb": 2048,
+        "storage_mb": 10240,
+        "gpus": 0
+      },
+      "category": "scientific-computing",
+      "verifier_dependency_profile": "not_configured",
+      "task_file_sha256": {
+        "README.md": "95408217b9c484ffc18a02a91b7706323b6902919c72b24866a32e27a23543e2",
+        "environment/Dockerfile": "e644a3343e339303230f1441b7383f076d21c69a5a7aa6430b4f8a9d27f15c8b",
+        "environment/protected.tar.gz.enc": "7fc9bf1347bb97da8dede04b2c58638f225070913e222d76627068aebd1b1df2",
+        "instruction.md": "e6d28c28acd76bef61817427c74c2a29e3e8d75f535cbd708afcbfd02c54013f",
+        "solution/solve.sh": "fe443a8082b5183ba1623311ba0ee958888aefa633eff93766b9ae14e9bdd84d",
+        "task.toml": "6ed7b37f74beb46e9d47b0f5fad53025667c9548baff4e96d467b56d429fab1a",
+        "tests/test.sh": "38b43560d173cc2b952c3a3e17b8a480216d84e33450515e047bcb0d806b1e0a",
+        "tests/test_outputs.py": "547dc6e107f034f41703722aeceb6d0236e3fb69116fc3f2fbaba11884de352f"
+      }
+    },
+    {
+      "name": "terminal-bench/db-wal-recovery",
+      "ref": "sha256:07e9193513bfce038b9f9d84962ecc85683eee0a642d14ee6fb7104b086ddab4",
+      "native_agent_timeout_seconds": 900,
+      "work_budget_seconds": 720,
+      "verifier_timeout_seconds": 900.0,
+      "docker_image": "alexgshaw/db-wal-recovery:20251031",
+      "resources": {
+        "cpus": 1,
+        "memory_mb": 2048,
+        "storage_mb": 10240,
+        "gpus": 0
+      },
+      "category": "file-operations",
+      "verifier_dependency_profile": "not_configured",
+      "task_file_sha256": {
+        "README.md": "600762031a1a02ecc3b46d70f7683f3ff930b47c8a3d0e27cc31b146158c3126",
+        "environment/Dockerfile": "63a260531c62f4ae72060db17f295c0aa03cfc6321f8162c324b7aa6b3f81302",
+        "environment/main.db": "68cb288eb29e76d814b4e7ee6887b37c415675e644f88173bc191348e290d119",
+        "environment/main.db-wal.encrypted": "68753bdce26196a53194260d001544cc6bb8f9f38e2f2f6826d09798e31e6e26",
+        "instruction.md": "5a20375162921b8911c67035d1905df7fd335d1bed03044271e608b2c3c93b26",
+        "solution/solve.sh": "ed8530c94103bf19783a1d41bc63fe4db478cdd9629f22fee25c45a45699f5e8",
+        "task.toml": "1cdc2c615e259bc10265efbe71b518332070c1c002e226566399e9788f18d0aa",
+        "tests/test.sh": "4770437ea96c3cc84684b4f99d55fb148fcac09f9ea1e8ef49de487716e6c334",
+        "tests/test_outputs.py": "88dc5c2a3abba75bca0e0f9136223ae5bb585d5984b8d9eba51a7829c1d2ba49"
+      }
+    },
+    {
+      "name": "terminal-bench/kv-store-grpc",
+      "ref": "sha256:973c5d4c111fb61a344457936f1c36400acd2d9e44389e7b319586fe23a7a307",
+      "native_agent_timeout_seconds": 900,
+      "work_budget_seconds": 720,
+      "verifier_timeout_seconds": 900.0,
+      "docker_image": "alexgshaw/kv-store-grpc:20251031",
+      "resources": {
+        "cpus": 1,
+        "memory_mb": 2048,
+        "storage_mb": 10240,
+        "gpus": 0
+      },
+      "category": "software-engineering",
+      "verifier_dependency_profile": "not_configured",
+      "task_file_sha256": {
+        "README.md": "51fdc2ad06f0adb1270164f247ea41799b2ee2ad15490e48ddbf5e79d3767aaa",
+        "environment/Dockerfile": "47c9982789092437f950736110b2bd18899c39b8fd20f3af3c95b848e564a197",
+        "instruction.md": "f4b3facc199d5b1c78db6e603ffa29ec1b93a77ff7d291e1a92b42e6bf582bed",
+        "solution/solve.sh": "6bee2938ac9731d1156b0ca02b28c3d5793bf2cbe5f995c863d397093af01cb7",
+        "task.toml": "6ff0f30a526bfff86a67b0b53439b9965229c0f368167dc1a5d05f34d82299a3",
+        "tests/test.sh": "c348467a076bf614d0a850e25b2263f85bbf7b4c7a8d344fbd5454a0611a0df8",
+        "tests/test_outputs.py": "115de2e964bbe689d07930527113422ef8a9e4be4eb958ad84594e3fcd2b1b0e"
+      }
+    },
+    {
+      "name": "terminal-bench/query-optimize",
+      "ref": "sha256:169496ea6843cb403b0860132675baf7aa0df0ac8b86221068e27d86395260f4",
+      "native_agent_timeout_seconds": 900,
+      "work_budget_seconds": 720,
+      "verifier_timeout_seconds": 1800.0,
+      "docker_image": "alexgshaw/query-optimize:20251031",
+      "resources": {
+        "cpus": 1,
+        "memory_mb": 2048,
+        "storage_mb": 10240,
+        "gpus": 0
+      },
+      "category": "data-science",
+      "verifier_dependency_profile": "not_configured",
+      "task_file_sha256": {
+        "README.md": "a82c53e333a0ac0b0b410b5e26cb8e7305dece31db84d340248ffb13d5a45160",
+        "environment/Dockerfile": "064818771f200e0517d22068f87229a7a95d0a82b0778c84ca7203bf1926a0b8",
+        "environment/task-deps/my-sql-query.sql": "43382792f0f438625d643e34abe144febd5669518f4ab1ab550a15c6d9314a21",
+        "instruction.md": "473e12d320d2732f679fb631c2a4dab4a01deea6775686718ae0c417b8a39447",
+        "solution/solve.sh": "0d95f3cedb287ff9d47c796cbaf812c300d1222c10fa80c49e8acb0e81eb6978",
+        "task.toml": "6e88f2a86df699594bcb1bf81df4ed694505f79c9420f671335b02524bbe3868",
+        "tests/golden.sql": "4da1885d56b3ffa7b7b22b422996d018795611f6963499d551582881b74b84fd",
+        "tests/test.sh": "4770437ea96c3cc84684b4f99d55fb148fcac09f9ea1e8ef49de487716e6c334",
+        "tests/test_outputs.py": "2068ae80c57af8d52a7ae510f0666732c057806e556b480970f36f14bae683f6"
+      }
+    },
+    {
+      "name": "terminal-bench/torch-pipeline-parallelism",
+      "ref": "sha256:db605337c749a872cea7b5b413429b3915bb4c3efe0f7875f0c46ce81bd8c4fb",
+      "native_agent_timeout_seconds": 900,
+      "work_budget_seconds": 720,
+      "verifier_timeout_seconds": 900.0,
+      "docker_image": "alexgshaw/torch-pipeline-parallelism:20251031",
+      "resources": {
+        "cpus": 1,
+        "memory_mb": 8192,
+        "storage_mb": 10240,
+        "gpus": 0
+      },
+      "category": "software-engineering",
+      "verifier_dependency_profile": "not_configured",
+      "task_file_sha256": {
+        "README.md": "297ddea07c230dfd246a8ab07337dd7abcd8e5e2a2b2e26590cb0a8af779839c",
+        "environment/Dockerfile": "c357ea41e4a4013c5c5980b8fef04af6a5ae250ccc6697cf3bad3127401a931f",
+        "instruction.md": "989bfcf0c54baf6c8cdc391743dbb2f9e858c5779dd85112536875d4d1b2dcd2",
+        "solution/solve.sh": "787198e4178d88b887697aefdd29819f29435ffdfe356545d92f58a30fb38b70",
+        "task.toml": "8b4f9f525c08afbbc4eca9168e46238a61ccb903486773e9ca4da64bb7ffe857",
+        "tests/test.sh": "cc9fd28b6590993fc37e4a99da771cbf04726c8c62fe308e802612241da69dea",
+        "tests/test_outputs.py": "23574246a407163a8e6581c84287cc2f1eb745efbaad2a317a5a28c39958c64f"
+      }
+    },
+    {
+      "name": "terminal-bench/regex-chess",
+      "ref": "sha256:e763e0ac1c9759081af0a4a82ba51b8cf9ae5485a93de3bbe42d7d344597bd78",
+      "native_agent_timeout_seconds": 3600,
+      "work_budget_seconds": 3420,
+      "verifier_timeout_seconds": 3600.0,
+      "docker_image": "alexgshaw/regex-chess:20251031",
+      "resources": {
+        "cpus": 1,
+        "memory_mb": 2048,
+        "storage_mb": 10240,
+        "gpus": 0
+      },
+      "category": "software-engineering",
+      "verifier_dependency_profile": "not_configured",
+      "task_file_sha256": {
+        "README.md": "9296cbee592fc996c58d228b51fee303647e5ecfae82d5666d42a40cb30ec355",
+        "environment/Dockerfile": "2616cd775252303839dfda0ee2f047eb862952b053b0103bd6b2460ef90c11b7",
+        "environment/check.py": "f9c416ae9ea01dcdd5f6a43170c8434b8c0a0b389a4a30876a2048e277d191c7",
+        "instruction.md": "a6ed3fcc54fc92d71a84af921942422d299ca89b410dcd92cd05a53d0d06436c",
+        "solution/solve.sh": "b728bf0ea392a3d244afbd0818d36933c8d3e190e47743c1382cd518d518a2e9",
+        "task.toml": "07fe52636fcc4bc9fff25d5630af145b97a5331219d719cc8decb4ca67ea21d4",
+        "tests/test.sh": "8f7f00a1c4dde7fcca6c918c768f2aba432b8c4007c15989716c68091950da87",
+        "tests/test_outputs.py": "4983c44d21c56dd8191381eff0304afc8b87bb714527ea9036b3319c015a30b4"
+      }
+    }
+  ],
+  "preflight": [
+    {
+      "arm": "main",
+      "config": "benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/install-only.json",
+      "config_sha256": "33559926b81dba3eff75f34ef962bd5f0223b8b1601660855dc9277eb8c1bf9c",
+      "job_name": "flash1017-install-only-v2-20261018"
+    }
+  ],
+  "queue": [
+    {
+      "arm": "main",
+      "repeat": 1,
+      "task": "terminal-bench/adaptive-rejection-sampler",
+      "config": "benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/adaptive-rejection-sampler.json",
+      "job_name": "flash1017-adaptive-rejection-sampler-20261018",
+      "config_sha256": "d9ca780ac2d7f9fcffbde5c621e834c360fc6d977e50e85702d59ec86d60d88e"
+    },
+    {
+      "arm": "main",
+      "repeat": 1,
+      "task": "terminal-bench/db-wal-recovery",
+      "config": "benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/db-wal-recovery.json",
+      "job_name": "flash1017-db-wal-recovery-20261018",
+      "config_sha256": "a0b6c8f879708ac1fa5b8161d3fc5430efe8fbc334ef643651c5a407731dcdc6"
+    },
+    {
+      "arm": "main",
+      "repeat": 1,
+      "task": "terminal-bench/kv-store-grpc",
+      "config": "benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/kv-store-grpc.json",
+      "job_name": "flash1017-kv-store-grpc-20261018",
+      "config_sha256": "be363d3b1e2a67cb63c134d57200e670ec0fc695a3f4191826e2fe6eefc336bf"
+    },
+    {
+      "arm": "main",
+      "repeat": 1,
+      "task": "terminal-bench/query-optimize",
+      "config": "benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/query-optimize.json",
+      "job_name": "flash1017-query-optimize-20261018",
+      "config_sha256": "35d7fb0c2d0e998d924b7c4ed644d7841a2a8e644febcb85cf3a06a26f64fa1e"
+    },
+    {
+      "arm": "main",
+      "repeat": 1,
+      "task": "terminal-bench/torch-pipeline-parallelism",
+      "config": "benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/torch-pipeline-parallelism.json",
+      "job_name": "flash1017-torch-pipeline-parallelism-20261018",
+      "config_sha256": "27f847ab5e77f92d4d0df85de79c4c65e3595dbf52e9e82315026230a7c4ca11"
+    },
+    {
+      "arm": "main",
+      "repeat": 1,
+      "task": "terminal-bench/regex-chess",
+      "config": "benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/regex-chess.json",
+      "job_name": "flash1017-regex-chess-20261018",
+      "config_sha256": "e357a1bfb8588d93ada2deb4e5f3ed5e0e130ba23ca850253d194d72268a7d37"
+    }
+  ],
+  "log_dir": "jobs/deepseek-v41-flash-1017-rerun-20261018-v2-orchestration",
+  "image_digests_file": "image-digests.json",
+  "image_digests_sha256": "1575f2ddb2e1d5cf088a01af5b43a5ed24271ea0bc6bed726c8f3a3a04fc712c",
+  "additional_file_sha256": {
+    "benchmarks/harbor/configs/dev34-main-baseline-20261012/dependency-release-manifest-v5.json": "bdf3d3ef8b7743700c5294c4a511c686a3928c178a77fd4abcba90ee4557ffcc",
+    "benchmarks/harbor/configs/dev34-main-baseline-20261012/transport-v3/apt-cache.conf": "d0e1ffb6dbd02c54978ba2df386326311a59f6d3774990180f22bcc6adfaedbc",
+    "benchmarks/harbor/configs/dev34-main-baseline-20261012/transport-v5/dependency-cache.compose.json": "60d778603976454bb407b8273626851357ee41489cdb0395b06dcc7aab9d151e",
+    "benchmarks/harbor/scripts/apt_index_proxy.py": "bc15f7adf1d7968d627ceca4683ab86742a7188a1b1955ba1fef37ba48e8a936",
+    "benchmarks/harbor/scripts/cache_image.py": "639709c03a7b0c3eb6aa78b4e87e6184b49b4b102853a31e33dd832d759180c0",
+    "benchmarks/harbor/scripts/run_comparison.py": "04f2106f0099a7b4d77e9206a36767f98a5e92c875ec78dd2de858c4936c12a8",
+    "benchmarks/harbor/scripts/uv_release_cache.py": "993cc35b1f3a9e8e171b662ab6d4524c5fe059f0d1f0e275a5bb1930111e0ff5",
+    "benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/dependency-cache.compose.json": "fc2c6a22fb28563171f21808cd103851c69ce02bd3aa8c8c770eea20af87f49f",
+    "benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/dependency-release-manifest.json": "77dfda40e45b870a8ce7f07a1e058d416d94a04ab3d27c2f21bc370f22151394",
+    "benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/python-agent-release-manifest.json": "78aecd7c397c469196f9d246a160bf1fc99ad2d51cb71320b7d26ebf839647ce"
+  },
+  "budget_policy": "Native task agent timeout minus 180 seconds, separate 1800-second setup. Native verifier per-attempt time limits; max three attempts, retry only timeouts against the same agent output.",
+  "execution_policy": "Exactly one fresh model trial for each of the six prior failures. Ascending work budget then name, consecutive pairs with concurrency two. No whole-trial retries or outcome-based replacements.",
+  "failure_policy": "Keep all setup failures and verifier errors separate from scored failures. Retain native zeros even when assertions never start, but classify them as verifier errors in analysis.",
+  "preflight_policy": "Fresh install-only checks for all six tasks before model work. All six verifier dependency profiles are not_configured; installation success does not certify verifier dependencies. Full trials repeat setup in fresh containers.",
+  "information_policy": "Reuse the exact previous baseline agent, adapter, prompt, task revisions, resources and transport configuration. Do not inject previous failure diagnoses or solutions. Freeze all six outputs before reviewing new trajectories or official test contents. Prior task/test exposure is disclosed.",
+  "scope": "Targeted exploratory rerun of the six failures in an exposed development pool after the provider model update. A single trial per task cannot isolate a checkpoint effect or establish reliability.",
+  "provider_checkpoint": {
+    "api_model": "deepseek-flash",
+    "reported_checkpoint": "DeepSeek-V4.1-Flash-1017",
+    "verification": "Official DeepSeek documentation states that deepseek-flash now serves this checkpoint. The API alias is not an immutable checkpoint pin.",
+    "sources": [
+      "https://api-docs.deepseek.com/news/news261017/",
+      "https://api-docs.deepseek.com/quick_start/agent_integrations/"
+    ],
+    "checked_at": "2026-10-18T13:16:26.524186+00:00"
+  },
+  "historical_baseline": {
+    "plan": "benchmarks/harbor/configs/dev34-main-baseline-20261012/experiment-v5.json",
+    "plan_sha256": "f7bc7b329767f4b8007956462f6cdbd4407d386af493524df27ba313c0a4e4c3",
+    "runtime_source_ref": "2eb99b608dfb39464d17d109a4a8f11afda607a9",
+    "original_scores": "28/34 passed; selected six tasks all failed."
+  },
+  "protocol_amendment": {
+    "prior_plan": "experiment.json",
+    "prior_plan_sha256": "9383b05eeca74a5161ac41c05c80e2311ea3f93d4d87a70185b33c55c3587d4f",
+    "reason": "The agent installer selected the original CPython 3.13.9+20251120 archive, absent from the old cache. Slow uncached GitHub transfers prompted a pre-model transport amendment. Cache the identical archive with original release metadata, filename, size and SHA256 verification.",
+    "archive_size": 34257108,
+    "archive_sha256": "1bc9a94ab2425fb6ae58e45f69cab61647759b57895649701390822d976fa923",
+    "model_runs_started_before_amendment": 0,
+    "original_agent_adapter_and_task_sources_unchanged": true,
+    "versions_package_sources_official_tests_resources_budgets_unchanged": true,
+    "interrupted_installation_evidence": "jobs/flash1017-preflight-interruption.json",
+    "original_preflight_job": "flash1017-install-only-20261018",
+    "new_preflight_policy": "Run all six installation-only checks afresh with the amended cache manifest; retain interrupted attempts separately."
+  }
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/experiment.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/experiment.json
new file mode 100644
index 0000000..fa0c616
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/experiment.json
@@ -0,0 +1,272 @@
+{
+  "schema_version": 1,
+  "dataset_ref": "sha256:7d7bdc1cbedad549fc1140404bd4dc45e5fd0ea7c4186773687d177ad3a0699a",
+  "model": "deepseek/deepseek-flash",
+  "endpoint": "https://api.deepseek.com/anthropic",
+  "arms": {
+    "main": {
+      "source_root": "jobs/dev34-baseline-main-source",
+      "source_ref": "2eb99b608dfb39464d17d109a4a8f11afda607a9",
+      "prompt_file": "main-prompt.txt",
+      "prompt_sha256": "37c4d9d44640c0051dc5a20e5954a067b98e8c51be07a48f080aa059e725c55d",
+      "prompt_words": 385
+    }
+  },
+  "max_concurrent_trials": 2,
+  "whole_trial_retries": 0,
+  "registered_at": "2026-10-18T13:16:26.523948+00:00",
+  "experiment": "deepseek-v41-flash-1017-rerun-20261018",
+  "selection_file": "selection.json",
+  "selection_sha256": "04c7a99606bc58a15ea6577548fdc81f12f8c6ef2770127743051ef2c4fd9bb7",
+  "total_model_runs": 6,
+  "tasks": [
+    {
+      "name": "terminal-bench/adaptive-rejection-sampler",
+      "ref": "sha256:bcaa2399985cd57666018025846289ab25e193ae0dd8fb7f0ffab2410c24d4de",
+      "native_agent_timeout_seconds": 900,
+      "work_budget_seconds": 720,
+      "verifier_timeout_seconds": 900.0,
+      "docker_image": "alexgshaw/adaptive-rejection-sampler:20251031",
+      "resources": {
+        "cpus": 1,
+        "memory_mb": 2048,
+        "storage_mb": 10240,
+        "gpus": 0
+      },
+      "category": "scientific-computing",
+      "verifier_dependency_profile": "not_configured",
+      "task_file_sha256": {
+        "README.md": "95408217b9c484ffc18a02a91b7706323b6902919c72b24866a32e27a23543e2",
+        "environment/Dockerfile": "e644a3343e339303230f1441b7383f076d21c69a5a7aa6430b4f8a9d27f15c8b",
+        "environment/protected.tar.gz.enc": "7fc9bf1347bb97da8dede04b2c58638f225070913e222d76627068aebd1b1df2",
+        "instruction.md": "e6d28c28acd76bef61817427c74c2a29e3e8d75f535cbd708afcbfd02c54013f",
+        "solution/solve.sh": "fe443a8082b5183ba1623311ba0ee958888aefa633eff93766b9ae14e9bdd84d",
+        "task.toml": "6ed7b37f74beb46e9d47b0f5fad53025667c9548baff4e96d467b56d429fab1a",
+        "tests/test.sh": "38b43560d173cc2b952c3a3e17b8a480216d84e33450515e047bcb0d806b1e0a",
+        "tests/test_outputs.py": "547dc6e107f034f41703722aeceb6d0236e3fb69116fc3f2fbaba11884de352f"
+      }
+    },
+    {
+      "name": "terminal-bench/db-wal-recovery",
+      "ref": "sha256:07e9193513bfce038b9f9d84962ecc85683eee0a642d14ee6fb7104b086ddab4",
+      "native_agent_timeout_seconds": 900,
+      "work_budget_seconds": 720,
+      "verifier_timeout_seconds": 900.0,
+      "docker_image": "alexgshaw/db-wal-recovery:20251031",
+      "resources": {
+        "cpus": 1,
+        "memory_mb": 2048,
+        "storage_mb": 10240,
+        "gpus": 0
+      },
+      "category": "file-operations",
+      "verifier_dependency_profile": "not_configured",
+      "task_file_sha256": {
+        "README.md": "600762031a1a02ecc3b46d70f7683f3ff930b47c8a3d0e27cc31b146158c3126",
+        "environment/Dockerfile": "63a260531c62f4ae72060db17f295c0aa03cfc6321f8162c324b7aa6b3f81302",
+        "environment/main.db": "68cb288eb29e76d814b4e7ee6887b37c415675e644f88173bc191348e290d119",
+        "environment/main.db-wal.encrypted": "68753bdce26196a53194260d001544cc6bb8f9f38e2f2f6826d09798e31e6e26",
+        "instruction.md": "5a20375162921b8911c67035d1905df7fd335d1bed03044271e608b2c3c93b26",
+        "solution/solve.sh": "ed8530c94103bf19783a1d41bc63fe4db478cdd9629f22fee25c45a45699f5e8",
+        "task.toml": "1cdc2c615e259bc10265efbe71b518332070c1c002e226566399e9788f18d0aa",
+        "tests/test.sh": "4770437ea96c3cc84684b4f99d55fb148fcac09f9ea1e8ef49de487716e6c334",
+        "tests/test_outputs.py": "88dc5c2a3abba75bca0e0f9136223ae5bb585d5984b8d9eba51a7829c1d2ba49"
+      }
+    },
+    {
+      "name": "terminal-bench/kv-store-grpc",
+      "ref": "sha256:973c5d4c111fb61a344457936f1c36400acd2d9e44389e7b319586fe23a7a307",
+      "native_agent_timeout_seconds": 900,
+      "work_budget_seconds": 720,
+      "verifier_timeout_seconds": 900.0,
+      "docker_image": "alexgshaw/kv-store-grpc:20251031",
+      "resources": {
+        "cpus": 1,
+        "memory_mb": 2048,
+        "storage_mb": 10240,
+        "gpus": 0
+      },
+      "category": "software-engineering",
+      "verifier_dependency_profile": "not_configured",
+      "task_file_sha256": {
+        "README.md": "51fdc2ad06f0adb1270164f247ea41799b2ee2ad15490e48ddbf5e79d3767aaa",
+        "environment/Dockerfile": "47c9982789092437f950736110b2bd18899c39b8fd20f3af3c95b848e564a197",
+        "instruction.md": "f4b3facc199d5b1c78db6e603ffa29ec1b93a77ff7d291e1a92b42e6bf582bed",
+        "solution/solve.sh": "6bee2938ac9731d1156b0ca02b28c3d5793bf2cbe5f995c863d397093af01cb7",
+        "task.toml": "6ff0f30a526bfff86a67b0b53439b9965229c0f368167dc1a5d05f34d82299a3",
+        "tests/test.sh": "c348467a076bf614d0a850e25b2263f85bbf7b4c7a8d344fbd5454a0611a0df8",
+        "tests/test_outputs.py": "115de2e964bbe689d07930527113422ef8a9e4be4eb958ad84594e3fcd2b1b0e"
+      }
+    },
+    {
+      "name": "terminal-bench/query-optimize",
+      "ref": "sha256:169496ea6843cb403b0860132675baf7aa0df0ac8b86221068e27d86395260f4",
+      "native_agent_timeout_seconds": 900,
+      "work_budget_seconds": 720,
+      "verifier_timeout_seconds": 1800.0,
+      "docker_image": "alexgshaw/query-optimize:20251031",
+      "resources": {
+        "cpus": 1,
+        "memory_mb": 2048,
+        "storage_mb": 10240,
+        "gpus": 0
+      },
+      "category": "data-science",
+      "verifier_dependency_profile": "not_configured",
+      "task_file_sha256": {
+        "README.md": "a82c53e333a0ac0b0b410b5e26cb8e7305dece31db84d340248ffb13d5a45160",
+        "environment/Dockerfile": "064818771f200e0517d22068f87229a7a95d0a82b0778c84ca7203bf1926a0b8",
+        "environment/task-deps/my-sql-query.sql": "43382792f0f438625d643e34abe144febd5669518f4ab1ab550a15c6d9314a21",
+        "instruction.md": "473e12d320d2732f679fb631c2a4dab4a01deea6775686718ae0c417b8a39447",
+        "solution/solve.sh": "0d95f3cedb287ff9d47c796cbaf812c300d1222c10fa80c49e8acb0e81eb6978",
+        "task.toml": "6e88f2a86df699594bcb1bf81df4ed694505f79c9420f671335b02524bbe3868",
+        "tests/golden.sql": "4da1885d56b3ffa7b7b22b422996d018795611f6963499d551582881b74b84fd",
+        "tests/test.sh": "4770437ea96c3cc84684b4f99d55fb148fcac09f9ea1e8ef49de487716e6c334",
+        "tests/test_outputs.py": "2068ae80c57af8d52a7ae510f0666732c057806e556b480970f36f14bae683f6"
+      }
+    },
+    {
+      "name": "terminal-bench/torch-pipeline-parallelism",
+      "ref": "sha256:db605337c749a872cea7b5b413429b3915bb4c3efe0f7875f0c46ce81bd8c4fb",
+      "native_agent_timeout_seconds": 900,
+      "work_budget_seconds": 720,
+      "verifier_timeout_seconds": 900.0,
+      "docker_image": "alexgshaw/torch-pipeline-parallelism:20251031",
+      "resources": {
+        "cpus": 1,
+        "memory_mb": 8192,
+        "storage_mb": 10240,
+        "gpus": 0
+      },
+      "category": "software-engineering",
+      "verifier_dependency_profile": "not_configured",
+      "task_file_sha256": {
+        "README.md": "297ddea07c230dfd246a8ab07337dd7abcd8e5e2a2b2e26590cb0a8af779839c",
+        "environment/Dockerfile": "c357ea41e4a4013c5c5980b8fef04af6a5ae250ccc6697cf3bad3127401a931f",
+        "instruction.md": "989bfcf0c54baf6c8cdc391743dbb2f9e858c5779dd85112536875d4d1b2dcd2",
+        "solution/solve.sh": "787198e4178d88b887697aefdd29819f29435ffdfe356545d92f58a30fb38b70",
+        "task.toml": "8b4f9f525c08afbbc4eca9168e46238a61ccb903486773e9ca4da64bb7ffe857",
+        "tests/test.sh": "cc9fd28b6590993fc37e4a99da771cbf04726c8c62fe308e802612241da69dea",
+        "tests/test_outputs.py": "23574246a407163a8e6581c84287cc2f1eb745efbaad2a317a5a28c39958c64f"
+      }
+    },
+    {
+      "name": "terminal-bench/regex-chess",
+      "ref": "sha256:e763e0ac1c9759081af0a4a82ba51b8cf9ae5485a93de3bbe42d7d344597bd78",
+      "native_agent_timeout_seconds": 3600,
+      "work_budget_seconds": 3420,
+      "verifier_timeout_seconds": 3600.0,
+      "docker_image": "alexgshaw/regex-chess:20251031",
+      "resources": {
+        "cpus": 1,
+        "memory_mb": 2048,
+        "storage_mb": 10240,
+        "gpus": 0
+      },
+      "category": "software-engineering",
+      "verifier_dependency_profile": "not_configured",
+      "task_file_sha256": {
+        "README.md": "9296cbee592fc996c58d228b51fee303647e5ecfae82d5666d42a40cb30ec355",
+        "environment/Dockerfile": "2616cd775252303839dfda0ee2f047eb862952b053b0103bd6b2460ef90c11b7",
+        "environment/check.py": "f9c416ae9ea01dcdd5f6a43170c8434b8c0a0b389a4a30876a2048e277d191c7",
+        "instruction.md": "a6ed3fcc54fc92d71a84af921942422d299ca89b410dcd92cd05a53d0d06436c",
+        "solution/solve.sh": "b728bf0ea392a3d244afbd0818d36933c8d3e190e47743c1382cd518d518a2e9",
+        "task.toml": "07fe52636fcc4bc9fff25d5630af145b97a5331219d719cc8decb4ca67ea21d4",
+        "tests/test.sh": "8f7f00a1c4dde7fcca6c918c768f2aba432b8c4007c15989716c68091950da87",
+        "tests/test_outputs.py": "4983c44d21c56dd8191381eff0304afc8b87bb714527ea9036b3319c015a30b4"
+      }
+    }
+  ],
+  "preflight": [
+    {
+      "arm": "main",
+      "config": "benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/install-only.json",
+      "config_sha256": "d431e06126191d5ce53d9015133ac00484fbd1876e87e065e4e4722b92072d67",
+      "job_name": "flash1017-install-only-20261018"
+    }
+  ],
+  "queue": [
+    {
+      "arm": "main",
+      "repeat": 1,
+      "task": "terminal-bench/adaptive-rejection-sampler",
+      "config": "benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/adaptive-rejection-sampler.json",
+      "job_name": "flash1017-adaptive-rejection-sampler-20261018",
+      "config_sha256": "0bf207305a6399d3cb34fa2641e4241b4ada0b64903aa7805d1d488aabf43fa0"
+    },
+    {
+      "arm": "main",
+      "repeat": 1,
+      "task": "terminal-bench/db-wal-recovery",
+      "config": "benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/db-wal-recovery.json",
+      "job_name": "flash1017-db-wal-recovery-20261018",
+      "config_sha256": "98d43e3daf212326518eb56b5c436a01effb82768c6ee49a581a65bebe28af8e"
+    },
+    {
+      "arm": "main",
+      "repeat": 1,
+      "task": "terminal-bench/kv-store-grpc",
+      "config": "benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/kv-store-grpc.json",
+      "job_name": "flash1017-kv-store-grpc-20261018",
+      "config_sha256": "ef2504381558a7df9baa22961ddf39e09e810e6c6982618cf3ae8e66c2a66691"
+    },
+    {
+      "arm": "main",
+      "repeat": 1,
+      "task": "terminal-bench/query-optimize",
+      "config": "benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/query-optimize.json",
+      "job_name": "flash1017-query-optimize-20261018",
+      "config_sha256": "2d836edcb1155bc43a4b0af29387cd4120c785e63d9c2359671235cc099b3719"
+    },
+    {
+      "arm": "main",
+      "repeat": 1,
+      "task": "terminal-bench/torch-pipeline-parallelism",
+      "config": "benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/torch-pipeline-parallelism.json",
+      "job_name": "flash1017-torch-pipeline-parallelism-20261018",
+      "config_sha256": "02f824e64249f601748039cb58f8837ae20fb90b063fa7f9a8614354801417df"
+    },
+    {
+      "arm": "main",
+      "repeat": 1,
+      "task": "terminal-bench/regex-chess",
+      "config": "benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/regex-chess.json",
+      "job_name": "flash1017-regex-chess-20261018",
+      "config_sha256": "3a9ed5db6443cc1a5e77dd417083db59ab9ed49cc24ac4e43778dee51094a4aa"
+    }
+  ],
+  "log_dir": "jobs/deepseek-v41-flash-1017-rerun-20261018-orchestration",
+  "image_digests_file": "image-digests.json",
+  "image_digests_sha256": "1575f2ddb2e1d5cf088a01af5b43a5ed24271ea0bc6bed726c8f3a3a04fc712c",
+  "additional_file_sha256": {
+    "benchmarks/harbor/configs/dev34-main-baseline-20261012/dependency-release-manifest-v5.json": "bdf3d3ef8b7743700c5294c4a511c686a3928c178a77fd4abcba90ee4557ffcc",
+    "benchmarks/harbor/configs/dev34-main-baseline-20261012/transport-v3/apt-cache.conf": "d0e1ffb6dbd02c54978ba2df386326311a59f6d3774990180f22bcc6adfaedbc",
+    "benchmarks/harbor/configs/dev34-main-baseline-20261012/transport-v5/dependency-cache.compose.json": "60d778603976454bb407b8273626851357ee41489cdb0395b06dcc7aab9d151e",
+    "benchmarks/harbor/scripts/apt_index_proxy.py": "bc15f7adf1d7968d627ceca4683ab86742a7188a1b1955ba1fef37ba48e8a936",
+    "benchmarks/harbor/scripts/cache_image.py": "639709c03a7b0c3eb6aa78b4e87e6184b49b4b102853a31e33dd832d759180c0",
+    "benchmarks/harbor/scripts/run_comparison.py": "04f2106f0099a7b4d77e9206a36767f98a5e92c875ec78dd2de858c4936c12a8",
+    "benchmarks/harbor/scripts/uv_release_cache.py": "993cc35b1f3a9e8e171b662ab6d4524c5fe059f0d1f0e275a5bb1930111e0ff5"
+  },
+  "budget_policy": "Native task agent timeout minus 180 seconds, separate 1800-second setup. Native verifier per-attempt time limits; max three attempts, retry only timeouts against the same agent output.",
+  "execution_policy": "Exactly one fresh model trial for each of the six prior failures. Ascending work budget then name, consecutive pairs with concurrency two. No whole-trial retries or outcome-based replacements.",
+  "failure_policy": "Keep all setup failures and verifier errors separate from scored failures. Retain native zeros even when assertions never start, but classify them as verifier errors in analysis.",
+  "preflight_policy": "Fresh install-only checks for all six tasks before model work. All six verifier dependency profiles are not_configured; installation success does not certify verifier dependencies. Full trials repeat setup in fresh containers.",
+  "information_policy": "Reuse the exact previous baseline agent, adapter, prompt, task revisions, resources and transport configuration. Do not inject previous failure diagnoses or solutions. Freeze all six outputs before reviewing new trajectories or official test contents. Prior task/test exposure is disclosed.",
+  "scope": "Targeted exploratory rerun of the six failures in an exposed development pool after the provider model update. A single trial per task cannot isolate a checkpoint effect or establish reliability.",
+  "provider_checkpoint": {
+    "api_model": "deepseek-flash",
+    "reported_checkpoint": "DeepSeek-V4.1-Flash-1017",
+    "verification": "Official DeepSeek documentation states that deepseek-flash now serves this checkpoint. The API alias is not an immutable checkpoint pin.",
+    "sources": [
+      "https://api-docs.deepseek.com/news/news261017/",
+      "https://api-docs.deepseek.com/quick_start/agent_integrations/"
+    ],
+    "checked_at": "2026-10-18T13:16:26.524186+00:00"
+  },
+  "historical_baseline": {
+    "plan": "benchmarks/harbor/configs/dev34-main-baseline-20261012/experiment-v5.json",
+    "plan_sha256": "f7bc7b329767f4b8007956462f6cdbd4407d386af493524df27ba313c0a4e4c3",
+    "runtime_source_ref": "2eb99b608dfb39464d17d109a4a8f11afda607a9",
+    "original_scores": "28/34 passed; selected six tasks all failed."
+  }
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/image-digests.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/image-digests.json
new file mode 100644
index 0000000..9289aca
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/image-digests.json
@@ -0,0 +1,60 @@
+{
+  "observed_at": "2026-10-18T13:16:26.429384+00:00",
+  "all_match_previous_baseline": true,
+  "images": [
+    {
+      "task": "terminal-bench/kv-store-grpc",
+      "image": "alexgshaw/kv-store-grpc:20251031",
+      "image_id": "sha256:3399400800dcb207634daa42bc1b052e831e285cc9d221eea66c47bc0fc79791",
+      "repo_digests": [
+        "alexgshaw/kv-store-grpc@sha256:3399400800dcb207634daa42bc1b052e831e285cc9d221eea66c47bc0fc79791"
+      ],
+      "size_bytes": 44171671
+    },
+    {
+      "task": "terminal-bench/regex-chess",
+      "image": "alexgshaw/regex-chess:20251031",
+      "image_id": "sha256:e3b6d61d2d3930fbbdc5f7bf091a7f8b7f6125e1951ef8a7dbe165b1ef62942c",
+      "repo_digests": [
+        "alexgshaw/regex-chess@sha256:e3b6d61d2d3930fbbdc5f7bf091a7f8b7f6125e1951ef8a7dbe165b1ef62942c"
+      ],
+      "size_bytes": 55625584
+    },
+    {
+      "task": "terminal-bench/torch-pipeline-parallelism",
+      "image": "alexgshaw/torch-pipeline-parallelism:20251031",
+      "image_id": "sha256:3cb7b39d86f1704b20e668851da0b08b34b3270c01d583f136d65190984a8d1b",
+      "repo_digests": [
+        "alexgshaw/torch-pipeline-parallelism@sha256:3cb7b39d86f1704b20e668851da0b08b34b3270c01d583f136d65190984a8d1b"
+      ],
+      "size_bytes": 29725531
+    },
+    {
+      "task": "terminal-bench/db-wal-recovery",
+      "image": "alexgshaw/db-wal-recovery:20251031",
+      "image_id": "sha256:0e33ea5ec823975d1bd6c3778395c9f94251dd88f571146057bff6adb7e4594e",
+      "repo_digests": [
+        "alexgshaw/db-wal-recovery@sha256:0e33ea5ec823975d1bd6c3778395c9f94251dd88f571146057bff6adb7e4594e"
+      ],
+      "size_bytes": 175276145
+    },
+    {
+      "task": "terminal-bench/query-optimize",
+      "image": "alexgshaw/query-optimize:20251031",
+      "image_id": "sha256:28de7da87004eab5b29bd5eaf1823579bc5b73ccea4b80349a5623185770da14",
+      "repo_digests": [
+        "alexgshaw/query-optimize@sha256:28de7da87004eab5b29bd5eaf1823579bc5b73ccea4b80349a5623185770da14"
+      ],
+      "size_bytes": 120494963
+    },
+    {
+      "task": "terminal-bench/adaptive-rejection-sampler",
+      "image": "alexgshaw/adaptive-rejection-sampler:20251031",
+      "image_id": "sha256:985bed39e8448bc95f138b5f766b59003a95c4be171c96af8e76448338f3dd8f",
+      "repo_digests": [
+        "alexgshaw/adaptive-rejection-sampler@sha256:985bed39e8448bc95f138b5f766b59003a95c4be171c96af8e76448338f3dd8f"
+      ],
+      "size_bytes": 29731756
+    }
+  ]
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/install-only.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/install-only.json
new file mode 100644
index 0000000..0dfd408
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/install-only.json
@@ -0,0 +1,49 @@
+{
+  "n_concurrent_trials": 2,
+  "n_attempts": 1,
+  "agents": [
+    {
+      "name": "harbor_adapter:NanoPyCodeAgent",
+      "model_name": "deepseek/deepseek-flash",
+      "override_setup_timeout_sec": 1800,
+      "kwargs": {
+        "max_turns": 100,
+        "max_tokens": 65536,
+        "time_budget_seconds": 720
+      }
+    }
+  ],
+  "tasks": [
+    {
+      "name": "terminal-bench/adaptive-rejection-sampler",
+      "ref": "sha256:bcaa2399985cd57666018025846289ab25e193ae0dd8fb7f0ffab2410c24d4de"
+    },
+    {
+      "name": "terminal-bench/db-wal-recovery",
+      "ref": "sha256:07e9193513bfce038b9f9d84962ecc85683eee0a642d14ee6fb7104b086ddab4"
+    },
+    {
+      "name": "terminal-bench/kv-store-grpc",
+      "ref": "sha256:973c5d4c111fb61a344457936f1c36400acd2d9e44389e7b319586fe23a7a307"
+    },
+    {
+      "name": "terminal-bench/query-optimize",
+      "ref": "sha256:169496ea6843cb403b0860132675baf7aa0df0ac8b86221068e27d86395260f4"
+    },
+    {
+      "name": "terminal-bench/torch-pipeline-parallelism",
+      "ref": "sha256:db605337c749a872cea7b5b413429b3915bb4c3efe0f7875f0c46ce81bd8c4fb"
+    },
+    {
+      "name": "terminal-bench/regex-chess",
+      "ref": "sha256:e763e0ac1c9759081af0a4a82ba51b8cf9ae5485a93de3bbe42d7d344597bd78"
+    }
+  ],
+  "environment": {
+    "type": "docker",
+    "extra_docker_compose": [
+      "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/configs/dev34-main-baseline-20261012/transport-v5/dependency-cache.compose.json"
+    ]
+  },
+  "install_only": true
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/kv-store-grpc.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/kv-store-grpc.json
new file mode 100644
index 0000000..eadc233
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/kv-store-grpc.json
@@ -0,0 +1,28 @@
+{
+  "n_concurrent_trials": 1,
+  "n_attempts": 1,
+  "agents": [
+    {
+      "name": "harbor_adapter:NanoPyCodeAgent",
+      "model_name": "deepseek/deepseek-flash",
+      "override_setup_timeout_sec": 1800,
+      "kwargs": {
+        "max_turns": 100,
+        "max_tokens": 65536,
+        "time_budget_seconds": 720
+      }
+    }
+  ],
+  "tasks": [
+    {
+      "name": "terminal-bench/kv-store-grpc",
+      "ref": "sha256:973c5d4c111fb61a344457936f1c36400acd2d9e44389e7b319586fe23a7a307"
+    }
+  ],
+  "environment": {
+    "type": "docker",
+    "extra_docker_compose": [
+      "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/configs/dev34-main-baseline-20261012/transport-v5/dependency-cache.compose.json"
+    ]
+  }
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/main-prompt.txt b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/main-prompt.txt
new file mode 100644
index 0000000..56116f5
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/main-prompt.txt
@@ -0,0 +1,9 @@
+You are nanoPyCodeAgent, a coding agent running non-interactively on a task handed to you up front. There is no user to reply to you: never ask a clarifying question, never stop to wait for confirmation, and never present a plan for approval — decide on your own and carry it out. Work the task through to the end, then check the result with the tools instead of assuming it worked. When it is done, answer with a short summary and no further tool calls: that reply is what ends the run. Runtime budget reminders report remaining time and model replies; when either is running low, prioritize saving the required deliverables, necessary verification, and a final summary. Do not start optional work after the requirements are satisfied.
+
+Identify the required deliverables, interfaces, and constraints. Treat examples as evidence, not as the full contract unless the task says so. Inspect relevant files, dependencies, and existing checks instead of guessing. Task-supplied workspace files, including source, executables, and tests, are available references unless the task restricts their use. Consult general documentation as needed. Do not seek or use task-specific reference solutions or hidden evaluation tests from outside the supplied materials unless the task explicitly requests them. Derive checks from the stated requirements.
+
+Verify the requested behavior through the real entry point when feasible. Check the content and meaning of outputs, not only command success or file existence. For logic or numerical work, choose a few representative and boundary cases justified by the task, using an independent reference, invariant, or consistency check when available. Do not derive every expectation from the implementation being tested or weaken requirements to make a check pass. Keep verification proportional to the task and remaining budget.
+
+When a check fails, inspect the error and relevant output, revise the underlying assumption or implementation, and rerun the affected check. Keep the task's constraints intact. Once appropriate checks support the required behavior, save the deliverables and finish. State what was verified and any remaining limitation without claiming unobserved success.
+
+Prefer the read tool for viewing files, the edit tool for changing part of an existing file, and the write tool for creating files or rewriting them whole. Use the bash tool to run commands, search with grep, and complete tasks that need real command output instead of guessing.
\ No newline at end of file
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/python-agent-release-manifest.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/python-agent-release-manifest.json
new file mode 100644
index 0000000..2a47077
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/python-agent-release-manifest.json
@@ -0,0 +1,16 @@
+{
+  "registered_at": "2026-10-18T13:25:07.185006+00:00",
+  "metadata_source": "Original public GitHub release API, fetched before this registration",
+  "assets": [
+    {
+      "version": "20251120",
+      "filename": "cpython-3.13.9+20251120-x86_64-unknown-linux-gnu-install_only_stripped.tar.gz",
+      "url": "https://github.com/astral-sh/python-build-standalone/releases/download/20251120/cpython-3.13.9%2B20251120-x86_64-unknown-linux-gnu-install_only_stripped.tar.gz",
+      "digest": "sha256:1bc9a94ab2425fb6ae58e45f69cab61647759b57895649701390822d976fa923",
+      "size": 34257108,
+      "asset_id": 318881608,
+      "release_metadata_sha256": "705690ab2fd466e90b4c868635c529f3c7fd0cb597d399e350f399e3f3e35e6c",
+      "repository": "astral-sh/python-build-standalone"
+    }
+  ]
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/query-optimize.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/query-optimize.json
new file mode 100644
index 0000000..95dc238
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/query-optimize.json
@@ -0,0 +1,28 @@
+{
+  "n_concurrent_trials": 1,
+  "n_attempts": 1,
+  "agents": [
+    {
+      "name": "harbor_adapter:NanoPyCodeAgent",
+      "model_name": "deepseek/deepseek-flash",
+      "override_setup_timeout_sec": 1800,
+      "kwargs": {
+        "max_turns": 100,
+        "max_tokens": 65536,
+        "time_budget_seconds": 720
+      }
+    }
+  ],
+  "tasks": [
+    {
+      "name": "terminal-bench/query-optimize",
+      "ref": "sha256:169496ea6843cb403b0860132675baf7aa0df0ac8b86221068e27d86395260f4"
+    }
+  ],
+  "environment": {
+    "type": "docker",
+    "extra_docker_compose": [
+      "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/configs/dev34-main-baseline-20261012/transport-v5/dependency-cache.compose.json"
+    ]
+  }
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/regex-chess.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/regex-chess.json
new file mode 100644
index 0000000..3bef66e
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/regex-chess.json
@@ -0,0 +1,28 @@
+{
+  "n_concurrent_trials": 1,
+  "n_attempts": 1,
+  "agents": [
+    {
+      "name": "harbor_adapter:NanoPyCodeAgent",
+      "model_name": "deepseek/deepseek-flash",
+      "override_setup_timeout_sec": 1800,
+      "kwargs": {
+        "max_turns": 100,
+        "max_tokens": 65536,
+        "time_budget_seconds": 3420
+      }
+    }
+  ],
+  "tasks": [
+    {
+      "name": "terminal-bench/regex-chess",
+      "ref": "sha256:e763e0ac1c9759081af0a4a82ba51b8cf9ae5485a93de3bbe42d7d344597bd78"
+    }
+  ],
+  "environment": {
+    "type": "docker",
+    "extra_docker_compose": [
+      "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/configs/dev34-main-baseline-20261012/transport-v5/dependency-cache.compose.json"
+    ]
+  }
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/selection.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/selection.json
new file mode 100644
index 0000000..ebb1fc2
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/selection.json
@@ -0,0 +1,49 @@
+{
+  "source_results": "benchmarks/harbor/results/dev34-main-baseline-20261012.json",
+  "source_results_sha256": "90873385a2277776f09a37c77db390432d223b28461a892f564d24ab4f9407c3",
+  "rule": "All six raw-zero trials from the latest completed 34-task baseline; no task substitutions.",
+  "previous_trials": [
+    {
+      "task": "terminal-bench/adaptive-rejection-sampler",
+      "trial": "adaptive-rejection-sampler__zWw9zAU",
+      "job": "dev34-adaptive-rejection-sampler-main-20261012",
+      "raw_reward": 0.0,
+      "classification": "scored_failure"
+    },
+    {
+      "task": "terminal-bench/db-wal-recovery",
+      "trial": "db-wal-recovery__BCqJJMc",
+      "job": "dev34-db-wal-recovery-main-20261012",
+      "raw_reward": 0.0,
+      "classification": "scored_failure"
+    },
+    {
+      "task": "terminal-bench/kv-store-grpc",
+      "trial": "kv-store-grpc__pUAoWuK",
+      "job": "dev34-kv-store-grpc-main-20261012",
+      "raw_reward": 0.0,
+      "classification": "scored_failure"
+    },
+    {
+      "task": "terminal-bench/query-optimize",
+      "trial": "query-optimize__ZVLSEuZ",
+      "job": "dev34-query-optimize-main-20261012",
+      "raw_reward": 0.0,
+      "classification": "scored_failure"
+    },
+    {
+      "task": "terminal-bench/torch-pipeline-parallelism",
+      "trial": "torch-pipeline-parallelism__Zr8yYNi",
+      "job": "dev34-torch-pipeline-parallelism-main-20261012",
+      "raw_reward": 0.0,
+      "classification": "scored_failure"
+    },
+    {
+      "task": "terminal-bench/regex-chess",
+      "trial": "regex-chess__g85Ci69",
+      "job": "dev34-regex-chess-main-20261012",
+      "raw_reward": 0.0,
+      "classification": "scored_failure"
+    }
+  ]
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/torch-pipeline-parallelism.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/torch-pipeline-parallelism.json
new file mode 100644
index 0000000..797f01d
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/torch-pipeline-parallelism.json
@@ -0,0 +1,28 @@
+{
+  "n_concurrent_trials": 1,
+  "n_attempts": 1,
+  "agents": [
+    {
+      "name": "harbor_adapter:NanoPyCodeAgent",
+      "model_name": "deepseek/deepseek-flash",
+      "override_setup_timeout_sec": 1800,
+      "kwargs": {
+        "max_turns": 100,
+        "max_tokens": 65536,
+        "time_budget_seconds": 720
+      }
+    }
+  ],
+  "tasks": [
+    {
+      "name": "terminal-bench/torch-pipeline-parallelism",
+      "ref": "sha256:db605337c749a872cea7b5b413429b3915bb4c3efe0f7875f0c46ce81bd8c4fb"
+    }
+  ],
+  "environment": {
+    "type": "docker",
+    "extra_docker_compose": [
+      "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/configs/dev34-main-baseline-20261012/transport-v5/dependency-cache.compose.json"
+    ]
+  }
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/adaptive-rejection-sampler.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/adaptive-rejection-sampler.json
new file mode 100644
index 0000000..e1db84d
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/adaptive-rejection-sampler.json
@@ -0,0 +1,28 @@
+{
+  "n_concurrent_trials": 1,
+  "n_attempts": 1,
+  "agents": [
+    {
+      "name": "harbor_adapter:NanoPyCodeAgent",
+      "model_name": "deepseek/deepseek-flash",
+      "override_setup_timeout_sec": 1800,
+      "kwargs": {
+        "max_turns": 100,
+        "max_tokens": 65536,
+        "time_budget_seconds": 720
+      }
+    }
+  ],
+  "tasks": [
+    {
+      "name": "terminal-bench/adaptive-rejection-sampler",
+      "ref": "sha256:bcaa2399985cd57666018025846289ab25e193ae0dd8fb7f0ffab2410c24d4de"
+    }
+  ],
+  "environment": {
+    "type": "docker",
+    "extra_docker_compose": [
+      "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/dependency-cache.compose.json"
+    ]
+  }
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/db-wal-recovery.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/db-wal-recovery.json
new file mode 100644
index 0000000..324a467
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/db-wal-recovery.json
@@ -0,0 +1,28 @@
+{
+  "n_concurrent_trials": 1,
+  "n_attempts": 1,
+  "agents": [
+    {
+      "name": "harbor_adapter:NanoPyCodeAgent",
+      "model_name": "deepseek/deepseek-flash",
+      "override_setup_timeout_sec": 1800,
+      "kwargs": {
+        "max_turns": 100,
+        "max_tokens": 65536,
+        "time_budget_seconds": 720
+      }
+    }
+  ],
+  "tasks": [
+    {
+      "name": "terminal-bench/db-wal-recovery",
+      "ref": "sha256:07e9193513bfce038b9f9d84962ecc85683eee0a642d14ee6fb7104b086ddab4"
+    }
+  ],
+  "environment": {
+    "type": "docker",
+    "extra_docker_compose": [
+      "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/dependency-cache.compose.json"
+    ]
+  }
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/dependency-cache.compose.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/dependency-cache.compose.json
new file mode 100644
index 0000000..5f754f9
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/dependency-cache.compose.json
@@ -0,0 +1,100 @@
+{
+  "services": {
+    "main": {
+      "volumes": [
+        {
+          "type": "bind",
+          "source": "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/configs/dev34-main-baseline-20261012/transport-v3/apt-cache.conf",
+          "target": "/etc/apt/apt.conf.d/99nanopy-index-cache",
+          "read_only": true
+        }
+      ],
+      "environment": {
+        "UV_INSTALLER_GITHUB_BASE_URL": "http://127.0.0.1:18743",
+        "UV_PYTHON_INSTALL_MIRROR": "http://127.0.0.1:18743/astral-sh/python-build-standalone/releases/download"
+      }
+    },
+    "apt-index-cache": {
+      "image": "alexgshaw/pytorch-model-cli@sha256:c996884b5fb5fb2c0c07d67da2c4f974f271f07234ed28e1a868240375c57a67",
+      "network_mode": "service:main",
+      "entrypoint": [
+        "python3",
+        "/proxy/apt_index_proxy.py",
+        "--bind",
+        "127.0.0.1",
+        "--port",
+        "18742",
+        "--cache",
+        "/cache",
+        "--workers",
+        "16"
+      ],
+      "user": "1000:1000",
+      "read_only": true,
+      "cap_drop": [
+        "ALL"
+      ],
+      "security_opt": [
+        "no-new-privileges:true"
+      ],
+      "mem_limit": "256m",
+      "cpus": 0.5,
+      "volumes": [
+        "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/scripts/apt_index_proxy.py:/proxy/apt_index_proxy.py:ro",
+        "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/scripts/cache_image.py:/proxy/cache_image.py:ro",
+        "/home/minix/projects/nanoPyCodeAgent/.cache/nanopy-harbor/apt-indexes:/cache"
+      ],
+      "healthcheck": {
+        "test": [
+          "CMD",
+          "python3",
+          "-c",
+          "import socket; socket.create_connection((\"127.0.0.1\",18742), timeout=2).close()"
+        ],
+        "interval": "1s",
+        "timeout": "3s",
+        "retries": 30
+      }
+    },
+    "uv-release-cache": {
+      "image": "alexgshaw/pytorch-model-cli@sha256:c996884b5fb5fb2c0c07d67da2c4f974f271f07234ed28e1a868240375c57a67",
+      "network_mode": "service:main",
+      "entrypoint": [
+        "python3",
+        "/proxy/uv_release_cache.py",
+        "serve",
+        "--cache",
+        "/cache",
+        "--manifest",
+        "/manifest.json"
+      ],
+      "user": "1000:1000",
+      "read_only": true,
+      "cap_drop": [
+        "ALL"
+      ],
+      "security_opt": [
+        "no-new-privileges:true"
+      ],
+      "cpus": 0.5,
+      "mem_limit": "256m",
+      "volumes": [
+        "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/scripts/uv_release_cache.py:/proxy/uv_release_cache.py:ro",
+        "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/scripts/cache_image.py:/proxy/cache_image.py:ro",
+        "/home/minix/projects/nanoPyCodeAgent/.cache/nanopy-harbor/uv-releases:/cache:ro",
+        "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/dependency-release-manifest.json:/manifest.json:ro"
+      ],
+      "healthcheck": {
+        "test": [
+          "CMD",
+          "python3",
+          "-c",
+          "import socket; socket.create_connection((\"127.0.0.1\",18743), timeout=2).close()"
+        ],
+        "interval": "1s",
+        "timeout": "3s",
+        "retries": 30
+      }
+    }
+  }
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/install-only.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/install-only.json
new file mode 100644
index 0000000..cb9ca03
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/install-only.json
@@ -0,0 +1,49 @@
+{
+  "n_concurrent_trials": 2,
+  "n_attempts": 1,
+  "agents": [
+    {
+      "name": "harbor_adapter:NanoPyCodeAgent",
+      "model_name": "deepseek/deepseek-flash",
+      "override_setup_timeout_sec": 1800,
+      "kwargs": {
+        "max_turns": 100,
+        "max_tokens": 65536,
+        "time_budget_seconds": 720
+      }
+    }
+  ],
+  "tasks": [
+    {
+      "name": "terminal-bench/adaptive-rejection-sampler",
+      "ref": "sha256:bcaa2399985cd57666018025846289ab25e193ae0dd8fb7f0ffab2410c24d4de"
+    },
+    {
+      "name": "terminal-bench/db-wal-recovery",
+      "ref": "sha256:07e9193513bfce038b9f9d84962ecc85683eee0a642d14ee6fb7104b086ddab4"
+    },
+    {
+      "name": "terminal-bench/kv-store-grpc",
+      "ref": "sha256:973c5d4c111fb61a344457936f1c36400acd2d9e44389e7b319586fe23a7a307"
+    },
+    {
+      "name": "terminal-bench/query-optimize",
+      "ref": "sha256:169496ea6843cb403b0860132675baf7aa0df0ac8b86221068e27d86395260f4"
+    },
+    {
+      "name": "terminal-bench/torch-pipeline-parallelism",
+      "ref": "sha256:db605337c749a872cea7b5b413429b3915bb4c3efe0f7875f0c46ce81bd8c4fb"
+    },
+    {
+      "name": "terminal-bench/regex-chess",
+      "ref": "sha256:e763e0ac1c9759081af0a4a82ba51b8cf9ae5485a93de3bbe42d7d344597bd78"
+    }
+  ],
+  "environment": {
+    "type": "docker",
+    "extra_docker_compose": [
+      "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/dependency-cache.compose.json"
+    ]
+  },
+  "install_only": true
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/kv-store-grpc.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/kv-store-grpc.json
new file mode 100644
index 0000000..b430b4a
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/kv-store-grpc.json
@@ -0,0 +1,28 @@
+{
+  "n_concurrent_trials": 1,
+  "n_attempts": 1,
+  "agents": [
+    {
+      "name": "harbor_adapter:NanoPyCodeAgent",
+      "model_name": "deepseek/deepseek-flash",
+      "override_setup_timeout_sec": 1800,
+      "kwargs": {
+        "max_turns": 100,
+        "max_tokens": 65536,
+        "time_budget_seconds": 720
+      }
+    }
+  ],
+  "tasks": [
+    {
+      "name": "terminal-bench/kv-store-grpc",
+      "ref": "sha256:973c5d4c111fb61a344457936f1c36400acd2d9e44389e7b319586fe23a7a307"
+    }
+  ],
+  "environment": {
+    "type": "docker",
+    "extra_docker_compose": [
+      "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/dependency-cache.compose.json"
+    ]
+  }
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/query-optimize.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/query-optimize.json
new file mode 100644
index 0000000..b23c304
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/query-optimize.json
@@ -0,0 +1,28 @@
+{
+  "n_concurrent_trials": 1,
+  "n_attempts": 1,
+  "agents": [
+    {
+      "name": "harbor_adapter:NanoPyCodeAgent",
+      "model_name": "deepseek/deepseek-flash",
+      "override_setup_timeout_sec": 1800,
+      "kwargs": {
+        "max_turns": 100,
+        "max_tokens": 65536,
+        "time_budget_seconds": 720
+      }
+    }
+  ],
+  "tasks": [
+    {
+      "name": "terminal-bench/query-optimize",
+      "ref": "sha256:169496ea6843cb403b0860132675baf7aa0df0ac8b86221068e27d86395260f4"
+    }
+  ],
+  "environment": {
+    "type": "docker",
+    "extra_docker_compose": [
+      "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/dependency-cache.compose.json"
+    ]
+  }
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/regex-chess.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/regex-chess.json
new file mode 100644
index 0000000..ca29596
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/regex-chess.json
@@ -0,0 +1,28 @@
+{
+  "n_concurrent_trials": 1,
+  "n_attempts": 1,
+  "agents": [
+    {
+      "name": "harbor_adapter:NanoPyCodeAgent",
+      "model_name": "deepseek/deepseek-flash",
+      "override_setup_timeout_sec": 1800,
+      "kwargs": {
+        "max_turns": 100,
+        "max_tokens": 65536,
+        "time_budget_seconds": 3420
+      }
+    }
+  ],
+  "tasks": [
+    {
+      "name": "terminal-bench/regex-chess",
+      "ref": "sha256:e763e0ac1c9759081af0a4a82ba51b8cf9ae5485a93de3bbe42d7d344597bd78"
+    }
+  ],
+  "environment": {
+    "type": "docker",
+    "extra_docker_compose": [
+      "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/dependency-cache.compose.json"
+    ]
+  }
+}
diff --git a/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/torch-pipeline-parallelism.json b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/torch-pipeline-parallelism.json
new file mode 100644
index 0000000..2be27bc
--- /dev/null
+++ b/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/torch-pipeline-parallelism.json
@@ -0,0 +1,28 @@
+{
+  "n_concurrent_trials": 1,
+  "n_attempts": 1,
+  "agents": [
+    {
+      "name": "harbor_adapter:NanoPyCodeAgent",
+      "model_name": "deepseek/deepseek-flash",
+      "override_setup_timeout_sec": 1800,
+      "kwargs": {
+        "max_turns": 100,
+        "max_tokens": 65536,
+        "time_budget_seconds": 720
+      }
+    }
+  ],
+  "tasks": [
+    {
+      "name": "terminal-bench/torch-pipeline-parallelism",
+      "ref": "sha256:db605337c749a872cea7b5b413429b3915bb4c3efe0f7875f0c46ce81bd8c4fb"
+    }
+  ],
+  "environment": {
+    "type": "docker",
+    "extra_docker_compose": [
+      "/home/minix/projects/nanoPyCodeAgent/benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/transport-v2/dependency-cache.compose.json"
+    ]
+  }
+}
diff --git a/benchmarks/harbor/reports/deepseek-v41-flash-1017-recovery-20261018.md b/benchmarks/harbor/reports/deepseek-v41-flash-1017-recovery-20261018.md
new file mode 100644
index 0000000..c545330
--- /dev/null
+++ b/benchmarks/harbor/reports/deepseek-v41-flash-1017-recovery-20261018.md
@@ -0,0 +1,97 @@
+# Flash 1017: authorized retry of the two unscored tasks
+
+**The sampler passed; pipeline parallelism failed.** Both now have valid
+official test results in this follow-up. The sampler completed on its first
+verification attempt. Pipeline verification timed out during dependency
+preparation on its first attempt, then completed on its second attempt with a
+valid failure score. No third attempt was needed or started.
+
+| Task | Official result | Tests passed / failed | Verifier attempts | Agent seconds |
+| --- | --- | ---: | --- | ---: |
+| adaptive-rejection-sampler | Passed, reward 1 | 9 / 0 | 1: completed | 591.4 |
+| torch-pipeline-parallelism | Failed, reward 0 | 2 / 2 | 1: timeout; 2: completed | 562.6 |
+
+The [initial six-trial run](deepseek-v41-flash-1017-rerun-20261018.md) left both
+tasks unscored because official dependencies failed to download. The user then
+authorized retrying these two tasks with a maximum of three attempts. The
+original containers had been deleted and their artifact manifests held no
+`/app` outputs, so this follow-up generated **one fresh model output per task**.
+It does not regrade or prove correctness of the original unscored submissions.
+Those results and frozen artifacts remain unchanged.
+
+The [registered plan](../configs/deepseek-v41-flash-1017-recovery-20261018/experiment.json)
+has SHA256
+`7ba0d4083cbc856a364417bc51025a8bd84434eb39319927664bfb4c018fba9d`.
+The [audited results](../results/deepseek-v41-flash-1017-recovery-20261018.json)
+retain trial IDs, all verifier attempts, actual pytest and native CTRF counts,
+source and wheel hashes, preserved-output hashes, authorization, and preflight
+and exposure findings. The
+[configuration README](../configs/deepseek-v41-flash-1017-recovery-20261018/README.md)
+documents execution and local evidence paths.
+
+Both fresh model trials used `deepseek/deepseek-flash` against the official
+DeepSeek endpoint. Agent Python sources are byte-identical to the initial
+evaluation, with the same 385-word prompt, pinned task revisions, original
+images, resources, and limits: 100 replies, 65,536 output tokens per reply,
+720 seconds of model work, and separate 1800-second setup. The evaluated
+checkout is frozen at `35fe1a2d6a786619a2adace1cfa9d0acca08d533`; its adapter
+adds optional pre-test uv network-download recovery. The user-authorized plan
+enables it explicitly. Ordinary runs still use the default timeout-only policy.
+
+The standard Harbor entry point ran two fresh installation-only checks before
+model work. Both succeeded, while both verifier dependency profiles remained
+explicitly **`not_configured`**. Installation success was not treated as verifier
+dependency certification. The model then ran once per task. Verification reused
+each fresh output and container, with at most three total attempts and the
+original **900-second per-attempt limit**. Valid official results, including
+zero, stop recovery immediately; setup failures and verifier errors remain
+separate from scored failures. Original package sources and test scripts were
+unchanged.
+
+The sampler installed R, submitted `ars.R`, and passed all nine official tests,
+including distribution sampling, sample files, modularity, input validation,
+and non-log-concavity handling. Verification took about 38.3 seconds, with
+8.55 seconds in pytest. It required no retry.
+
+Pipeline's first attempt reached its 900-second limit while uv was downloading
+PyTorch/CUDA dependencies; pytest had not started. The second attempt reused
+the same container and cache, finished dependency installation, and executed
+all four official cases. The existence and no-hooks checks passed. Both
+world-size cases failed the backward-signal comparison at **`lm_head.bwd`,
+microbatch 0**: rank 0 for world size 1 and rank 1 for world size 2, each with
+maximum difference **0.03571479767560959**. The agent reported successful
+self-checks on its own CPU PyTorch/Transformers stack, but these did not
+establish compliance with the official comparison. This is a valid scored
+failure, not a verifier installation error. Native CTRF deduplicates the
+parameterized cases into three test names; actual pytest execution was
+**two passed and two failed**. Both representations are retained.
+
+Pipeline's second attempt took 828.9 seconds, including 76.46 seconds in pytest.
+The retry budget was therefore respected: one native timeout retry, no
+whole-trial retries, and no attempt beyond the valid failure. The optional
+dependency-download-error rule did not trigger in these native trials; its
+behavior is covered by regression tests.
+
+All outputs froze at **2026-10-18 15:06:01 UTC**, after both trials ended and
+before new trajectory or official-test-content review. Thirteen official
+pytest cases executed: 11 passed and two failed, with no skips or errors.
+Both `/app` artifact collections succeeded, preserving submitted code and
+sample files locally for future inspection or verification. Compact artifact
+tree digests and exact submitted-code/sample hashes are recorded in the results.
+All original image identities were unchanged and trial containers were removed.
+A review of both frozen trajectories found no observed benchmark-answer
+retrieval, hidden verifier content, or provider credentials. Pipeline listed
+an empty verifier-log directory and inspected its own bootstrap/startup logs;
+other network accesses concerned public dependency sources.
+
+The retry implementation passed all **178 Harbor workflow tests**, including
+bounded retries, archival, exhausted errors remaining unscored, default behavior,
+and stopping on valid results. The implementation's three CI checks passed.
+The English and Chinese adapter READMEs both document the optional mode.
+
+Model jobs occupied about **39.4 minutes** of wall time, excluding preflight.
+This remains an exploratory follow-up on previously exposed tasks with fresh
+model outputs and live dependency resolution. It does not isolate a checkpoint
+effect or provide a fresh score for the full 34-task pool. Raw trajectories,
+logs, and `/app` contents remain local; only derived findings and hashes are
+published.
diff --git a/benchmarks/harbor/reports/deepseek-v41-flash-1017-rerun-20261018.md b/benchmarks/harbor/reports/deepseek-v41-flash-1017-rerun-20261018.md
new file mode 100644
index 0000000..9960d86
--- /dev/null
+++ b/benchmarks/harbor/reports/deepseek-v41-flash-1017-rerun-20261018.md
@@ -0,0 +1,111 @@
+# DeepSeek Flash 1017: six previous failures rerun
+
+This report preserves the initial six trials. A later
+[authorized follow-up](deepseek-v41-flash-1017-recovery-20261018.md) generated
+fresh outputs for the two unscored tasks: the sampler passed and pipeline
+parallelism failed after official tests executed. The initial results below
+remain unchanged.
+
+Three previous failures passed: **WAL recovery, gRPC KV store, and SQL
+optimization**. Regex chess failed. The sampler and pipeline-parallelism
+trials are **unscored verifier errors**: dependency downloads failed before
+pytest started. Their native reward zeros are preserved in the results but
+are not counted as model failures.
+
+These are fresh, single trials of all six failures from the
+[October 12 baseline](dev34-main-baseline-20261012.md), using the same agent and
+adapter source, prompt, task revisions, resources, and work budgets. The
+[machine-readable results](../results/deepseek-v41-flash-1017-rerun-20261018.json)
+contain trial IDs, raw rewards, analytical scores, timing, source and artifact
+hashes, preflight history, and audit findings. The
+[final registration](../configs/deepseek-v41-flash-1017-rerun-20261018/experiment-v2.json)
+has SHA256
+`9c9306f3eb17c3044987587a7329671160ee902a43c2b67be6de4f857e4c96c2`.
+
+| Task | Previous result | New result | Official tests passed / failed | Agent seconds |
+| --- | --- | --- | ---: | ---: |
+| adaptive-rejection-sampler | Failed | Unscored: NumPy download timeout | Not started | 627.6 |
+| db-wal-recovery | Failed | Passed | 7 / 0 | 76.2 |
+| kv-store-grpc | Failed | Passed | 7 / 0 | 304.2 |
+| query-optimize | Failed | Passed | 6 / 0 | 287.0 |
+| torch-pipeline-parallelism | Failed | Unscored: CUDA dependency download timeout | Not started | 613.9 |
+| regex-chess | Failed | Failed: required file missing | 0 / 4 | 644.8 |
+
+Twenty-four official test cases executed across four tasks: 20 passed and four
+failed. All six model trials completed without setup errors. There were zero
+whole-trial retries and zero verifier retries. This incomplete, selected subset
+does not provide a fresh score for the entire 34-task development pool.
+
+The run used `deepseek/deepseek-flash` at the official
+`https://api.deepseek.com/anthropic` endpoint. DeepSeek's
+[API documentation](https://api-docs.deepseek.com/quick_start/agent_integrations/)
+and [October 17 announcement](https://api-docs.deepseek.com/news/news261017/)
+identify `deepseek-flash` as **DeepSeek-V4.1-Flash-1017**. This mapping was checked
+before execution; the alias is not an immutable checkpoint pin.
+
+The evaluated source remained
+`2eb99b608dfb39464d17d109a4a8f11afda607a9`, with the original 385-word headless
+prompt. Every model trial installed the previous baseline's exact wheel,
+SHA256 `f4dcc18758e584696f955ae57fd18c751599b70e8b497ca2db7ef9e2b16e9350`.
+Limits remained 100 replies and 65,536 output tokens per request; work budgets
+were 720 seconds for five tasks and 3420 seconds for regex chess. Consecutive
+pairs ran in ascending budget/name order with concurrency two. The standard
+Harbor entry point built and recorded every source snapshot and wheel.
+Temperature and reasoning effort remained unset.
+
+WAL recovery preserved the original database and sidecar before opening SQLite,
+decoded the WAL, and recovered all 11 records. gRPC passed the official request
+contract tests. SQL passed correctness, database-preservation, and the official
+optimized-reference performance test. The failure diagnoses from the previous
+run were not injected into any new task instruction.
+
+Regex chess stopped after about 10.7 minutes because of response truncation,
+well before its 57-minute work deadline. ATIF steps 5 and 8 both ended with
+`stop_reason=max_tokens`, reporting 65,536 and 65,535 completion tokens. The
+agent used its one automatic truncation recovery between those responses.
+Neither response produced an executed file-writing tool call, so `/app/re.json`
+was absent and all four official tests failed with `FileNotFoundError`.
+This run evaluated no submitted chess implementation; it does not establish
+whether the previous castling bug was fixed.
+
+The sampler's original verifier failed to download `numpy==2.3.3` because of a
+network timeout. Pipeline verification failed to download
+`nvidia-cusparse-cu12==12.5.4.2`, required by its original `torch==2.7.0`
+dependency. Neither verifier reached pytest or produced a CTRF test report.
+The original wrappers still returned native reward zero without a Harbor
+exception. The analytical scores are therefore null, classified as
+`verifier_error`. These were download errors within completed attempts of
+363.0 and 159.8 seconds, not verifier-process timeouts; the registered
+timeout-only policy did not permit another attempt. The submitted R sampler
+and pipeline implementation remain officially unverified. Agent-created R
+tests and a custom torch stub do not replace official verification.
+
+Before model work, two installation-only trials were administratively
+interrupted after slow downloads of uv's normally selected original CPython
+3.13.9+20251120 archive. A transport amendment added a verified cache entry for
+that identical archive, then all six fresh installation checks succeeded.
+The archive's upstream size, filename, and SHA256 were verified with the
+repository's existing cache utility. All six verifier dependency profiles were
+explicitly **`not_configured`**; successful agent installation did not certify
+their verifier dependencies. The
+[configuration README](../configs/deepseek-v41-flash-1017-rerun-20261018/README.md)
+documents both plans and the retained interruption evidence. Original images,
+package sources, official test scripts, resources, and verifier time limits
+were preserved.
+
+All six native trials ended before trajectory or official-test-content review.
+Outputs froze at **2026-10-18 14:09:39 UTC**. Source snapshots, installed wheel
+digests, task/configuration hashes, preflight evidence, and queue barriers were
+verified. Both registered plans and all 14 Harbor configurations validated.
+All six original image IDs and registry digests remained unchanged; no scored
+trial containers remained. A scan and review of the six frozen trajectories
+found no observed benchmark-answer retrieval, hidden verifier access, or
+provider credential exposure. Raw logs and trajectories remain local; only
+derived findings and evidence hashes are published.
+
+The model jobs occupied approximately **40.1 minutes** of wall time, excluding
+preflight. This is an exploratory rerun of previously exposed development tasks
+selected because they failed. Single trials, live dependency resolution, and
+changing service conditions cannot isolate a checkpoint effect or establish
+reliability. The supported conclusion is three newly observed passes, one
+valid scored failure, and two unresolved official evaluations.
diff --git a/benchmarks/harbor/results/deepseek-v41-flash-1017-recovery-20261018.json b/benchmarks/harbor/results/deepseek-v41-flash-1017-recovery-20261018.json
new file mode 100644
index 0000000..dd9be0e
--- /dev/null
+++ b/benchmarks/harbor/results/deepseek-v41-flash-1017-recovery-20261018.json
@@ -0,0 +1,551 @@
+{
+  "schema_version": 1,
+  "experiment": "deepseek-v41-flash-1017-recovery-20261018",
+  "collected_at": "2026-10-18T15:06:01.383578+00:00",
+  "cohort": "authorized_followup_of_two_unscored_development_tasks",
+  "dataset_ref": "sha256:7d7bdc1cbedad549fc1140404bd4dc45e5fd0ea7c4186773687d177ad3a0699a",
+  "model": "deepseek/deepseek-flash",
+  "endpoint": "https://api.deepseek.com/anthropic",
+  "runtime_source_ref": "35fe1a2d6a786619a2adace1cfa9d0acca08d533",
+  "plan_path": "benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/experiment.json",
+  "plan_sha256": "7ba0d4083cbc856a364417bc51025a8bd84434eb39319927664bfb4c018fba9d",
+  "model_output_freeze": {
+    "observed_at": "2026-10-18T15:06:01.314545+00:00",
+    "plan_path": "benchmarks/harbor/configs/deepseek-v41-flash-1017-recovery-20261018/experiment.json",
+    "plan_sha256": "7ba0d4083cbc856a364417bc51025a8bd84434eb39319927664bfb4c018fba9d",
+    "runtime_source_ref": "35fe1a2d6a786619a2adace1cfa9d0acca08d533",
+    "boundary": "Both queued native trials ended before new trajectory or hidden-test content review. Setup errors, if any, have no model output and remain separate.",
+    "trials": [
+      {
+        "task": "terminal-bench/adaptive-rejection-sampler",
+        "trial": "adaptive-rejection-sampler__v3bi7q6",
+        "finished_at": "2026-10-18T14:37:47.983031Z",
+        "model_started": true,
+        "trajectory_sha256": "199a3ecd99d805672a8fbd23735d224d6c85fbd598695b1d4de60d36bfeac261",
+        "raw_rewards": {
+          "reward": 1.0
+        },
+        "exception_type": null,
+        "files": {
+          "jobs/flash1017-recovery-orchestration/flash1017-recovery-adaptive-rejection-sampler-20261018.json": "cc470922d79504a61f7b3e0365db5b0cb9395cf389c7ecc1cac38be954813f49",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/agent/bootstrap.json": "ab0fd54609830cbb821308672c74215feb68f8f369542e9be758228f7dcc43bb",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/agent/nanopycodeagent.txt": "9c33098ea3f1f74921fa385f3b8628b2cf6289301df582126cd30be7df300997",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/agent/trajectory.json": "199a3ecd99d805672a8fbd23735d224d6c85fbd598695b1d4de60d36bfeac261",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/artifacts/app/ars.R": "bdd2f472d3cb277ed7f8cf40c8602a412df1dd600ffa116f86d6ecb79059c691",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/artifacts/app/exponential_samples.txt": "42c2e8ce1b239d95c11bd9a133f0b4532fa5075a772defbf5d6112a3b0a4ba74",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/artifacts/app/normal_samples.txt": "1469fd064c67f57a194f13b486e917f9fe261ffca33a83ea6bb653f7cedd2dfc",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/artifacts/manifest.json": "502e6df5db116b0d94d1f1c7cd9a32405cab81979a51f3dbecb38c424c665aa6",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/config.json": "fb0cf9aecc9cda1bd79b0f8f09fe303688ed70f0f072e4e4afa2dd20639c56aa",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/lock.json": "6577d0d09dd44cba05e0680551cec770f8afc77bcffb479fdca2a1f1b362161e",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/result.json": "e78a62618d3625b0616674d6270616bd143d4fb7acde9c94658f3dd52e7e2379",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/trial.log": "2e5c1438295e0a77b03d5917f99e44d3f6b91b276ace757a5abb82cc94115af7",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/verifier/attempts/1/ctrf.json": "bb1a28d8381f9faefb237261c4768736060469a5404a2a6097e9f4f9c01ff000",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/verifier/attempts/1/reward.txt": "4355a46b19d348dc2f57c046f8ef63d4538ebb936000f3c9ee954a27460dd865",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/verifier/attempts/1/test-stdout.txt": "79b0ca14edef47743b08e6e94427a79cb03c8b231b20598aada9470437511f77",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/verifier/ctrf.json": "bb1a28d8381f9faefb237261c4768736060469a5404a2a6097e9f4f9c01ff000",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/verifier/retry-summary.json": "26c96b5931ae1e7ff4e86001b889c4e1934e08b3c7f4380a1c4f00715231d973",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/verifier/reward.txt": "4355a46b19d348dc2f57c046f8ef63d4538ebb936000f3c9ee954a27460dd865",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/verifier/test-stdout.txt": "79b0ca14edef47743b08e6e94427a79cb03c8b231b20598aada9470437511f77"
+        },
+        "artifact_inventory": {
+          "tree_sha256": "10df4e3d1be3d4c29ba6d631e3afcb7ac1218da2dc7a2c5e098f99be81069ea7",
+          "entries": 4,
+          "regular_file_bytes": 310907,
+          "hash_recipe": "SHA256 of compact JSON [relative_path, file_or_symlink, content_sha256_or_link_target] entries sorted by path; symlinks are not followed."
+        }
+      },
+      {
+        "task": "terminal-bench/torch-pipeline-parallelism",
+        "trial": "torch-pipeline-parallelism__ASK4Vjz",
+        "finished_at": "2026-10-18T15:05:42.487531Z",
+        "model_started": true,
+        "trajectory_sha256": "d579b59466eaec4ebcdf0c24f51ece4e0fe78957cffc9446da776d9732e4c90d",
+        "raw_rewards": {
+          "reward": 0.0
+        },
+        "exception_type": null,
+        "files": {
+          "jobs/flash1017-recovery-orchestration/flash1017-recovery-torch-pipeline-parallelism-20261018.json": "9ae9ba153dfa2f438827058096a5e068eb5c322e1f51d954c77ddb3a5c699b85",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/agent/bootstrap.json": "ab0fd54609830cbb821308672c74215feb68f8f369542e9be758228f7dcc43bb",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/agent/nanopycodeagent.txt": "f86373e326cd647b3a8f80215b1a995828011d8eccaee40fd068fc24e40d48b8",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/agent/trajectory.json": "d579b59466eaec4ebcdf0c24f51ece4e0fe78957cffc9446da776d9732e4c90d",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/artifacts/app/pipeline_parallel.py": "1ac320ac344f1b7a6827c7350041d507be8c78184951b9a0e9776bd0ed89b6ef",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/artifacts/manifest.json": "502e6df5db116b0d94d1f1c7cd9a32405cab81979a51f3dbecb38c424c665aa6",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/config.json": "3f59f758dd8ba5085f389946c72f68ba48116ebcfd82b61220b210f939064c56",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/lock.json": "e799139a792ac94aa035b759a9200b4f1a918e49df42c03cb9f2e60e268a0f95",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/result.json": "493eed193c955b521e6c81e008bd7633996f8eef000a2e1e3383088d7fa4dcd3",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/trial.log": "ba4fc53101d582f309f83682c7822b7fe12db2f6612efca6eaade93e22dcc6b1",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/verifier/attempts/1/test-stdout.txt": "1fca2b129f1887d7ba83351cba952121d322f010c513bddbae7aa40ba2594236",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/verifier/attempts/2/ctrf.json": "15a983b890b8a8e5c4a54faccb22a90c169acd55fd06478cab7312a03ce30944",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/verifier/attempts/2/reward.txt": "9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/verifier/attempts/2/test-stdout.txt": "0bddff2743a055b821c34c5c1d02f6bb95895ba6c62482a52d8a2dbbe8b921a7",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/verifier/ctrf.json": "15a983b890b8a8e5c4a54faccb22a90c169acd55fd06478cab7312a03ce30944",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/verifier/retry-summary.json": "8f0b19e524e9564445ef8a73b0a61ff71dff95a150aba3aed9a6836ce8c768ed",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/verifier/reward.txt": "9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa",
+          "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/verifier/test-stdout.txt": "0bddff2743a055b821c34c5c1d02f6bb95895ba6c62482a52d8a2dbbe8b921a7"
+        },
+        "artifact_inventory": {
+          "tree_sha256": "ba32935d9221fdf174c3cb815d11ca4dca467ef0ddbf4c6bf8d32a4b72249433",
+          "entries": 3,
+          "regular_file_bytes": 31252,
+          "hash_recipe": "SHA256 of compact JSON [relative_path, file_or_symlink, content_sha256_or_link_target] entries sorted by path; symlinks are not followed."
+        }
+      }
+    ]
+  },
+  "integrity": {
+    "registration_unchanged": true,
+    "plan_source_prompt_configs_tasks_unchanged": true,
+    "verified_preflight_tasks": 2,
+    "source_snapshots_verified": 2,
+    "wheel_digests_verified": 2,
+    "frozen_native_trial_artifacts_verified": 2,
+    "queue_pair_barriers_verified": true,
+    "whole_model_trial_retries": 0,
+    "outcome_based_replacements": 0
+  },
+  "aggregate": {
+    "trials": 2,
+    "models_started": 2,
+    "raw_passes": 1,
+    "raw_zeros": 1,
+    "classifications": {
+      "scored_pass": 1,
+      "scored_failure": 1
+    },
+    "verifier_timeout_retries": 1,
+    "estimated_known_cost_usd": 0.138438564,
+    "cost_complete": true,
+    "usage_incomplete_tasks": [],
+    "recorded_tokens": {
+      "prompt_tokens": 1652399,
+      "completion_tokens": 94203,
+      "cached_tokens": 1599744
+    },
+    "summed_stage_seconds": {
+      "environment_setup": 4.241263999999999,
+      "agent_setup": 82.273162,
+      "agent_execution": 1153.967016,
+      "verifier": 1766.249108
+    },
+    "model_job_wall_seconds": 2363.087596,
+    "dependency_download_retries": 0,
+    "official_pytest_counts": {
+      "collected": 13,
+      "passed": 11,
+      "failed": 2,
+      "skipped": 0,
+      "error": 0
+    },
+    "verifier_attempts_total": 3,
+    "retry_limit_respected": true,
+    "setup_errors": 0,
+    "verifier_errors": 0
+  },
+  "trials": [
+    {
+      "order": 1,
+      "task": "terminal-bench/adaptive-rejection-sampler",
+      "task_ref": "sha256:bcaa2399985cd57666018025846289ab25e193ae0dd8fb7f0ffab2410c24d4de",
+      "trial": "adaptive-rejection-sampler__v3bi7q6",
+      "job": "flash1017-recovery-adaptive-rejection-sampler-20261018",
+      "raw_reward": 1.0,
+      "classification": "scored_pass",
+      "exception_type": null,
+      "official_test_summary": {
+        "tests": 9,
+        "passed": 9,
+        "failed": 0,
+        "skipped": 0,
+        "pending": 0,
+        "other": 0,
+        "start": 1792334238.2265651,
+        "stop": 1792334246.7767608
+      },
+      "started_at": "2026-10-18T14:26:19.399935Z",
+      "finished_at": "2026-10-18T14:37:47.983031Z",
+      "seconds": {
+        "environment_setup": 2.185485,
+        "agent_setup": 35.034209,
+        "agent_execution": 591.356955,
+        "verifier": 38.513416
+      },
+      "work_budget_seconds": 720,
+      "verifier_timeout_seconds": 900.0,
+      "verifier_attempts": [
+        {
+          "attempt": 1,
+          "status": "completed",
+          "duration_seconds": 38.279680895007914,
+          "return_code": 0
+        }
+      ],
+      "verifier_dependency_profile": {
+        "profile": null,
+        "status": "not_configured"
+      },
+      "bootstrap_status": "ready",
+      "model_started": true,
+      "trajectory_terminal_outcome": "completed",
+      "recorded_agent_steps": 23,
+      "tool_calls": 23,
+      "incomplete_model_steps": [],
+      "accounting": {
+        "usage_complete": true,
+        "cost_complete": true,
+        "cost_is_estimated": true,
+        "recorded_tokens": {
+          "prompt_tokens": 891663,
+          "completion_tokens": 56214,
+          "cached_tokens": 852352
+        },
+        "estimated_known_cost_usd": 0.084364212
+      },
+      "task_checksum": "a4726253ad23c3fc46b5cb9c3d21025874e367b18643a3920f636a1fa89e1207",
+      "source_ref": "35fe1a2d6a786619a2adace1cfa9d0acca08d533",
+      "source_sha256": {
+        "nanopycodeagent/__init__.py": "b922299f65b3db2fb1514df5aeda4413e599f595d2595f45a353b4dc551a1eb1",
+        "nanopycodeagent/agent.py": "aa58960803be77797c2addf65d7cc785a4be54f40ad58d53b9eea9fca1cf4b56",
+        "nanopycodeagent/atif.py": "ca4793b971d125d904888b858452db8db30e21f0d9b87818133062fb6cebe270",
+        "nanopycodeagent/bash_tool.py": "dcf0366fb9ebdca258d4c40b1a50b78b2ec79fc4b85cbb253ac9a9ce9534574f",
+        "nanopycodeagent/cli.py": "e335789b3f09440abd441547a2e3e2feb32b90c68e9ba5fc8237de8f6bac2932",
+        "nanopycodeagent/cost.py": "30a8df17942c424796e8532f45e1cb0784ed6541cea44284b863db3dae2f7b70",
+        "nanopycodeagent/deadline.py": "26a861721e6897a4e02e6ce364cb3ef1f1dc7d4ad9fb2f772404276f5c9e9d5f",
+        "nanopycodeagent/edit_tool.py": "78e2bec911279082b081366ab05cb4cbb37d8910b20b6245cccd4b0424489d0d",
+        "nanopycodeagent/event_journal.py": "8974919c2d460afb52bc57ec37e7e271851290c0845d01b4ad0f9f760cde0814",
+        "nanopycodeagent/read_tool.py": "e0075da19f2e1c6ccf34a3056a0aec92398e5afe67869937a096c6f1e04181c3",
+        "nanopycodeagent/settings.py": "491bf5c6eca50eb4c5fe04b7bfe62a8bfc7d2e1fad05b8eeac1680c2fcf3674e",
+        "nanopycodeagent/terminal.py": "8d27b40976b2b6d476e5106dfea982b055115435b4ba16b7f758412a05b75f04",
+        "nanopycodeagent/tool_validation.py": "3a1a85a7222dd1504bdf58f37e65a51d11de13d16c4e7df42c72b4cd58b5e5e5",
+        "nanopycodeagent/transport.py": "152d3ff78be0e3b2d92e0abbbfb652d73466aff57ad316c06dc0e072411167fe",
+        "nanopycodeagent/write_tool.py": "7671e83464fa1e942f25742fc573d0712e32053b5fc0291257884b027f7f6206",
+        "harbor_adapter/__init__.py": "513ad49c721b3e73293095c6a0486a791483e9aa2b51e0e9ad4b7816af07e76b",
+        "harbor_adapter/__main__.py": "db89d582fba45aad72147317f595531af49c434e3b23c65f22d7491f0b10551d",
+        "harbor_adapter/_verifier_runner.py": "d718f7c5528c156df549d68ab2deff247a289d08a453c6d083b72d0e9c7013dc",
+        "harbor_adapter/adapter.py": "d1aa154ec6972d706f0725cfa69c56635506d4499dff2a129e2c957ea7034c02",
+        "harbor_adapter/bootstrap.py": "f499bfddfc3d29338cae15a7f8f6e9c66d19eb69ad9577e607ced4b1aa057be2",
+        "harbor_adapter/verifier.py": "44bf247282437a0092bd7e649fd7fa9c776b6d34a0c14ffedcc10b37a47608ef"
+      },
+      "wheel_sha256": "f1113a2b382ae8a9ae8f9d231b8e35ce98e71f910959d6c8df8e92e6f200c700",
+      "provenance_checks_passed": true,
+      "evidence_sha256": {
+        "jobs/flash1017-recovery-orchestration/flash1017-recovery-adaptive-rejection-sampler-20261018.json": "cc470922d79504a61f7b3e0365db5b0cb9395cf389c7ecc1cac38be954813f49",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/agent/bootstrap.json": "ab0fd54609830cbb821308672c74215feb68f8f369542e9be758228f7dcc43bb",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/agent/nanopycodeagent.txt": "9c33098ea3f1f74921fa385f3b8628b2cf6289301df582126cd30be7df300997",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/agent/trajectory.json": "199a3ecd99d805672a8fbd23735d224d6c85fbd598695b1d4de60d36bfeac261",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/artifacts/app/ars.R": "bdd2f472d3cb277ed7f8cf40c8602a412df1dd600ffa116f86d6ecb79059c691",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/artifacts/app/exponential_samples.txt": "42c2e8ce1b239d95c11bd9a133f0b4532fa5075a772defbf5d6112a3b0a4ba74",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/artifacts/app/normal_samples.txt": "1469fd064c67f57a194f13b486e917f9fe261ffca33a83ea6bb653f7cedd2dfc",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/artifacts/manifest.json": "502e6df5db116b0d94d1f1c7cd9a32405cab81979a51f3dbecb38c424c665aa6",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/config.json": "fb0cf9aecc9cda1bd79b0f8f09fe303688ed70f0f072e4e4afa2dd20639c56aa",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/lock.json": "6577d0d09dd44cba05e0680551cec770f8afc77bcffb479fdca2a1f1b362161e",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/result.json": "e78a62618d3625b0616674d6270616bd143d4fb7acde9c94658f3dd52e7e2379",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/trial.log": "2e5c1438295e0a77b03d5917f99e44d3f6b91b276ace757a5abb82cc94115af7",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/verifier/attempts/1/ctrf.json": "bb1a28d8381f9faefb237261c4768736060469a5404a2a6097e9f4f9c01ff000",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/verifier/attempts/1/reward.txt": "4355a46b19d348dc2f57c046f8ef63d4538ebb936000f3c9ee954a27460dd865",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/verifier/attempts/1/test-stdout.txt": "79b0ca14edef47743b08e6e94427a79cb03c8b231b20598aada9470437511f77",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/verifier/ctrf.json": "bb1a28d8381f9faefb237261c4768736060469a5404a2a6097e9f4f9c01ff000",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/verifier/retry-summary.json": "26c96b5931ae1e7ff4e86001b889c4e1934e08b3c7f4380a1c4f00715231d973",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/verifier/reward.txt": "4355a46b19d348dc2f57c046f8ef63d4538ebb936000f3c9ee954a27460dd865",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__v3bi7q6/verifier/test-stdout.txt": "79b0ca14edef47743b08e6e94427a79cb03c8b231b20598aada9470437511f77",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018-input/manifest.json": "47fe90c4df1ca3a54c36f00fa490c9b8aa84780e96789f24c250e56e1e7dfddf",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018-input/config.json": "69c83d4c0d3b9554cb67f63af7d5952bd81315a53ef7fbdea3070ffebd7e9263",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-adaptive-rejection-sampler-20261018-input/working-tree.patch": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
+      },
+      "prior_unscored_trial": {
+        "task": "terminal-bench/adaptive-rejection-sampler",
+        "trial": "adaptive-rejection-sampler__Bim6vqF",
+        "classification": "verifier_error",
+        "raw_reward": 0.0,
+        "score": null
+      },
+      "official_pytest_summary": {
+        "executed": true,
+        "collected": 9,
+        "passed": 9,
+        "failed": 0,
+        "skipped": 0,
+        "error": 0,
+        "summary_line": "============================== 9 passed in 8.55s ===============================",
+        "source": "verifier/test-stdout.txt"
+      },
+      "score": 1.0,
+      "preserved_artifacts": {
+        "tree_sha256": "10df4e3d1be3d4c29ba6d631e3afcb7ac1218da2dc7a2c5e098f99be81069ea7",
+        "entries": 4,
+        "regular_file_bytes": 310907,
+        "hash_recipe": "SHA256 of compact JSON [relative_path, file_or_symlink, content_sha256_or_link_target] entries sorted by path; symlinks are not followed."
+      },
+      "artifact_manifest": [
+        {
+          "source": "/logs/artifacts",
+          "destination": "artifacts/logs/artifacts",
+          "type": "directory",
+          "status": "empty",
+          "service": null
+        },
+        {
+          "source": "/app",
+          "destination": "artifacts/app",
+          "type": "directory",
+          "status": "ok",
+          "service": null
+        }
+      ],
+      "audit": {
+        "benchmark_material_exposure": "not_observed_in_review",
+        "provider_credentials_observed": false
+      },
+      "submitted_code_sha256": "bdd2f472d3cb277ed7f8cf40c8602a412df1dd600ffa116f86d6ecb79059c691",
+      "retry_limit_respected": true
+    },
+    {
+      "order": 2,
+      "task": "terminal-bench/torch-pipeline-parallelism",
+      "task_ref": "sha256:db605337c749a872cea7b5b413429b3915bb4c3efe0f7875f0c46ce81bd8c4fb",
+      "trial": "torch-pipeline-parallelism__ASK4Vjz",
+      "job": "flash1017-recovery-torch-pipeline-parallelism-20261018",
+      "raw_reward": 0.0,
+      "classification": "scored_failure",
+      "exception_type": null,
+      "official_test_summary": {
+        "tests": 3,
+        "passed": 2,
+        "failed": 1,
+        "skipped": 0,
+        "pending": 0,
+        "other": 0,
+        "start": 1792335842.1998892,
+        "stop": 1792335918.6577215
+      },
+      "started_at": "2026-10-18T14:26:19.788639Z",
+      "finished_at": "2026-10-18T15:05:42.487531Z",
+      "seconds": {
+        "environment_setup": 2.055779,
+        "agent_setup": 47.238953,
+        "agent_execution": 562.610061,
+        "verifier": 1727.735692
+      },
+      "work_budget_seconds": 720,
+      "verifier_timeout_seconds": 900.0,
+      "verifier_attempts": [
+        {
+          "attempt": 1,
+          "status": "timeout",
+          "duration_seconds": 900.0019356679986,
+          "phase_at_timeout": "dependency_preparation",
+          "official_assertions_started": false
+        },
+        {
+          "attempt": 2,
+          "status": "completed",
+          "duration_seconds": 828.9257361209893,
+          "return_code": 0,
+          "official_assertions_started": true
+        }
+      ],
+      "verifier_dependency_profile": {
+        "profile": null,
+        "status": "not_configured"
+      },
+      "bootstrap_status": "ready",
+      "model_started": true,
+      "trajectory_terminal_outcome": "completed",
+      "recorded_agent_steps": 27,
+      "tool_calls": 31,
+      "incomplete_model_steps": [],
+      "accounting": {
+        "usage_complete": true,
+        "cost_complete": true,
+        "cost_is_estimated": true,
+        "recorded_tokens": {
+          "prompt_tokens": 760736,
+          "completion_tokens": 37989,
+          "cached_tokens": 747392
+        },
+        "estimated_known_cost_usd": 0.054074352
+      },
+      "task_checksum": "31e565558cd22598a3c2069f771d9d0eee26fb843b77f7b03e00ad4c336fbce9",
+      "source_ref": "35fe1a2d6a786619a2adace1cfa9d0acca08d533",
+      "source_sha256": {
+        "nanopycodeagent/__init__.py": "b922299f65b3db2fb1514df5aeda4413e599f595d2595f45a353b4dc551a1eb1",
+        "nanopycodeagent/agent.py": "aa58960803be77797c2addf65d7cc785a4be54f40ad58d53b9eea9fca1cf4b56",
+        "nanopycodeagent/atif.py": "ca4793b971d125d904888b858452db8db30e21f0d9b87818133062fb6cebe270",
+        "nanopycodeagent/bash_tool.py": "dcf0366fb9ebdca258d4c40b1a50b78b2ec79fc4b85cbb253ac9a9ce9534574f",
+        "nanopycodeagent/cli.py": "e335789b3f09440abd441547a2e3e2feb32b90c68e9ba5fc8237de8f6bac2932",
+        "nanopycodeagent/cost.py": "30a8df17942c424796e8532f45e1cb0784ed6541cea44284b863db3dae2f7b70",
+        "nanopycodeagent/deadline.py": "26a861721e6897a4e02e6ce364cb3ef1f1dc7d4ad9fb2f772404276f5c9e9d5f",
+        "nanopycodeagent/edit_tool.py": "78e2bec911279082b081366ab05cb4cbb37d8910b20b6245cccd4b0424489d0d",
+        "nanopycodeagent/event_journal.py": "8974919c2d460afb52bc57ec37e7e271851290c0845d01b4ad0f9f760cde0814",
+        "nanopycodeagent/read_tool.py": "e0075da19f2e1c6ccf34a3056a0aec92398e5afe67869937a096c6f1e04181c3",
+        "nanopycodeagent/settings.py": "491bf5c6eca50eb4c5fe04b7bfe62a8bfc7d2e1fad05b8eeac1680c2fcf3674e",
+        "nanopycodeagent/terminal.py": "8d27b40976b2b6d476e5106dfea982b055115435b4ba16b7f758412a05b75f04",
+        "nanopycodeagent/tool_validation.py": "3a1a85a7222dd1504bdf58f37e65a51d11de13d16c4e7df42c72b4cd58b5e5e5",
+        "nanopycodeagent/transport.py": "152d3ff78be0e3b2d92e0abbbfb652d73466aff57ad316c06dc0e072411167fe",
+        "nanopycodeagent/write_tool.py": "7671e83464fa1e942f25742fc573d0712e32053b5fc0291257884b027f7f6206",
+        "harbor_adapter/__init__.py": "513ad49c721b3e73293095c6a0486a791483e9aa2b51e0e9ad4b7816af07e76b",
+        "harbor_adapter/__main__.py": "db89d582fba45aad72147317f595531af49c434e3b23c65f22d7491f0b10551d",
+        "harbor_adapter/_verifier_runner.py": "d718f7c5528c156df549d68ab2deff247a289d08a453c6d083b72d0e9c7013dc",
+        "harbor_adapter/adapter.py": "d1aa154ec6972d706f0725cfa69c56635506d4499dff2a129e2c957ea7034c02",
+        "harbor_adapter/bootstrap.py": "f499bfddfc3d29338cae15a7f8f6e9c66d19eb69ad9577e607ced4b1aa057be2",
+        "harbor_adapter/verifier.py": "44bf247282437a0092bd7e649fd7fa9c776b6d34a0c14ffedcc10b37a47608ef"
+      },
+      "wheel_sha256": "f1113a2b382ae8a9ae8f9d231b8e35ce98e71f910959d6c8df8e92e6f200c700",
+      "provenance_checks_passed": true,
+      "evidence_sha256": {
+        "jobs/flash1017-recovery-orchestration/flash1017-recovery-torch-pipeline-parallelism-20261018.json": "9ae9ba153dfa2f438827058096a5e068eb5c322e1f51d954c77ddb3a5c699b85",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/agent/bootstrap.json": "ab0fd54609830cbb821308672c74215feb68f8f369542e9be758228f7dcc43bb",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/agent/nanopycodeagent.txt": "f86373e326cd647b3a8f80215b1a995828011d8eccaee40fd068fc24e40d48b8",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/agent/trajectory.json": "d579b59466eaec4ebcdf0c24f51ece4e0fe78957cffc9446da776d9732e4c90d",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/artifacts/app/pipeline_parallel.py": "1ac320ac344f1b7a6827c7350041d507be8c78184951b9a0e9776bd0ed89b6ef",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/artifacts/manifest.json": "502e6df5db116b0d94d1f1c7cd9a32405cab81979a51f3dbecb38c424c665aa6",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/config.json": "3f59f758dd8ba5085f389946c72f68ba48116ebcfd82b61220b210f939064c56",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/lock.json": "e799139a792ac94aa035b759a9200b4f1a918e49df42c03cb9f2e60e268a0f95",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/result.json": "493eed193c955b521e6c81e008bd7633996f8eef000a2e1e3383088d7fa4dcd3",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/trial.log": "ba4fc53101d582f309f83682c7822b7fe12db2f6612efca6eaade93e22dcc6b1",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/verifier/attempts/1/test-stdout.txt": "1fca2b129f1887d7ba83351cba952121d322f010c513bddbae7aa40ba2594236",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/verifier/attempts/2/ctrf.json": "15a983b890b8a8e5c4a54faccb22a90c169acd55fd06478cab7312a03ce30944",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/verifier/attempts/2/reward.txt": "9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/verifier/attempts/2/test-stdout.txt": "0bddff2743a055b821c34c5c1d02f6bb95895ba6c62482a52d8a2dbbe8b921a7",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/verifier/ctrf.json": "15a983b890b8a8e5c4a54faccb22a90c169acd55fd06478cab7312a03ce30944",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/verifier/retry-summary.json": "8f0b19e524e9564445ef8a73b0a61ff71dff95a150aba3aed9a6836ce8c768ed",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/verifier/reward.txt": "9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__ASK4Vjz/verifier/test-stdout.txt": "0bddff2743a055b821c34c5c1d02f6bb95895ba6c62482a52d8a2dbbe8b921a7",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018-input/manifest.json": "974861542f065b10aa6e0e3c6827c1b27832a0a4a48693a020e49ae82f2c9454",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018-input/config.json": "fa2e90319d72c6a1d3a156ad79fbcba434165778cb2e88a664f8ae3bb4adff68",
+        "jobs/flash1017-recovery-source/jobs/flash1017-recovery-torch-pipeline-parallelism-20261018-input/working-tree.patch": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
+      },
+      "prior_unscored_trial": {
+        "task": "terminal-bench/torch-pipeline-parallelism",
+        "trial": "torch-pipeline-parallelism__q96EdZw",
+        "classification": "verifier_error",
+        "raw_reward": 0.0,
+        "score": null
+      },
+      "official_pytest_summary": {
+        "executed": true,
+        "collected": 4,
+        "passed": 2,
+        "failed": 2,
+        "skipped": 0,
+        "error": 0,
+        "summary_line": "==================== 2 failed, 2 passed in 76.46s (0:01:16) ====================",
+        "source": "verifier/test-stdout.txt"
+      },
+      "score": 0.0,
+      "preserved_artifacts": {
+        "tree_sha256": "ba32935d9221fdf174c3cb815d11ca4dca467ef0ddbf4c6bf8d32a4b72249433",
+        "entries": 3,
+        "regular_file_bytes": 31252,
+        "hash_recipe": "SHA256 of compact JSON [relative_path, file_or_symlink, content_sha256_or_link_target] entries sorted by path; symlinks are not followed."
+      },
+      "artifact_manifest": [
+        {
+          "source": "/logs/artifacts",
+          "destination": "artifacts/logs/artifacts",
+          "type": "directory",
+          "status": "empty",
+          "service": null
+        },
+        {
+          "source": "/app",
+          "destination": "artifacts/app",
+          "type": "directory",
+          "status": "ok",
+          "service": null
+        }
+      ],
+      "audit": {
+        "benchmark_material_exposure": "not_observed_in_review",
+        "provider_credentials_observed": false,
+        "reviewed_local_access": {
+          "atif_steps": [
+            5,
+            7,
+            9,
+            11
+          ],
+          "finding": "The agent searched public installed files, listed /logs while the verifier directory was empty, and read its own bootstrap/startup logs. No hidden verifier contents or credentials were returned."
+        }
+      },
+      "failure_analysis": {
+        "cause": "Both world-size execution tests reached the submitted code and failed the official backward-signal comparison at lm_head.bwd for microbatch zero.",
+        "official_evidence": "World size 1: rank 0 mismatch; world size 2: rank 1 mismatch; maximum difference 0.03571479767560959 in both cases. Two structural tests pass.",
+        "agent_evidence": "The agent reported successful hook-based self-checks against its own full-batch reference on its installed CPU PyTorch/Transformers stack. Those self-checks did not reproduce the official backward-signal mismatch.",
+        "limitation": "This is a valid scored failure after dependency preparation recovered; the original unscored submission was not restored or evaluated.",
+        "submitted_code_sha256": "1ac320ac344f1b7a6827c7350041d507be8c78184951b9a0e9776bd0ed89b6ef"
+      },
+      "retry_limit_respected": true
+    }
+  ],
+  "publication_policy": "Only derived results and hashes are published. Raw logs and trajectories remain local and require credential redaction before sharing.",
+  "authorization": {
+    "request": "Retry the two unscored tasks, at most three attempts.",
+    "scope": "This follow-up explicitly enables retrying recognized pre-test uv network download failures in addition to native verifier timeouts.",
+    "limit": "At most three total verifier attempts per fresh model output, each with the original 900-second limit. Stop immediately on an executed official test result, including failure.",
+    "original_workspace_availability": "The original containers were deleted and artifact manifests contain no /app outputs; exact original workspace regrading is unavailable.",
+    "new_model_policy": "One fresh model trial per selected task, no whole-trial retries. Within each new trial, reuse its output and container for all authorized verifier attempts.",
+    "artifact_policy": "Collect the complete /app directory for future verification without another model run."
+  },
+  "provider_checkpoint": {
+    "api_model": "deepseek-flash",
+    "reported_checkpoint": "DeepSeek-V4.1-Flash-1017",
+    "verification": "Official DeepSeek documentation states that deepseek-flash now serves this checkpoint. The API alias is not an immutable checkpoint pin.",
+    "sources": [
+      "https://api-docs.deepseek.com/news/news261017/",
+      "https://api-docs.deepseek.com/quick_start/agent_integrations/"
+    ],
+    "checked_at": "2026-10-18T13:16:26.524186+00:00"
+  },
+  "prior_run": "deepseek-v41-flash-1017-rerun-20261018",
+  "preflight_history": {
+    "attempts": 2,
+    "setup_passed": 2,
+    "verifier_profiles_not_configured": 2,
+    "models_or_official_assertions_executed": false,
+    "trials": [
+      {
+        "task": "terminal-bench/adaptive-rejection-sampler",
+        "trial": "adaptive-rejection-sampler__aTArNpp",
+        "result_path": "jobs/flash1017-recovery-source/jobs/flash1017-recovery-install-only-20261018/adaptive-rejection-sampler__aTArNpp/result.json",
+        "result_sha256": "96154b9dd9b82b2cfea76bbd6d87616d054c199221b3e3f3adb986f8e6b29acf",
+        "bootstrap_sha256": "ab0fd54609830cbb821308672c74215feb68f8f369542e9be758228f7dcc43bb",
+        "manifest_sha256": "0db0eff9593da546b363d2b8f6dc5c33be1a06b4ad3ebb9d66b1a9fb7367cbc5",
+        "input_config_sha256": "359a14efc9a1ff2ee2b8d764de6b2a42e5754122a438296b994757e82de915c9",
+        "status": "setup_passed",
+        "verifier_dependency_profile": "not_configured"
+      },
+      {
+        "task": "terminal-bench/torch-pipeline-parallelism",
+        "trial": "torch-pipeline-parallelism__WVwoToA",
+        "result_path": "jobs/flash1017-recovery-source/jobs/flash1017-recovery-install-only-20261018/torch-pipeline-parallelism__WVwoToA/result.json",
+        "result_sha256": "cba4318de561742f16dbd842d47df94ed873230e3060ce43b6da666ed5af0839",
+        "bootstrap_sha256": "ab0fd54609830cbb821308672c74215feb68f8f369542e9be758228f7dcc43bb",
+        "manifest_sha256": "0db0eff9593da546b363d2b8f6dc5c33be1a06b4ad3ebb9d66b1a9fb7367cbc5",
+        "input_config_sha256": "359a14efc9a1ff2ee2b8d764de6b2a42e5754122a438296b994757e82de915c9",
+        "status": "setup_passed",
+        "verifier_dependency_profile": "not_configured"
+      }
+    ]
+  },
+  "original_results_unchanged": true,
+  "agent_source_matches_original_run": true,
+  "final_image_integrity": {
+    "observed_at": "2026-10-18T15:06:52.650992+00:00",
+    "original_task_images_verified": 2,
+    "image_ids_and_registry_digests_unchanged": true,
+    "remaining_scored_trial_containers": []
+  },
+  "post_freeze_audit": {
+    "completed_at": "2026-10-18T15:08:33.164753+00:00",
+    "official_assertions_executed_tasks": 2,
+    "benchmark_material_exposure_observed": false,
+    "provider_credentials_observed_in_trajectories": false,
+    "method": "Review both frozen trajectories and official outputs, including every verifier attempt. Scan exact configured credentials and benchmark/test-path references; inspect dependency and installed-file accesses.",
+    "network_access": "Observed external accesses concerned PyPI, the public PyPA installer and official PyTorch CPU wheels. No benchmark-answer retrieval was observed.",
+    "limitation": "Retrospective trace review cannot prove absence of all exposure or training contamination. Both tasks were previously exposed development tasks."
+  },
+  "comparison_note": "These are fresh model outputs because the initial containers were deleted without /app artifacts. They resolve follow-up outcomes, not the correctness of the two original unscored submissions. A single selected-task follow-up does not isolate checkpoint causality or update the full development-pool score.",
+  "retry_policy_observation": "Only the native timeout retry was exercised in these trials. The explicitly enabled dependency-download-error recovery did not trigger; its behavior is covered by regression tests.",
+  "accounting_note": "Costs are unchanged agent token-based estimates, not independently verified provider billing."
+}
diff --git a/benchmarks/harbor/results/deepseek-v41-flash-1017-rerun-20261018.json b/benchmarks/harbor/results/deepseek-v41-flash-1017-rerun-20261018.json
new file mode 100644
index 0000000..6a22149
--- /dev/null
+++ b/benchmarks/harbor/results/deepseek-v41-flash-1017-rerun-20261018.json
@@ -0,0 +1,1211 @@
+{
+  "schema_version": 1,
+  "experiment": "deepseek-v41-flash-1017-rerun-20261018",
+  "collected_at": "2026-10-18T14:09:40.003848+00:00",
+  "cohort": "six_previous_failures_from_existing_34_task_development_pool",
+  "dataset_ref": "sha256:7d7bdc1cbedad549fc1140404bd4dc45e5fd0ea7c4186773687d177ad3a0699a",
+  "model": "deepseek/deepseek-flash",
+  "endpoint": "https://api.deepseek.com/anthropic",
+  "runtime_source_ref": "2eb99b608dfb39464d17d109a4a8f11afda607a9",
+  "plan_path": "benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/experiment-v2.json",
+  "plan_sha256": "9c9306f3eb17c3044987587a7329671160ee902a43c2b67be6de4f857e4c96c2",
+  "model_output_freeze": {
+    "observed_at": "2026-10-18T14:09:39.938135+00:00",
+    "plan_path": "benchmarks/harbor/configs/deepseek-v41-flash-1017-rerun-20261018/experiment-v2.json",
+    "plan_sha256": "9c9306f3eb17c3044987587a7329671160ee902a43c2b67be6de4f857e4c96c2",
+    "runtime_source_ref": "2eb99b608dfb39464d17d109a4a8f11afda607a9",
+    "boundary": "All six queued native trials ended before new trajectory or hidden-test content review. Setup errors, if any, have no model output and remain separate.",
+    "trials": [
+      {
+        "task": "terminal-bench/adaptive-rejection-sampler",
+        "trial": "adaptive-rejection-sampler__Bim6vqF",
+        "finished_at": "2026-10-18T13:46:54.802088Z",
+        "model_started": true,
+        "trajectory_sha256": "df034659657412897dda47cec614b6d0b803904e5a8651e72eded8b1183aeb34",
+        "raw_rewards": {
+          "reward": 0.0
+        },
+        "exception_type": null,
+        "files": {
+          "jobs/deepseek-v41-flash-1017-rerun-20261018-v2-orchestration/flash1017-adaptive-rejection-sampler-20261018.json": "de95b04d9d0e73c76de3a36e507d39fc2f6e9ef4e6adbd873fa6fee4b8939785",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/agent/bootstrap.json": "fcebfd336f1b466235db7c3fe8ba5ab12f89adf5748f461affa9f8ab75ebc7a6",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/agent/nanopycodeagent.txt": "ef9164ce3172461c2a71bcb8f9c8ea1068131693c51daf293576bc16efba93e7",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/agent/trajectory.json": "df034659657412897dda47cec614b6d0b803904e5a8651e72eded8b1183aeb34",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/artifacts/manifest.json": "c1db653f487fb773c4f9247162ab3314e7f0584a486914b9874c4b3d6949e42d",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/config.json": "b972b206adf474688083b92425631dd1c70300b01690c24e99d7f7e0134c965b",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/lock.json": "1792b7a3b658f4cf272e44f902f584e9ad9b4070ef331b41e03bf8e26152d27e",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/result.json": "c57278943fe30ca5f82d9cf68ffa0ee0bdda61612b4e8e9a3ac4b278d8b19337",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/trial.log": "5f1050a5bc4629de43f0dfbfb4df3e53486deaea3acb4fac98e366700e89719b",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/verifier/attempts/1/reward.txt": "9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/verifier/attempts/1/test-stdout.txt": "9910c1025fef05f9ee04ff45a0364ffc8d88c7a159ad65b40992444fbbd164ad",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/verifier/retry-summary.json": "b70d48f409e1b4b4e89e921302518c03904dcb6304c481ea4054a8b0b9bdaa89",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/verifier/reward.txt": "9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/verifier/test-stdout.txt": "9910c1025fef05f9ee04ff45a0364ffc8d88c7a159ad65b40992444fbbd164ad"
+        }
+      },
+      {
+        "task": "terminal-bench/db-wal-recovery",
+        "trial": "db-wal-recovery__iXi6Qjy",
+        "finished_at": "2026-10-18T13:31:54.497971Z",
+        "model_started": true,
+        "trajectory_sha256": "cc05c9b3c5f217e4067348143b8824495029f8394d450b92273fde2d839bad19",
+        "raw_rewards": {
+          "reward": 1.0
+        },
+        "exception_type": null,
+        "files": {
+          "jobs/deepseek-v41-flash-1017-rerun-20261018-v2-orchestration/flash1017-db-wal-recovery-20261018.json": "aba7a7e2b507d135aad4cc372617e704dd533e1317e98199ab7c1bf5f92c493e",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/agent/bootstrap.json": "51366101df4e7f6425f8fe241b10713287e6f89274173cba25c6ffe708d61dc3",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/agent/nanopycodeagent.txt": "1e5138fd52619c17dc98a132a6b3f5955d05fc9792cc57a2bc1150a3e80ee9cb",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/agent/trajectory.json": "cc05c9b3c5f217e4067348143b8824495029f8394d450b92273fde2d839bad19",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/artifacts/manifest.json": "c1db653f487fb773c4f9247162ab3314e7f0584a486914b9874c4b3d6949e42d",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/config.json": "d88662c669c196d97b3c77a39bbd7d6a10313757b46860feb2258dc4cd40a7c2",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/lock.json": "2d54030d3f9476b33afbe16bf6395230b174e3b248036e76b04fb7ef192b5834",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/result.json": "0a0bfb5de1b212e41bb17b328994f31208ce0eb587151267ca5686b412adb86b",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/trial.log": "331fe02968e34e1c0cecfbf532a32dc552ccdf4232a28c9bc0a0cb266d41c5eb",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/verifier/attempts/1/ctrf.json": "c276c93a192d5ea8078837a7adce56070d661dc2b9cfefe18d63e0ea4746a365",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/verifier/attempts/1/reward.txt": "4355a46b19d348dc2f57c046f8ef63d4538ebb936000f3c9ee954a27460dd865",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/verifier/attempts/1/test-stdout.txt": "f6502d849c96dabb43cda87d0b8d453fb11f08d3a7e28137fbc0c25c4baa030b",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/verifier/ctrf.json": "c276c93a192d5ea8078837a7adce56070d661dc2b9cfefe18d63e0ea4746a365",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/verifier/retry-summary.json": "917b2a0fdfccb2017728dc353860862d45fe1bc59c97fa2547d20eb371fa7656",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/verifier/reward.txt": "4355a46b19d348dc2f57c046f8ef63d4538ebb936000f3c9ee954a27460dd865",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/verifier/test-stdout.txt": "f6502d849c96dabb43cda87d0b8d453fb11f08d3a7e28137fbc0c25c4baa030b"
+        }
+      },
+      {
+        "task": "terminal-bench/kv-store-grpc",
+        "trial": "kv-store-grpc__4TFUA4j",
+        "finished_at": "2026-10-18T13:54:26.858875Z",
+        "model_started": true,
+        "trajectory_sha256": "71d056ce715961081a775ca4461ae1e6cc278fb124dedcea3261e410f7da9264",
+        "raw_rewards": {
+          "reward": 1.0
+        },
+        "exception_type": null,
+        "files": {
+          "jobs/deepseek-v41-flash-1017-rerun-20261018-v2-orchestration/flash1017-kv-store-grpc-20261018.json": "07372fed21de0bfce73dee8b5a71a592be1ae2c5b8ad1db78b5382d2a42cd2a0",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/agent/bootstrap.json": "9abae99e4800f37de01cc8c2eb9d13d0437c60d0c1c0724f648d551c7f382b77",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/agent/nanopycodeagent.txt": "bbabb30397f5ef5da2ea434aca3c1ed9d5ebaa39583585e73cc030801faa315e",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/agent/trajectory.json": "71d056ce715961081a775ca4461ae1e6cc278fb124dedcea3261e410f7da9264",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/artifacts/manifest.json": "c1db653f487fb773c4f9247162ab3314e7f0584a486914b9874c4b3d6949e42d",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/config.json": "93a23d07f2b21646fab7cf47ca38b5fa46c1f58f445e82360d3dadc9c3eedb64",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/lock.json": "3b059c7147009b9eb88fa82bc54c4dd131b04163255a02742140b95db77ca9a7",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/result.json": "81a490fc5dc6b51fc898db40a6f6d678d01575f9da353baf3ffa012c400cfd42",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/trial.log": "786c8e5d96f1ff6e98e74b79d218027a9cd136a89a5ca65e211951fad0cf74a8",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/verifier/attempts/1/ctrf.json": "d4239ddd3b1d9eec2aac414a8fd6cace832f0c854e6a09f36c67ac7d269a1e01",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/verifier/attempts/1/reward.txt": "4355a46b19d348dc2f57c046f8ef63d4538ebb936000f3c9ee954a27460dd865",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/verifier/attempts/1/test-stdout.txt": "bc35e4c8cb26c8751d4f863cc24d30bab201e4eb17ddb3ff2e4bb5a826bc8986",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/verifier/ctrf.json": "d4239ddd3b1d9eec2aac414a8fd6cace832f0c854e6a09f36c67ac7d269a1e01",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/verifier/retry-summary.json": "3fc4da194d0fdde7d6e00f0eb458186f5e5746eb62b0bce74d9734403e513e3d",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/verifier/reward.txt": "4355a46b19d348dc2f57c046f8ef63d4538ebb936000f3c9ee954a27460dd865",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/verifier/test-stdout.txt": "bc35e4c8cb26c8751d4f863cc24d30bab201e4eb17ddb3ff2e4bb5a826bc8986"
+        }
+      },
+      {
+        "task": "terminal-bench/query-optimize",
+        "trial": "query-optimize__ioRN6AM",
+        "finished_at": "2026-10-18T13:55:27.515015Z",
+        "model_started": true,
+        "trajectory_sha256": "126b264a80801d55a7546d651b619c8b35d3b19c05a1639b3b01c1a4478e6094",
+        "raw_rewards": {
+          "reward": 1.0
+        },
+        "exception_type": null,
+        "files": {
+          "jobs/deepseek-v41-flash-1017-rerun-20261018-v2-orchestration/flash1017-query-optimize-20261018.json": "ac61d4144c34cd9e5b698332360a09d161d53213ffdfab0b11e955ef6807fdb1",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/agent/bootstrap.json": "4a93e54276aa8f2cc86d505dad8f483806c983d5c140d69bb855ed60c1cd32b6",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/agent/nanopycodeagent.txt": "65c651a52e18f1354c47b506be980d39cd2ce8c0660a8000b97c1404a74fa435",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/agent/trajectory.json": "126b264a80801d55a7546d651b619c8b35d3b19c05a1639b3b01c1a4478e6094",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/artifacts/manifest.json": "c1db653f487fb773c4f9247162ab3314e7f0584a486914b9874c4b3d6949e42d",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/config.json": "099122689c4c6a00d300285a2efe5ec43259762ed8bc106988dc3327fc162563",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/lock.json": "53d10aefdfebbcb1070f1c9fbbbdb83c96c73d870a6a76db1653dfe25d0fc1be",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/result.json": "4c67b5595e3129d4ac1cabe5c7b4765fe4dad0c1c40a472db78e232d102d327b",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/trial.log": "5bd01f98bd6b0cd4684ced0582ec10d90c9456f262e332feb01fc3690a4f43f6",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/verifier/attempts/1/ctrf.json": "476de5d9485e1e00b2ad4be31fce52726cd533e914d84678ddaad969b354186d",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/verifier/attempts/1/reward.txt": "4355a46b19d348dc2f57c046f8ef63d4538ebb936000f3c9ee954a27460dd865",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/verifier/attempts/1/test-stdout.txt": "f23f9656c0e03ae2f2a33a6fa7d4d3c09a6cca9f12c8275c3d9f7fd4c3cd2632",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/verifier/ctrf.json": "476de5d9485e1e00b2ad4be31fce52726cd533e914d84678ddaad969b354186d",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/verifier/retry-summary.json": "81defcbe259dbe4add5d866536697ec0487aa00ab1fec5c7b47b044779ee90ae",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/verifier/reward.txt": "4355a46b19d348dc2f57c046f8ef63d4538ebb936000f3c9ee954a27460dd865",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/verifier/test-stdout.txt": "f23f9656c0e03ae2f2a33a6fa7d4d3c09a6cca9f12c8275c3d9f7fd4c3cd2632"
+        }
+      },
+      {
+        "task": "terminal-bench/torch-pipeline-parallelism",
+        "trial": "torch-pipeline-parallelism__q96EdZw",
+        "finished_at": "2026-10-18T14:09:27.485651Z",
+        "model_started": true,
+        "trajectory_sha256": "b62861207ac59c1440f3f53327dd0807d0c285a6880b38e82b255b1c4b95358e",
+        "raw_rewards": {
+          "reward": 0.0
+        },
+        "exception_type": null,
+        "files": {
+          "jobs/deepseek-v41-flash-1017-rerun-20261018-v2-orchestration/flash1017-torch-pipeline-parallelism-20261018.json": "a831faf281a4035fb80fda236ebc3e9f4cf2ad969ecf7026c88085f7b529118a",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/agent/bootstrap.json": "fcebfd336f1b466235db7c3fe8ba5ab12f89adf5748f461affa9f8ab75ebc7a6",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/agent/nanopycodeagent.txt": "8bcd1e4a7a01d1532a4c0fd81d1fdef340aea3a3b1255ba68b156fb2f43326df",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/agent/trajectory.json": "b62861207ac59c1440f3f53327dd0807d0c285a6880b38e82b255b1c4b95358e",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/artifacts/manifest.json": "c1db653f487fb773c4f9247162ab3314e7f0584a486914b9874c4b3d6949e42d",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/config.json": "1d2e5b78495ca7eb5c7d0d0b1dc1026610467f7b146492f767c9b52ad63d05b7",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/lock.json": "9a4647a5911abb58b20f8de8f23b95c10af20241eb50a33e188bb1af430afed1",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/result.json": "0a4eb83eb1dca3a6bbcfd74cdf3865e36d7ebcd113662394522b73e36b5dd9d0",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/trial.log": "0da597fa9cfbd01ca30021dd0b518d99d7ea6615bcf88d03a6f10bb0d3236e24",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/verifier/attempts/1/reward.txt": "9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/verifier/attempts/1/test-stdout.txt": "553b6078d900c0f346061308b0cac3010f0e4231fa68df5741422951d2b55293",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/verifier/retry-summary.json": "ca8c4424c03a7ef529e0e8076622048fd016df1b4a3ead75d7faf94fab6f6f2f",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/verifier/reward.txt": "9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/verifier/test-stdout.txt": "553b6078d900c0f346061308b0cac3010f0e4231fa68df5741422951d2b55293"
+        }
+      },
+      {
+        "task": "terminal-bench/regex-chess",
+        "trial": "regex-chess__BbMPdq7",
+        "finished_at": "2026-10-18T14:08:19.573882Z",
+        "model_started": true,
+        "trajectory_sha256": "c2c4272da108a1288c8ac0659c1fd4602bfddec009dd6394b4199085b72d95f5",
+        "raw_rewards": {
+          "reward": 0.0
+        },
+        "exception_type": null,
+        "files": {
+          "jobs/deepseek-v41-flash-1017-rerun-20261018-v2-orchestration/flash1017-regex-chess-20261018.json": "9257870036a57a2b99937edfcac5fceb3d5ccd020f6266f799c16fb0a3e64458",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/agent/bootstrap.json": "9abae99e4800f37de01cc8c2eb9d13d0437c60d0c1c0724f648d551c7f382b77",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/agent/nanopycodeagent.txt": "d5469e3e71ed9379e67ea53c31f938ac02e393bac9d262ed001e20edb0ce2880",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/agent/trajectory.json": "c2c4272da108a1288c8ac0659c1fd4602bfddec009dd6394b4199085b72d95f5",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/artifacts/manifest.json": "c1db653f487fb773c4f9247162ab3314e7f0584a486914b9874c4b3d6949e42d",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/config.json": "5646757537b53bd73c27d4bfb5976a7897750b8ff6c5ee8e10dc5258c3387084",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/lock.json": "25fb9f9020944b5515ff1abd1911a16dbbc6a297eca9974291c030301d73081b",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/result.json": "a77fe9b3aefafbd0898f517394289d60ece1c7d044d290eaa537e9a91dad28ae",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/trial.log": "ef21738d5435b0a881ea5555e88d382a711f8103dbd9e6e66f1d08c0d635ed14",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/verifier/attempts/1/ctrf.json": "1c6d5890a24a89e28193baa90caa33bcfc43aa63716511807d71df7b6152c12a",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/verifier/attempts/1/reward.txt": "9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/verifier/attempts/1/test-stdout.txt": "52c9df468950afcf550257d9866293ce7156da67c10d162f505cab682d599279",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/verifier/ctrf.json": "1c6d5890a24a89e28193baa90caa33bcfc43aa63716511807d71df7b6152c12a",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/verifier/retry-summary.json": "f43790b8b976ec87f3666a9e94bd4b06cf1a8557d8549eb9fc107a4cf5fa94f0",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/verifier/reward.txt": "9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa",
+          "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/verifier/test-stdout.txt": "52c9df468950afcf550257d9866293ce7156da67c10d162f505cab682d599279"
+        }
+      }
+    ]
+  },
+  "integrity": {
+    "registration_unchanged": true,
+    "plan_source_prompt_configs_tasks_unchanged": true,
+    "verified_preflight_tasks": 6,
+    "source_snapshots_verified": 6,
+    "wheel_digests_verified": 6,
+    "frozen_native_trial_artifacts_verified": 6,
+    "queue_pair_barriers_verified": true,
+    "whole_model_trial_retries": 0,
+    "outcome_based_replacements": 0
+  },
+  "aggregate": {
+    "trials": 6,
+    "models_started": 6,
+    "raw_passes": 3,
+    "raw_zeros": 3,
+    "classifications": {
+      "verifier_error": 2,
+      "scored_pass": 3,
+      "scored_failure": 1
+    },
+    "verifier_timeout_retries": 0,
+    "estimated_known_cost_usd": 0.374230884,
+    "cost_complete": true,
+    "usage_incomplete_tasks": [],
+    "recorded_tokens": {
+      "prompt_tokens": 2509747,
+      "completion_tokens": 288510,
+      "cached_tokens": 2465664
+    },
+    "summed_stage_seconds": {
+      "environment_setup": 12.846128,
+      "agent_setup": 227.443329,
+      "agent_execution": 2553.69249,
+      "verifier": 826.736545
+    },
+    "model_job_wall_seconds": 2404.499523,
+    "valid_scored_trials": 4,
+    "scored_passes": 3,
+    "scored_failures": 1,
+    "verifier_errors": 2,
+    "setup_errors_in_model_phase": 0,
+    "native_raw_zeros_include_verifier_errors": true,
+    "official_pytest_counts": {
+      "collected": 24,
+      "passed": 20,
+      "failed": 4,
+      "skipped": 0,
+      "error": 0
+    }
+  },
+  "trials": [
+    {
+      "order": 1,
+      "task": "terminal-bench/adaptive-rejection-sampler",
+      "task_ref": "sha256:bcaa2399985cd57666018025846289ab25e193ae0dd8fb7f0ffab2410c24d4de",
+      "trial": "adaptive-rejection-sampler__Bim6vqF",
+      "job": "flash1017-adaptive-rejection-sampler-20261018",
+      "raw_reward": 0.0,
+      "classification": "verifier_error",
+      "exception_type": null,
+      "official_test_summary": null,
+      "started_at": "2026-10-18T13:29:24.154763Z",
+      "finished_at": "2026-10-18T13:46:54.802088Z",
+      "seconds": {
+        "environment_setup": 2.099497,
+        "agent_setup": 37.036393,
+        "agent_execution": 627.580774,
+        "verifier": 362.621739
+      },
+      "work_budget_seconds": 720,
+      "verifier_timeout_seconds": 900.0,
+      "verifier_attempts": [
+        {
+          "attempt": 1,
+          "status": "completed",
+          "duration_seconds": 362.9866722779989,
+          "return_code": 0
+        }
+      ],
+      "verifier_dependency_profile": {
+        "profile": null,
+        "status": "not_configured"
+      },
+      "bootstrap_status": "ready",
+      "model_started": true,
+      "trajectory_terminal_outcome": "completed",
+      "recorded_agent_steps": 23,
+      "tool_calls": 23,
+      "incomplete_model_steps": [],
+      "accounting": {
+        "usage_complete": true,
+        "cost_complete": true,
+        "cost_is_estimated": true,
+        "recorded_tokens": {
+          "prompt_tokens": 1012326,
+          "completion_tokens": 79009,
+          "cached_tokens": 1002752
+        },
+        "estimated_known_cost_usd": 0.103699512
+      },
+      "task_checksum": "a4726253ad23c3fc46b5cb9c3d21025874e367b18643a3920f636a1fa89e1207",
+      "source_ref": "2eb99b608dfb39464d17d109a4a8f11afda607a9",
+      "source_sha256": {
+        "nanopycodeagent/__init__.py": "b922299f65b3db2fb1514df5aeda4413e599f595d2595f45a353b4dc551a1eb1",
+        "nanopycodeagent/agent.py": "aa58960803be77797c2addf65d7cc785a4be54f40ad58d53b9eea9fca1cf4b56",
+        "nanopycodeagent/atif.py": "ca4793b971d125d904888b858452db8db30e21f0d9b87818133062fb6cebe270",
+        "nanopycodeagent/bash_tool.py": "dcf0366fb9ebdca258d4c40b1a50b78b2ec79fc4b85cbb253ac9a9ce9534574f",
+        "nanopycodeagent/cli.py": "e335789b3f09440abd441547a2e3e2feb32b90c68e9ba5fc8237de8f6bac2932",
+        "nanopycodeagent/cost.py": "30a8df17942c424796e8532f45e1cb0784ed6541cea44284b863db3dae2f7b70",
+        "nanopycodeagent/deadline.py": "26a861721e6897a4e02e6ce364cb3ef1f1dc7d4ad9fb2f772404276f5c9e9d5f",
+        "nanopycodeagent/edit_tool.py": "78e2bec911279082b081366ab05cb4cbb37d8910b20b6245cccd4b0424489d0d",
+        "nanopycodeagent/event_journal.py": "8974919c2d460afb52bc57ec37e7e271851290c0845d01b4ad0f9f760cde0814",
+        "nanopycodeagent/read_tool.py": "e0075da19f2e1c6ccf34a3056a0aec92398e5afe67869937a096c6f1e04181c3",
+        "nanopycodeagent/settings.py": "491bf5c6eca50eb4c5fe04b7bfe62a8bfc7d2e1fad05b8eeac1680c2fcf3674e",
+        "nanopycodeagent/terminal.py": "8d27b40976b2b6d476e5106dfea982b055115435b4ba16b7f758412a05b75f04",
+        "nanopycodeagent/tool_validation.py": "3a1a85a7222dd1504bdf58f37e65a51d11de13d16c4e7df42c72b4cd58b5e5e5",
+        "nanopycodeagent/transport.py": "152d3ff78be0e3b2d92e0abbbfb652d73466aff57ad316c06dc0e072411167fe",
+        "nanopycodeagent/write_tool.py": "7671e83464fa1e942f25742fc573d0712e32053b5fc0291257884b027f7f6206",
+        "harbor_adapter/__init__.py": "513ad49c721b3e73293095c6a0486a791483e9aa2b51e0e9ad4b7816af07e76b",
+        "harbor_adapter/__main__.py": "db89d582fba45aad72147317f595531af49c434e3b23c65f22d7491f0b10551d",
+        "harbor_adapter/_verifier_runner.py": "2eb73c3e76d7360f2fa16018f0e933797bf982e589236095b5f98750f3807a7c",
+        "harbor_adapter/adapter.py": "d1aa154ec6972d706f0725cfa69c56635506d4499dff2a129e2c957ea7034c02",
+        "harbor_adapter/bootstrap.py": "f499bfddfc3d29338cae15a7f8f6e9c66d19eb69ad9577e607ced4b1aa057be2",
+        "harbor_adapter/verifier.py": "ec26fdefd69be1c5c450c9ba69d749d4a9d320a2a170b8db34421f0573d9b493"
+      },
+      "wheel_sha256": "f4dcc18758e584696f955ae57fd18c751599b70e8b497ca2db7ef9e2b16e9350",
+      "provenance_checks_passed": true,
+      "evidence_sha256": {
+        "jobs/deepseek-v41-flash-1017-rerun-20261018-v2-orchestration/flash1017-adaptive-rejection-sampler-20261018.json": "de95b04d9d0e73c76de3a36e507d39fc2f6e9ef4e6adbd873fa6fee4b8939785",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/agent/bootstrap.json": "fcebfd336f1b466235db7c3fe8ba5ab12f89adf5748f461affa9f8ab75ebc7a6",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/agent/nanopycodeagent.txt": "ef9164ce3172461c2a71bcb8f9c8ea1068131693c51daf293576bc16efba93e7",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/agent/trajectory.json": "df034659657412897dda47cec614b6d0b803904e5a8651e72eded8b1183aeb34",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/artifacts/manifest.json": "c1db653f487fb773c4f9247162ab3314e7f0584a486914b9874c4b3d6949e42d",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/config.json": "b972b206adf474688083b92425631dd1c70300b01690c24e99d7f7e0134c965b",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/lock.json": "1792b7a3b658f4cf272e44f902f584e9ad9b4070ef331b41e03bf8e26152d27e",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/result.json": "c57278943fe30ca5f82d9cf68ffa0ee0bdda61612b4e8e9a3ac4b278d8b19337",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/trial.log": "5f1050a5bc4629de43f0dfbfb4df3e53486deaea3acb4fac98e366700e89719b",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/verifier/attempts/1/reward.txt": "9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/verifier/attempts/1/test-stdout.txt": "9910c1025fef05f9ee04ff45a0364ffc8d88c7a159ad65b40992444fbbd164ad",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/verifier/retry-summary.json": "b70d48f409e1b4b4e89e921302518c03904dcb6304c481ea4054a8b0b9bdaa89",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/verifier/reward.txt": "9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018/adaptive-rejection-sampler__Bim6vqF/verifier/test-stdout.txt": "9910c1025fef05f9ee04ff45a0364ffc8d88c7a159ad65b40992444fbbd164ad",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018-input/manifest.json": "0e088e72039a3c6ab83f477fdb17fe08df4866f9269bef6ad1d1cab289f60177",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018-input/config.json": "0116a0a6f7072afe48d16bf7043c152c10126872eddc33606c7214956d16bc69",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-adaptive-rejection-sampler-20261018-input/working-tree.patch": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
+      },
+      "previous_baseline": {
+        "task": "terminal-bench/adaptive-rejection-sampler",
+        "trial": "adaptive-rejection-sampler__zWw9zAU",
+        "job": "dev34-adaptive-rejection-sampler-main-20261012",
+        "raw_reward": 0.0,
+        "classification": "scored_failure"
+      },
+      "official_pytest_summary": {
+        "executed": false,
+        "collected": 0,
+        "passed": 0,
+        "failed": 0,
+        "skipped": 0,
+        "error": 0,
+        "summary_line": null,
+        "source": "verifier/test-stdout.txt"
+      },
+      "audit": {
+        "benchmark_material_exposure": "not_observed_in_review",
+        "credential_observation_steps": []
+      },
+      "score": null,
+      "infrastructure_failure": {
+        "stage": "verifier_dependency_installation",
+        "kind": "download_timeout",
+        "dependency": "numpy==2.3.3",
+        "official_assertions_started": false,
+        "native_reward_preserved": 0.0,
+        "native_exception_type": null,
+        "retry_eligible": false,
+        "reason": "Dependency download failed before pytest started. The original wrapper returned reward zero without a Harbor exception. The verifier attempt itself did not time out, so timeout-only retry policy does not permit another attempt."
+      },
+      "agent_output_note": "The model submitted /app/ars.R and ran its own R tests; official correctness remains unverified."
+    },
+    {
+      "order": 2,
+      "task": "terminal-bench/db-wal-recovery",
+      "task_ref": "sha256:07e9193513bfce038b9f9d84962ecc85683eee0a642d14ee6fb7104b086ddab4",
+      "trial": "db-wal-recovery__iXi6Qjy",
+      "job": "flash1017-db-wal-recovery-20261018",
+      "raw_reward": 1.0,
+      "classification": "scored_pass",
+      "exception_type": null,
+      "official_test_summary": {
+        "tests": 7,
+        "passed": 7,
+        "failed": 0,
+        "skipped": 0,
+        "pending": 0,
+        "other": 0,
+        "start": 1792330293.4841444,
+        "stop": 1792330293.5027974
+      },
+      "started_at": "2026-10-18T13:29:22.986128Z",
+      "finished_at": "2026-10-18T13:31:54.497971Z",
+      "seconds": {
+        "environment_setup": 2.229671,
+        "agent_setup": 35.811538,
+        "agent_execution": 76.170112,
+        "verifier": 16.120535
+      },
+      "work_budget_seconds": 720,
+      "verifier_timeout_seconds": 900.0,
+      "verifier_attempts": [
+        {
+          "attempt": 1,
+          "status": "completed",
+          "duration_seconds": 15.898172123997938,
+          "return_code": 0
+        }
+      ],
+      "verifier_dependency_profile": {
+        "profile": null,
+        "status": "not_configured"
+      },
+      "bootstrap_status": "ready",
+      "model_started": true,
+      "trajectory_terminal_outcome": "completed",
+      "recorded_agent_steps": 14,
+      "tool_calls": 20,
+      "incomplete_model_steps": [],
+      "accounting": {
+        "usage_complete": true,
+        "cost_complete": true,
+        "cost_is_estimated": true,
+        "recorded_tokens": {
+          "prompt_tokens": 178023,
+          "completion_tokens": 13271,
+          "cached_tokens": 167424
+        },
+        "estimated_known_cost_usd": 0.020109444
+      },
+      "task_checksum": "c18abdc4fdc3a01bf374c55a9700708fe6a9662077d29db81abb692f0a3c5f6f",
+      "source_ref": "2eb99b608dfb39464d17d109a4a8f11afda607a9",
+      "source_sha256": {
+        "nanopycodeagent/__init__.py": "b922299f65b3db2fb1514df5aeda4413e599f595d2595f45a353b4dc551a1eb1",
+        "nanopycodeagent/agent.py": "aa58960803be77797c2addf65d7cc785a4be54f40ad58d53b9eea9fca1cf4b56",
+        "nanopycodeagent/atif.py": "ca4793b971d125d904888b858452db8db30e21f0d9b87818133062fb6cebe270",
+        "nanopycodeagent/bash_tool.py": "dcf0366fb9ebdca258d4c40b1a50b78b2ec79fc4b85cbb253ac9a9ce9534574f",
+        "nanopycodeagent/cli.py": "e335789b3f09440abd441547a2e3e2feb32b90c68e9ba5fc8237de8f6bac2932",
+        "nanopycodeagent/cost.py": "30a8df17942c424796e8532f45e1cb0784ed6541cea44284b863db3dae2f7b70",
+        "nanopycodeagent/deadline.py": "26a861721e6897a4e02e6ce364cb3ef1f1dc7d4ad9fb2f772404276f5c9e9d5f",
+        "nanopycodeagent/edit_tool.py": "78e2bec911279082b081366ab05cb4cbb37d8910b20b6245cccd4b0424489d0d",
+        "nanopycodeagent/event_journal.py": "8974919c2d460afb52bc57ec37e7e271851290c0845d01b4ad0f9f760cde0814",
+        "nanopycodeagent/read_tool.py": "e0075da19f2e1c6ccf34a3056a0aec92398e5afe67869937a096c6f1e04181c3",
+        "nanopycodeagent/settings.py": "491bf5c6eca50eb4c5fe04b7bfe62a8bfc7d2e1fad05b8eeac1680c2fcf3674e",
+        "nanopycodeagent/terminal.py": "8d27b40976b2b6d476e5106dfea982b055115435b4ba16b7f758412a05b75f04",
+        "nanopycodeagent/tool_validation.py": "3a1a85a7222dd1504bdf58f37e65a51d11de13d16c4e7df42c72b4cd58b5e5e5",
+        "nanopycodeagent/transport.py": "152d3ff78be0e3b2d92e0abbbfb652d73466aff57ad316c06dc0e072411167fe",
+        "nanopycodeagent/write_tool.py": "7671e83464fa1e942f25742fc573d0712e32053b5fc0291257884b027f7f6206",
+        "harbor_adapter/__init__.py": "513ad49c721b3e73293095c6a0486a791483e9aa2b51e0e9ad4b7816af07e76b",
+        "harbor_adapter/__main__.py": "db89d582fba45aad72147317f595531af49c434e3b23c65f22d7491f0b10551d",
+        "harbor_adapter/_verifier_runner.py": "2eb73c3e76d7360f2fa16018f0e933797bf982e589236095b5f98750f3807a7c",
+        "harbor_adapter/adapter.py": "d1aa154ec6972d706f0725cfa69c56635506d4499dff2a129e2c957ea7034c02",
+        "harbor_adapter/bootstrap.py": "f499bfddfc3d29338cae15a7f8f6e9c66d19eb69ad9577e607ced4b1aa057be2",
+        "harbor_adapter/verifier.py": "ec26fdefd69be1c5c450c9ba69d749d4a9d320a2a170b8db34421f0573d9b493"
+      },
+      "wheel_sha256": "f4dcc18758e584696f955ae57fd18c751599b70e8b497ca2db7ef9e2b16e9350",
+      "provenance_checks_passed": true,
+      "evidence_sha256": {
+        "jobs/deepseek-v41-flash-1017-rerun-20261018-v2-orchestration/flash1017-db-wal-recovery-20261018.json": "aba7a7e2b507d135aad4cc372617e704dd533e1317e98199ab7c1bf5f92c493e",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/agent/bootstrap.json": "51366101df4e7f6425f8fe241b10713287e6f89274173cba25c6ffe708d61dc3",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/agent/nanopycodeagent.txt": "1e5138fd52619c17dc98a132a6b3f5955d05fc9792cc57a2bc1150a3e80ee9cb",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/agent/trajectory.json": "cc05c9b3c5f217e4067348143b8824495029f8394d450b92273fde2d839bad19",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/artifacts/manifest.json": "c1db653f487fb773c4f9247162ab3314e7f0584a486914b9874c4b3d6949e42d",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/config.json": "d88662c669c196d97b3c77a39bbd7d6a10313757b46860feb2258dc4cd40a7c2",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/lock.json": "2d54030d3f9476b33afbe16bf6395230b174e3b248036e76b04fb7ef192b5834",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/result.json": "0a0bfb5de1b212e41bb17b328994f31208ce0eb587151267ca5686b412adb86b",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/trial.log": "331fe02968e34e1c0cecfbf532a32dc552ccdf4232a28c9bc0a0cb266d41c5eb",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/verifier/attempts/1/ctrf.json": "c276c93a192d5ea8078837a7adce56070d661dc2b9cfefe18d63e0ea4746a365",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/verifier/attempts/1/reward.txt": "4355a46b19d348dc2f57c046f8ef63d4538ebb936000f3c9ee954a27460dd865",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/verifier/attempts/1/test-stdout.txt": "f6502d849c96dabb43cda87d0b8d453fb11f08d3a7e28137fbc0c25c4baa030b",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/verifier/ctrf.json": "c276c93a192d5ea8078837a7adce56070d661dc2b9cfefe18d63e0ea4746a365",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/verifier/retry-summary.json": "917b2a0fdfccb2017728dc353860862d45fe1bc59c97fa2547d20eb371fa7656",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/verifier/reward.txt": "4355a46b19d348dc2f57c046f8ef63d4538ebb936000f3c9ee954a27460dd865",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018/db-wal-recovery__iXi6Qjy/verifier/test-stdout.txt": "f6502d849c96dabb43cda87d0b8d453fb11f08d3a7e28137fbc0c25c4baa030b",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018-input/manifest.json": "f0274834b2015c69c3ed58125e05b21e9e93eabcd1f5ddb73e074b3a33cbbc03",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018-input/config.json": "b1f07d918c5026c0567b948a886636c68078e8757f1fd234b6467328a920a0f3",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-db-wal-recovery-20261018-input/working-tree.patch": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
+      },
+      "previous_baseline": {
+        "task": "terminal-bench/db-wal-recovery",
+        "trial": "db-wal-recovery__BCqJJMc",
+        "job": "dev34-db-wal-recovery-main-20261012",
+        "raw_reward": 0.0,
+        "classification": "scored_failure"
+      },
+      "official_pytest_summary": {
+        "executed": true,
+        "collected": 7,
+        "passed": 7,
+        "failed": 0,
+        "skipped": 0,
+        "error": 0,
+        "summary_line": "============================== 7 passed in 0.02s ===============================",
+        "source": "verifier/test-stdout.txt"
+      },
+      "audit": {
+        "benchmark_material_exposure": "not_observed_in_review",
+        "credential_observation_steps": []
+      },
+      "score": 1.0
+    },
+    {
+      "order": 3,
+      "task": "terminal-bench/kv-store-grpc",
+      "task_ref": "sha256:973c5d4c111fb61a344457936f1c36400acd2d9e44389e7b319586fe23a7a307",
+      "trial": "kv-store-grpc__4TFUA4j",
+      "job": "flash1017-kv-store-grpc-20261018",
+      "raw_reward": 1.0,
+      "classification": "scored_pass",
+      "exception_type": null,
+      "official_test_summary": {
+        "tests": 7,
+        "passed": 7,
+        "failed": 0,
+        "skipped": 0,
+        "pending": 0,
+        "other": 0,
+        "start": 1792331645.8044658,
+        "stop": 1792331645.8659093
+      },
+      "started_at": "2026-10-18T13:47:03.465276Z",
+      "finished_at": "2026-10-18T13:54:26.858875Z",
+      "seconds": {
+        "environment_setup": 2.121689,
+        "agent_setup": 43.168861,
+        "agent_execution": 304.229454,
+        "verifier": 72.677772
+      },
+      "work_budget_seconds": 720,
+      "verifier_timeout_seconds": 900.0,
+      "verifier_attempts": [
+        {
+          "attempt": 1,
+          "status": "completed",
+          "duration_seconds": 72.45762456799275,
+          "return_code": 0
+        }
+      ],
+      "verifier_dependency_profile": {
+        "profile": null,
+        "status": "not_configured"
+      },
+      "bootstrap_status": "ready",
+      "model_started": true,
+      "trajectory_terminal_outcome": "completed",
+      "recorded_agent_steps": 16,
+      "tool_calls": 17,
+      "incomplete_model_steps": [],
+      "accounting": {
+        "usage_complete": true,
+        "cost_complete": true,
+        "cost_is_estimated": true,
+        "recorded_tokens": {
+          "prompt_tokens": 107418,
+          "completion_tokens": 5776,
+          "cached_tokens": 101504
+        },
+        "estimated_known_cost_usd": 0.009314424
+      },
+      "task_checksum": "901a4dd5c3078d5155043875a0d07f3583f462c73f919b52855c025e680b1edd",
+      "source_ref": "2eb99b608dfb39464d17d109a4a8f11afda607a9",
+      "source_sha256": {
+        "nanopycodeagent/__init__.py": "b922299f65b3db2fb1514df5aeda4413e599f595d2595f45a353b4dc551a1eb1",
+        "nanopycodeagent/agent.py": "aa58960803be77797c2addf65d7cc785a4be54f40ad58d53b9eea9fca1cf4b56",
+        "nanopycodeagent/atif.py": "ca4793b971d125d904888b858452db8db30e21f0d9b87818133062fb6cebe270",
+        "nanopycodeagent/bash_tool.py": "dcf0366fb9ebdca258d4c40b1a50b78b2ec79fc4b85cbb253ac9a9ce9534574f",
+        "nanopycodeagent/cli.py": "e335789b3f09440abd441547a2e3e2feb32b90c68e9ba5fc8237de8f6bac2932",
+        "nanopycodeagent/cost.py": "30a8df17942c424796e8532f45e1cb0784ed6541cea44284b863db3dae2f7b70",
+        "nanopycodeagent/deadline.py": "26a861721e6897a4e02e6ce364cb3ef1f1dc7d4ad9fb2f772404276f5c9e9d5f",
+        "nanopycodeagent/edit_tool.py": "78e2bec911279082b081366ab05cb4cbb37d8910b20b6245cccd4b0424489d0d",
+        "nanopycodeagent/event_journal.py": "8974919c2d460afb52bc57ec37e7e271851290c0845d01b4ad0f9f760cde0814",
+        "nanopycodeagent/read_tool.py": "e0075da19f2e1c6ccf34a3056a0aec92398e5afe67869937a096c6f1e04181c3",
+        "nanopycodeagent/settings.py": "491bf5c6eca50eb4c5fe04b7bfe62a8bfc7d2e1fad05b8eeac1680c2fcf3674e",
+        "nanopycodeagent/terminal.py": "8d27b40976b2b6d476e5106dfea982b055115435b4ba16b7f758412a05b75f04",
+        "nanopycodeagent/tool_validation.py": "3a1a85a7222dd1504bdf58f37e65a51d11de13d16c4e7df42c72b4cd58b5e5e5",
+        "nanopycodeagent/transport.py": "152d3ff78be0e3b2d92e0abbbfb652d73466aff57ad316c06dc0e072411167fe",
+        "nanopycodeagent/write_tool.py": "7671e83464fa1e942f25742fc573d0712e32053b5fc0291257884b027f7f6206",
+        "harbor_adapter/__init__.py": "513ad49c721b3e73293095c6a0486a791483e9aa2b51e0e9ad4b7816af07e76b",
+        "harbor_adapter/__main__.py": "db89d582fba45aad72147317f595531af49c434e3b23c65f22d7491f0b10551d",
+        "harbor_adapter/_verifier_runner.py": "2eb73c3e76d7360f2fa16018f0e933797bf982e589236095b5f98750f3807a7c",
+        "harbor_adapter/adapter.py": "d1aa154ec6972d706f0725cfa69c56635506d4499dff2a129e2c957ea7034c02",
+        "harbor_adapter/bootstrap.py": "f499bfddfc3d29338cae15a7f8f6e9c66d19eb69ad9577e607ced4b1aa057be2",
+        "harbor_adapter/verifier.py": "ec26fdefd69be1c5c450c9ba69d749d4a9d320a2a170b8db34421f0573d9b493"
+      },
+      "wheel_sha256": "f4dcc18758e584696f955ae57fd18c751599b70e8b497ca2db7ef9e2b16e9350",
+      "provenance_checks_passed": true,
+      "evidence_sha256": {
+        "jobs/deepseek-v41-flash-1017-rerun-20261018-v2-orchestration/flash1017-kv-store-grpc-20261018.json": "07372fed21de0bfce73dee8b5a71a592be1ae2c5b8ad1db78b5382d2a42cd2a0",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/agent/bootstrap.json": "9abae99e4800f37de01cc8c2eb9d13d0437c60d0c1c0724f648d551c7f382b77",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/agent/nanopycodeagent.txt": "bbabb30397f5ef5da2ea434aca3c1ed9d5ebaa39583585e73cc030801faa315e",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/agent/trajectory.json": "71d056ce715961081a775ca4461ae1e6cc278fb124dedcea3261e410f7da9264",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/artifacts/manifest.json": "c1db653f487fb773c4f9247162ab3314e7f0584a486914b9874c4b3d6949e42d",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/config.json": "93a23d07f2b21646fab7cf47ca38b5fa46c1f58f445e82360d3dadc9c3eedb64",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/lock.json": "3b059c7147009b9eb88fa82bc54c4dd131b04163255a02742140b95db77ca9a7",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/result.json": "81a490fc5dc6b51fc898db40a6f6d678d01575f9da353baf3ffa012c400cfd42",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/trial.log": "786c8e5d96f1ff6e98e74b79d218027a9cd136a89a5ca65e211951fad0cf74a8",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/verifier/attempts/1/ctrf.json": "d4239ddd3b1d9eec2aac414a8fd6cace832f0c854e6a09f36c67ac7d269a1e01",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/verifier/attempts/1/reward.txt": "4355a46b19d348dc2f57c046f8ef63d4538ebb936000f3c9ee954a27460dd865",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/verifier/attempts/1/test-stdout.txt": "bc35e4c8cb26c8751d4f863cc24d30bab201e4eb17ddb3ff2e4bb5a826bc8986",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/verifier/ctrf.json": "d4239ddd3b1d9eec2aac414a8fd6cace832f0c854e6a09f36c67ac7d269a1e01",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/verifier/retry-summary.json": "3fc4da194d0fdde7d6e00f0eb458186f5e5746eb62b0bce74d9734403e513e3d",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/verifier/reward.txt": "4355a46b19d348dc2f57c046f8ef63d4538ebb936000f3c9ee954a27460dd865",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018/kv-store-grpc__4TFUA4j/verifier/test-stdout.txt": "bc35e4c8cb26c8751d4f863cc24d30bab201e4eb17ddb3ff2e4bb5a826bc8986",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018-input/manifest.json": "61af435e1c0f1dc1c1eafc669ffb25f9f536e3dd96374ea8ff5a77266a0f3a01",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018-input/config.json": "c386524cf52cb451ba0ec07ef1108e69bab02fea5d52006758a3b15bf448348d",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-kv-store-grpc-20261018-input/working-tree.patch": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
+      },
+      "previous_baseline": {
+        "task": "terminal-bench/kv-store-grpc",
+        "trial": "kv-store-grpc__pUAoWuK",
+        "job": "dev34-kv-store-grpc-main-20261012",
+        "raw_reward": 0.0,
+        "classification": "scored_failure"
+      },
+      "official_pytest_summary": {
+        "executed": true,
+        "collected": 7,
+        "passed": 7,
+        "failed": 0,
+        "skipped": 0,
+        "error": 0,
+        "summary_line": "============================== 7 passed in 0.06s ===============================",
+        "source": "verifier/test-stdout.txt"
+      },
+      "audit": {
+        "benchmark_material_exposure": "not_observed_in_review",
+        "credential_observation_steps": [],
+        "reviewed_local_access": {
+          "atif_steps": [
+            13
+          ],
+          "finding": "The /tests substring occurs in an agent-authored server comment (clients/tests); it is not a command reading hidden verifier files."
+        }
+      },
+      "score": 1.0
+    },
+    {
+      "order": 4,
+      "task": "terminal-bench/query-optimize",
+      "task_ref": "sha256:169496ea6843cb403b0860132675baf7aa0df0ac8b86221068e27d86395260f4",
+      "trial": "query-optimize__ioRN6AM",
+      "job": "flash1017-query-optimize-20261018",
+      "raw_reward": 1.0,
+      "classification": "scored_pass",
+      "exception_type": null,
+      "official_test_summary": {
+        "tests": 6,
+        "passed": 6,
+        "failed": 0,
+        "skipped": 0,
+        "pending": 0,
+        "other": 0,
+        "start": 1792331571.5115156,
+        "stop": 1792331706.4421964
+      },
+      "started_at": "2026-10-18T13:47:02.874010Z",
+      "finished_at": "2026-10-18T13:55:27.515015Z",
+      "seconds": {
+        "environment_setup": 2.204782,
+        "agent_setup": 38.606783,
+        "agent_execution": 287.02429,
+        "verifier": 155.544866
+      },
+      "work_budget_seconds": 720,
+      "verifier_timeout_seconds": 1800.0,
+      "verifier_attempts": [
+        {
+          "attempt": 1,
+          "status": "completed",
+          "duration_seconds": 154.93728369200835,
+          "return_code": 0
+        }
+      ],
+      "verifier_dependency_profile": {
+        "profile": null,
+        "status": "not_configured"
+      },
+      "bootstrap_status": "ready",
+      "model_started": true,
+      "trajectory_terminal_outcome": "completed",
+      "recorded_agent_steps": 16,
+      "tool_calls": 22,
+      "incomplete_model_steps": [],
+      "accounting": {
+        "usage_complete": true,
+        "cost_complete": true,
+        "cost_is_estimated": true,
+        "recorded_tokens": {
+          "prompt_tokens": 196642,
+          "completion_tokens": 14949,
+          "cached_tokens": 190592
+        },
+        "estimated_known_cost_usd": 0.020897352
+      },
+      "task_checksum": "593405b6a2f6970f0f4c2ca2aff455a3b32bce3eecf8b506cfeeb0fe93b14838",
+      "source_ref": "2eb99b608dfb39464d17d109a4a8f11afda607a9",
+      "source_sha256": {
+        "nanopycodeagent/__init__.py": "b922299f65b3db2fb1514df5aeda4413e599f595d2595f45a353b4dc551a1eb1",
+        "nanopycodeagent/agent.py": "aa58960803be77797c2addf65d7cc785a4be54f40ad58d53b9eea9fca1cf4b56",
+        "nanopycodeagent/atif.py": "ca4793b971d125d904888b858452db8db30e21f0d9b87818133062fb6cebe270",
+        "nanopycodeagent/bash_tool.py": "dcf0366fb9ebdca258d4c40b1a50b78b2ec79fc4b85cbb253ac9a9ce9534574f",
+        "nanopycodeagent/cli.py": "e335789b3f09440abd441547a2e3e2feb32b90c68e9ba5fc8237de8f6bac2932",
+        "nanopycodeagent/cost.py": "30a8df17942c424796e8532f45e1cb0784ed6541cea44284b863db3dae2f7b70",
+        "nanopycodeagent/deadline.py": "26a861721e6897a4e02e6ce364cb3ef1f1dc7d4ad9fb2f772404276f5c9e9d5f",
+        "nanopycodeagent/edit_tool.py": "78e2bec911279082b081366ab05cb4cbb37d8910b20b6245cccd4b0424489d0d",
+        "nanopycodeagent/event_journal.py": "8974919c2d460afb52bc57ec37e7e271851290c0845d01b4ad0f9f760cde0814",
+        "nanopycodeagent/read_tool.py": "e0075da19f2e1c6ccf34a3056a0aec92398e5afe67869937a096c6f1e04181c3",
+        "nanopycodeagent/settings.py": "491bf5c6eca50eb4c5fe04b7bfe62a8bfc7d2e1fad05b8eeac1680c2fcf3674e",
+        "nanopycodeagent/terminal.py": "8d27b40976b2b6d476e5106dfea982b055115435b4ba16b7f758412a05b75f04",
+        "nanopycodeagent/tool_validation.py": "3a1a85a7222dd1504bdf58f37e65a51d11de13d16c4e7df42c72b4cd58b5e5e5",
+        "nanopycodeagent/transport.py": "152d3ff78be0e3b2d92e0abbbfb652d73466aff57ad316c06dc0e072411167fe",
+        "nanopycodeagent/write_tool.py": "7671e83464fa1e942f25742fc573d0712e32053b5fc0291257884b027f7f6206",
+        "harbor_adapter/__init__.py": "513ad49c721b3e73293095c6a0486a791483e9aa2b51e0e9ad4b7816af07e76b",
+        "harbor_adapter/__main__.py": "db89d582fba45aad72147317f595531af49c434e3b23c65f22d7491f0b10551d",
+        "harbor_adapter/_verifier_runner.py": "2eb73c3e76d7360f2fa16018f0e933797bf982e589236095b5f98750f3807a7c",
+        "harbor_adapter/adapter.py": "d1aa154ec6972d706f0725cfa69c56635506d4499dff2a129e2c957ea7034c02",
+        "harbor_adapter/bootstrap.py": "f499bfddfc3d29338cae15a7f8f6e9c66d19eb69ad9577e607ced4b1aa057be2",
+        "harbor_adapter/verifier.py": "ec26fdefd69be1c5c450c9ba69d749d4a9d320a2a170b8db34421f0573d9b493"
+      },
+      "wheel_sha256": "f4dcc18758e584696f955ae57fd18c751599b70e8b497ca2db7ef9e2b16e9350",
+      "provenance_checks_passed": true,
+      "evidence_sha256": {
+        "jobs/deepseek-v41-flash-1017-rerun-20261018-v2-orchestration/flash1017-query-optimize-20261018.json": "ac61d4144c34cd9e5b698332360a09d161d53213ffdfab0b11e955ef6807fdb1",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/agent/bootstrap.json": "4a93e54276aa8f2cc86d505dad8f483806c983d5c140d69bb855ed60c1cd32b6",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/agent/nanopycodeagent.txt": "65c651a52e18f1354c47b506be980d39cd2ce8c0660a8000b97c1404a74fa435",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/agent/trajectory.json": "126b264a80801d55a7546d651b619c8b35d3b19c05a1639b3b01c1a4478e6094",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/artifacts/manifest.json": "c1db653f487fb773c4f9247162ab3314e7f0584a486914b9874c4b3d6949e42d",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/config.json": "099122689c4c6a00d300285a2efe5ec43259762ed8bc106988dc3327fc162563",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/lock.json": "53d10aefdfebbcb1070f1c9fbbbdb83c96c73d870a6a76db1653dfe25d0fc1be",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/result.json": "4c67b5595e3129d4ac1cabe5c7b4765fe4dad0c1c40a472db78e232d102d327b",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/trial.log": "5bd01f98bd6b0cd4684ced0582ec10d90c9456f262e332feb01fc3690a4f43f6",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/verifier/attempts/1/ctrf.json": "476de5d9485e1e00b2ad4be31fce52726cd533e914d84678ddaad969b354186d",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/verifier/attempts/1/reward.txt": "4355a46b19d348dc2f57c046f8ef63d4538ebb936000f3c9ee954a27460dd865",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/verifier/attempts/1/test-stdout.txt": "f23f9656c0e03ae2f2a33a6fa7d4d3c09a6cca9f12c8275c3d9f7fd4c3cd2632",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/verifier/ctrf.json": "476de5d9485e1e00b2ad4be31fce52726cd533e914d84678ddaad969b354186d",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/verifier/retry-summary.json": "81defcbe259dbe4add5d866536697ec0487aa00ab1fec5c7b47b044779ee90ae",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/verifier/reward.txt": "4355a46b19d348dc2f57c046f8ef63d4538ebb936000f3c9ee954a27460dd865",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018/query-optimize__ioRN6AM/verifier/test-stdout.txt": "f23f9656c0e03ae2f2a33a6fa7d4d3c09a6cca9f12c8275c3d9f7fd4c3cd2632",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018-input/manifest.json": "2a105f512b3346a6f23a8d3c7565c5df4a7ee95d912c555f44c44109cdfaf349",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018-input/config.json": "6ecd9cc1fc7d85410fe28fe036394b7fe96a03f41c3869542bba4dfd1f2f715f",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-query-optimize-20261018-input/working-tree.patch": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
+      },
+      "previous_baseline": {
+        "task": "terminal-bench/query-optimize",
+        "trial": "query-optimize__ZVLSEuZ",
+        "job": "dev34-query-optimize-main-20261012",
+        "raw_reward": 0.0,
+        "classification": "scored_failure"
+      },
+      "official_pytest_summary": {
+        "executed": true,
+        "collected": 6,
+        "passed": 6,
+        "failed": 0,
+        "skipped": 0,
+        "error": 0,
+        "summary_line": "======================== 6 passed in 134.51s (0:02:14) =========================",
+        "source": "verifier/test-stdout.txt"
+      },
+      "audit": {
+        "benchmark_material_exposure": "not_observed_in_review",
+        "credential_observation_steps": []
+      },
+      "score": 1.0
+    },
+    {
+      "order": 5,
+      "task": "terminal-bench/torch-pipeline-parallelism",
+      "task_ref": "sha256:db605337c749a872cea7b5b413429b3915bb4c3efe0f7875f0c46ce81bd8c4fb",
+      "trial": "torch-pipeline-parallelism__q96EdZw",
+      "job": "flash1017-torch-pipeline-parallelism-20261018",
+      "raw_reward": 0.0,
+      "classification": "verifier_error",
+      "exception_type": null,
+      "official_test_summary": null,
+      "started_at": "2026-10-18T13:55:34.684057Z",
+      "finished_at": "2026-10-18T14:09:27.485651Z",
+      "seconds": {
+        "environment_setup": 2.04244,
+        "agent_setup": 35.133596,
+        "agent_execution": 613.924691,
+        "verifier": 160.443647
+      },
+      "work_budget_seconds": 720,
+      "verifier_timeout_seconds": 900.0,
+      "verifier_attempts": [
+        {
+          "attempt": 1,
+          "status": "completed",
+          "duration_seconds": 159.8422649980057,
+          "return_code": 0
+        }
+      ],
+      "verifier_dependency_profile": {
+        "profile": null,
+        "status": "not_configured"
+      },
+      "bootstrap_status": "ready",
+      "model_started": true,
+      "trajectory_terminal_outcome": "completed",
+      "recorded_agent_steps": 29,
+      "tool_calls": 31,
+      "incomplete_model_steps": [],
+      "accounting": {
+        "usage_complete": true,
+        "cost_complete": true,
+        "cost_is_estimated": true,
+        "recorded_tokens": {
+          "prompt_tokens": 1005856,
+          "completion_tokens": 44296,
+          "cached_tokens": 996224
+        },
+        "estimated_known_cost_usd": 0.062022144
+      },
+      "task_checksum": "31e565558cd22598a3c2069f771d9d0eee26fb843b77f7b03e00ad4c336fbce9",
+      "source_ref": "2eb99b608dfb39464d17d109a4a8f11afda607a9",
+      "source_sha256": {
+        "nanopycodeagent/__init__.py": "b922299f65b3db2fb1514df5aeda4413e599f595d2595f45a353b4dc551a1eb1",
+        "nanopycodeagent/agent.py": "aa58960803be77797c2addf65d7cc785a4be54f40ad58d53b9eea9fca1cf4b56",
+        "nanopycodeagent/atif.py": "ca4793b971d125d904888b858452db8db30e21f0d9b87818133062fb6cebe270",
+        "nanopycodeagent/bash_tool.py": "dcf0366fb9ebdca258d4c40b1a50b78b2ec79fc4b85cbb253ac9a9ce9534574f",
+        "nanopycodeagent/cli.py": "e335789b3f09440abd441547a2e3e2feb32b90c68e9ba5fc8237de8f6bac2932",
+        "nanopycodeagent/cost.py": "30a8df17942c424796e8532f45e1cb0784ed6541cea44284b863db3dae2f7b70",
+        "nanopycodeagent/deadline.py": "26a861721e6897a4e02e6ce364cb3ef1f1dc7d4ad9fb2f772404276f5c9e9d5f",
+        "nanopycodeagent/edit_tool.py": "78e2bec911279082b081366ab05cb4cbb37d8910b20b6245cccd4b0424489d0d",
+        "nanopycodeagent/event_journal.py": "8974919c2d460afb52bc57ec37e7e271851290c0845d01b4ad0f9f760cde0814",
+        "nanopycodeagent/read_tool.py": "e0075da19f2e1c6ccf34a3056a0aec92398e5afe67869937a096c6f1e04181c3",
+        "nanopycodeagent/settings.py": "491bf5c6eca50eb4c5fe04b7bfe62a8bfc7d2e1fad05b8eeac1680c2fcf3674e",
+        "nanopycodeagent/terminal.py": "8d27b40976b2b6d476e5106dfea982b055115435b4ba16b7f758412a05b75f04",
+        "nanopycodeagent/tool_validation.py": "3a1a85a7222dd1504bdf58f37e65a51d11de13d16c4e7df42c72b4cd58b5e5e5",
+        "nanopycodeagent/transport.py": "152d3ff78be0e3b2d92e0abbbfb652d73466aff57ad316c06dc0e072411167fe",
+        "nanopycodeagent/write_tool.py": "7671e83464fa1e942f25742fc573d0712e32053b5fc0291257884b027f7f6206",
+        "harbor_adapter/__init__.py": "513ad49c721b3e73293095c6a0486a791483e9aa2b51e0e9ad4b7816af07e76b",
+        "harbor_adapter/__main__.py": "db89d582fba45aad72147317f595531af49c434e3b23c65f22d7491f0b10551d",
+        "harbor_adapter/_verifier_runner.py": "2eb73c3e76d7360f2fa16018f0e933797bf982e589236095b5f98750f3807a7c",
+        "harbor_adapter/adapter.py": "d1aa154ec6972d706f0725cfa69c56635506d4499dff2a129e2c957ea7034c02",
+        "harbor_adapter/bootstrap.py": "f499bfddfc3d29338cae15a7f8f6e9c66d19eb69ad9577e607ced4b1aa057be2",
+        "harbor_adapter/verifier.py": "ec26fdefd69be1c5c450c9ba69d749d4a9d320a2a170b8db34421f0573d9b493"
+      },
+      "wheel_sha256": "f4dcc18758e584696f955ae57fd18c751599b70e8b497ca2db7ef9e2b16e9350",
+      "provenance_checks_passed": true,
+      "evidence_sha256": {
+        "jobs/deepseek-v41-flash-1017-rerun-20261018-v2-orchestration/flash1017-torch-pipeline-parallelism-20261018.json": "a831faf281a4035fb80fda236ebc3e9f4cf2ad969ecf7026c88085f7b529118a",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/agent/bootstrap.json": "fcebfd336f1b466235db7c3fe8ba5ab12f89adf5748f461affa9f8ab75ebc7a6",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/agent/nanopycodeagent.txt": "8bcd1e4a7a01d1532a4c0fd81d1fdef340aea3a3b1255ba68b156fb2f43326df",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/agent/trajectory.json": "b62861207ac59c1440f3f53327dd0807d0c285a6880b38e82b255b1c4b95358e",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/artifacts/manifest.json": "c1db653f487fb773c4f9247162ab3314e7f0584a486914b9874c4b3d6949e42d",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/config.json": "1d2e5b78495ca7eb5c7d0d0b1dc1026610467f7b146492f767c9b52ad63d05b7",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/lock.json": "9a4647a5911abb58b20f8de8f23b95c10af20241eb50a33e188bb1af430afed1",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/result.json": "0a4eb83eb1dca3a6bbcfd74cdf3865e36d7ebcd113662394522b73e36b5dd9d0",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/trial.log": "0da597fa9cfbd01ca30021dd0b518d99d7ea6615bcf88d03a6f10bb0d3236e24",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/verifier/attempts/1/reward.txt": "9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/verifier/attempts/1/test-stdout.txt": "553b6078d900c0f346061308b0cac3010f0e4231fa68df5741422951d2b55293",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/verifier/retry-summary.json": "ca8c4424c03a7ef529e0e8076622048fd016df1b4a3ead75d7faf94fab6f6f2f",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/verifier/reward.txt": "9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018/torch-pipeline-parallelism__q96EdZw/verifier/test-stdout.txt": "553b6078d900c0f346061308b0cac3010f0e4231fa68df5741422951d2b55293",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018-input/manifest.json": "0df1a9d2170d5420fa3bfa8c8d121e9decfe586d57f411153ad4c7579b9c5aee",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018-input/config.json": "1be7e64d299acf2f818001876b99e98f57fce1bacfcace6752f83f338e07b78f",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-torch-pipeline-parallelism-20261018-input/working-tree.patch": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
+      },
+      "previous_baseline": {
+        "task": "terminal-bench/torch-pipeline-parallelism",
+        "trial": "torch-pipeline-parallelism__Zr8yYNi",
+        "job": "dev34-torch-pipeline-parallelism-main-20261012",
+        "raw_reward": 0.0,
+        "classification": "scored_failure"
+      },
+      "official_pytest_summary": {
+        "executed": false,
+        "collected": 0,
+        "passed": 0,
+        "failed": 0,
+        "skipped": 0,
+        "error": 0,
+        "summary_line": null,
+        "source": "verifier/test-stdout.txt"
+      },
+      "audit": {
+        "benchmark_material_exposure": "not_observed_in_review",
+        "credential_observation_steps": [],
+        "reviewed_upstream_reference": {
+          "finding": "HuggingFace appears as the model library name in reasoning or submitted code; no external benchmark fetch was observed."
+        }
+      },
+      "score": null,
+      "infrastructure_failure": {
+        "stage": "verifier_dependency_installation",
+        "kind": "download_timeout",
+        "dependency": "nvidia-cusparse-cu12==12.5.4.2",
+        "official_assertions_started": false,
+        "native_reward_preserved": 0.0,
+        "native_exception_type": null,
+        "retry_eligible": false,
+        "reason": "Dependency download failed before pytest started. The original wrapper returned reward zero without a Harbor exception. The verifier attempt itself did not time out, so timeout-only retry policy does not permit another attempt."
+      },
+      "agent_output_note": "The model submitted /app/pipeline_parallel.py and exercised a custom torch stub; real PyTorch/Transformers execution and official correctness remain unverified."
+    },
+    {
+      "order": 6,
+      "task": "terminal-bench/regex-chess",
+      "task_ref": "sha256:e763e0ac1c9759081af0a4a82ba51b8cf9ae5485a93de3bbe42d7d344597bd78",
+      "trial": "regex-chess__BbMPdq7",
+      "job": "flash1017-regex-chess-20261018",
+      "raw_reward": 0.0,
+      "classification": "scored_failure",
+      "exception_type": null,
+      "official_test_summary": {
+        "tests": 4,
+        "passed": 0,
+        "failed": 4,
+        "skipped": 0,
+        "pending": 0,
+        "other": 0,
+        "start": 1792332478.3056858,
+        "stop": 1792332478.429377
+      },
+      "started_at": "2026-10-18T13:55:34.303731Z",
+      "finished_at": "2026-10-18T14:08:19.573882Z",
+      "seconds": {
+        "environment_setup": 2.148049,
+        "agent_setup": 37.686158,
+        "agent_execution": 644.763169,
+        "verifier": 59.327986
+      },
+      "work_budget_seconds": 3420,
+      "verifier_timeout_seconds": 3600.0,
+      "verifier_attempts": [
+        {
+          "attempt": 1,
+          "status": "completed",
+          "duration_seconds": 59.124319335998734,
+          "return_code": 0
+        }
+      ],
+      "verifier_dependency_profile": {
+        "profile": null,
+        "status": "not_configured"
+      },
+      "bootstrap_status": "ready",
+      "model_started": true,
+      "trajectory_terminal_outcome": "response_truncated",
+      "recorded_agent_steps": 3,
+      "tool_calls": 2,
+      "incomplete_model_steps": [],
+      "accounting": {
+        "usage_complete": true,
+        "cost_complete": true,
+        "cost_is_estimated": true,
+        "recorded_tokens": {
+          "prompt_tokens": 9482,
+          "completion_tokens": 131209,
+          "cached_tokens": 7168
+        },
+        "estimated_known_cost_usd": 0.158188008
+      },
+      "task_checksum": "64993ab339d437cea9397a033e825663a77d5527a42ef3c2fb2ad25133dcbd4e",
+      "source_ref": "2eb99b608dfb39464d17d109a4a8f11afda607a9",
+      "source_sha256": {
+        "nanopycodeagent/__init__.py": "b922299f65b3db2fb1514df5aeda4413e599f595d2595f45a353b4dc551a1eb1",
+        "nanopycodeagent/agent.py": "aa58960803be77797c2addf65d7cc785a4be54f40ad58d53b9eea9fca1cf4b56",
+        "nanopycodeagent/atif.py": "ca4793b971d125d904888b858452db8db30e21f0d9b87818133062fb6cebe270",
+        "nanopycodeagent/bash_tool.py": "dcf0366fb9ebdca258d4c40b1a50b78b2ec79fc4b85cbb253ac9a9ce9534574f",
+        "nanopycodeagent/cli.py": "e335789b3f09440abd441547a2e3e2feb32b90c68e9ba5fc8237de8f6bac2932",
+        "nanopycodeagent/cost.py": "30a8df17942c424796e8532f45e1cb0784ed6541cea44284b863db3dae2f7b70",
+        "nanopycodeagent/deadline.py": "26a861721e6897a4e02e6ce364cb3ef1f1dc7d4ad9fb2f772404276f5c9e9d5f",
+        "nanopycodeagent/edit_tool.py": "78e2bec911279082b081366ab05cb4cbb37d8910b20b6245cccd4b0424489d0d",
+        "nanopycodeagent/event_journal.py": "8974919c2d460afb52bc57ec37e7e271851290c0845d01b4ad0f9f760cde0814",
+        "nanopycodeagent/read_tool.py": "e0075da19f2e1c6ccf34a3056a0aec92398e5afe67869937a096c6f1e04181c3",
+        "nanopycodeagent/settings.py": "491bf5c6eca50eb4c5fe04b7bfe62a8bfc7d2e1fad05b8eeac1680c2fcf3674e",
+        "nanopycodeagent/terminal.py": "8d27b40976b2b6d476e5106dfea982b055115435b4ba16b7f758412a05b75f04",
+        "nanopycodeagent/tool_validation.py": "3a1a85a7222dd1504bdf58f37e65a51d11de13d16c4e7df42c72b4cd58b5e5e5",
+        "nanopycodeagent/transport.py": "152d3ff78be0e3b2d92e0abbbfb652d73466aff57ad316c06dc0e072411167fe",
+        "nanopycodeagent/write_tool.py": "7671e83464fa1e942f25742fc573d0712e32053b5fc0291257884b027f7f6206",
+        "harbor_adapter/__init__.py": "513ad49c721b3e73293095c6a0486a791483e9aa2b51e0e9ad4b7816af07e76b",
+        "harbor_adapter/__main__.py": "db89d582fba45aad72147317f595531af49c434e3b23c65f22d7491f0b10551d",
+        "harbor_adapter/_verifier_runner.py": "2eb73c3e76d7360f2fa16018f0e933797bf982e589236095b5f98750f3807a7c",
+        "harbor_adapter/adapter.py": "d1aa154ec6972d706f0725cfa69c56635506d4499dff2a129e2c957ea7034c02",
+        "harbor_adapter/bootstrap.py": "f499bfddfc3d29338cae15a7f8f6e9c66d19eb69ad9577e607ced4b1aa057be2",
+        "harbor_adapter/verifier.py": "ec26fdefd69be1c5c450c9ba69d749d4a9d320a2a170b8db34421f0573d9b493"
+      },
+      "wheel_sha256": "f4dcc18758e584696f955ae57fd18c751599b70e8b497ca2db7ef9e2b16e9350",
+      "provenance_checks_passed": true,
+      "evidence_sha256": {
+        "jobs/deepseek-v41-flash-1017-rerun-20261018-v2-orchestration/flash1017-regex-chess-20261018.json": "9257870036a57a2b99937edfcac5fceb3d5ccd020f6266f799c16fb0a3e64458",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/agent/bootstrap.json": "9abae99e4800f37de01cc8c2eb9d13d0437c60d0c1c0724f648d551c7f382b77",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/agent/nanopycodeagent.txt": "d5469e3e71ed9379e67ea53c31f938ac02e393bac9d262ed001e20edb0ce2880",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/agent/trajectory.json": "c2c4272da108a1288c8ac0659c1fd4602bfddec009dd6394b4199085b72d95f5",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/artifacts/manifest.json": "c1db653f487fb773c4f9247162ab3314e7f0584a486914b9874c4b3d6949e42d",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/config.json": "5646757537b53bd73c27d4bfb5976a7897750b8ff6c5ee8e10dc5258c3387084",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/lock.json": "25fb9f9020944b5515ff1abd1911a16dbbc6a297eca9974291c030301d73081b",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/result.json": "a77fe9b3aefafbd0898f517394289d60ece1c7d044d290eaa537e9a91dad28ae",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/trial.log": "ef21738d5435b0a881ea5555e88d382a711f8103dbd9e6e66f1d08c0d635ed14",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/verifier/attempts/1/ctrf.json": "1c6d5890a24a89e28193baa90caa33bcfc43aa63716511807d71df7b6152c12a",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/verifier/attempts/1/reward.txt": "9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/verifier/attempts/1/test-stdout.txt": "52c9df468950afcf550257d9866293ce7156da67c10d162f505cab682d599279",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/verifier/ctrf.json": "1c6d5890a24a89e28193baa90caa33bcfc43aa63716511807d71df7b6152c12a",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/verifier/retry-summary.json": "f43790b8b976ec87f3666a9e94bd4b06cf1a8557d8549eb9fc107a4cf5fa94f0",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/verifier/reward.txt": "9a271f2a916b0b6ee6cecb2426f0b3206ef074578be55d9bc94f6f3fe3ab86aa",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018/regex-chess__BbMPdq7/verifier/test-stdout.txt": "52c9df468950afcf550257d9866293ce7156da67c10d162f505cab682d599279",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018-input/manifest.json": "88852db83196729996f806157ba3c1fd1b985ea5a10d59f275dad2a7c87a284e",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018-input/config.json": "15effe28966d978050d5d431dc53dd7dee702795f81f1243f66ddb1879b6fc31",
+        "jobs/dev34-baseline-main-source/jobs/flash1017-regex-chess-20261018-input/working-tree.patch": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
+      },
+      "previous_baseline": {
+        "task": "terminal-bench/regex-chess",
+        "trial": "regex-chess__g85Ci69",
+        "job": "dev34-regex-chess-main-20261012",
+        "raw_reward": 0.0,
+        "classification": "scored_failure"
+      },
+      "official_pytest_summary": {
+        "executed": true,
+        "collected": 4,
+        "passed": 0,
+        "failed": 4,
+        "skipped": 0,
+        "error": 0,
+        "summary_line": "============================== 4 failed in 0.12s ===============================",
+        "source": "verifier/test-stdout.txt"
+      },
+      "audit": {
+        "benchmark_material_exposure": "not_observed_in_review",
+        "credential_observation_steps": []
+      },
+      "score": 0.0,
+      "failure_analysis": {
+        "cause": "Both substantive model responses reached their output limit; the bounded truncation recovery also truncated. No /app/re.json was submitted.",
+        "official_evidence": "All four official tests executed and failed with FileNotFoundError for /app/re.json.",
+        "atif_steps": [
+          3,
+          5,
+          6,
+          8
+        ],
+        "truncated_responses": [
+          {
+            "step_id": 5,
+            "stop_reason": "max_tokens",
+            "completion_tokens": 65536
+          },
+          {
+            "step_id": 8,
+            "stop_reason": "max_tokens",
+            "completion_tokens": 65535
+          }
+        ],
+        "automatic_truncation_recoveries": 1,
+        "model_phase_seconds": 644.763169,
+        "remaining_work_budget_seconds_approx": 2775.236831,
+        "limitation": "The run stopped under the truncation policy, not the wall-clock deadline. No submitted chess implementation was evaluated; this does not demonstrate the same castling bug as the previous run."
+      }
+    }
+  ],
+  "publication_policy": "Only derived results and hashes are published. Raw logs and trajectories remain local and require credential redaction before sharing.",
+  "provider_checkpoint": {
+    "api_model": "deepseek-flash",
+    "reported_checkpoint": "DeepSeek-V4.1-Flash-1017",
+    "verification": "Official DeepSeek documentation states that deepseek-flash now serves this checkpoint. The API alias is not an immutable checkpoint pin.",
+    "sources": [
+      "https://api-docs.deepseek.com/news/news261017/",
+      "https://api-docs.deepseek.com/quick_start/agent_integrations/"
+    ],
+    "checked_at": "2026-10-18T13:16:26.524186+00:00"
+  },
+  "historical_baseline": {
+    "plan": "benchmarks/harbor/configs/dev34-main-baseline-20261012/experiment-v5.json",
+    "plan_sha256": "f7bc7b329767f4b8007956462f6cdbd4407d386af493524df27ba313c0a4e4c3",
+    "runtime_source_ref": "2eb99b608dfb39464d17d109a4a8f11afda607a9",
+    "original_scores": "28/34 passed; selected six tasks all failed."
+  },
+  "preflight_history": {
+    "install_only_attempts": 8,
+    "successful": 6,
+    "setup_errors": 0,
+    "cancelled_or_administratively_interrupted": 2,
+    "unstarted_original_tasks": 4,
+    "verifier_profile_not_configured_tasks": 6,
+    "original_interruption_record_path": "jobs/flash1017-preflight-interruption.json",
+    "original_interruption_record_sha256": "d16fd7516d61f021806718089cee95c0954c98bb86adc2b6d21b42529fd62c12",
+    "trials": [
+      {
+        "job": "flash1017-install-only-20261018",
+        "task": "terminal-bench/adaptive-rejection-sampler",
+        "trial": "adaptive-rejection-sampler__jrPDadu",
+        "classification": "administratively_interrupted",
+        "model_started": false,
+        "official_assertions_started": false,
+        "result_path": "jobs/dev34-baseline-main-source/jobs/flash1017-install-only-20261018/adaptive-rejection-sampler__jrPDadu/result.json",
+        "result_sha256": "fa2796bb099495c620d472573a9f806deb8acee760e2e0a1db89c63740890f21",
+        "bootstrap_sha256": "52b89cf8662783f2c672e713937a3827ece980f03d4014f290f19f4577e40d15",
+        "manifest_sha256": "347d37f67c5bd204617583de36c2204e562415ef9b4af0b935ae6113fcca86a3",
+        "input_config_sha256": "4db70c5c01691c8a948a99591ddfb15681578f97cd1291991c64568701818cab"
+      },
+      {
+        "job": "flash1017-install-only-20261018",
+        "task": "terminal-bench/db-wal-recovery",
+        "trial": "db-wal-recovery__ufVDrhK",
+        "classification": "administratively_interrupted",
+        "model_started": false,
+        "official_assertions_started": false,
+        "result_path": "jobs/dev34-baseline-main-source/jobs/flash1017-install-only-20261018/db-wal-recovery__ufVDrhK/result.json",
+        "result_sha256": "0107eb2bc2646779f8d9132a9f7c3882c02caa875d801d8d2ba09ca179c8a1ac",
+        "bootstrap_sha256": "71b66a4b9f8828e3f24907dcf037bc12b80b7f0470daad60c2893f5b7d84423b",
+        "manifest_sha256": "347d37f67c5bd204617583de36c2204e562415ef9b4af0b935ae6113fcca86a3",
+        "input_config_sha256": "4db70c5c01691c8a948a99591ddfb15681578f97cd1291991c64568701818cab"
+      },
+      {
+        "job": "flash1017-install-only-v2-20261018",
+        "task": "terminal-bench/adaptive-rejection-sampler",
+        "trial": "adaptive-rejection-sampler__QEJGbeG",
+        "classification": "setup_passed",
+        "model_started": false,
+        "official_assertions_started": false,
+        "result_path": "jobs/dev34-baseline-main-source/jobs/flash1017-install-only-v2-20261018/adaptive-rejection-sampler__QEJGbeG/result.json",
+        "result_sha256": "0791ae260d5a336863391a26e70d678fe017537c61eb9e8082f1c7b8df3fefda",
+        "bootstrap_sha256": "fcebfd336f1b466235db7c3fe8ba5ab12f89adf5748f461affa9f8ab75ebc7a6",
+        "manifest_sha256": "72aa6516e6c4cd285af8b0b7c44cabc2545f0194baa4b6e20286daee4bfce240",
+        "input_config_sha256": "6fc60e637d55c82e22218277b63f76fdb3402ed4c45a8cd710a9a299ab78b549"
+      },
+      {
+        "job": "flash1017-install-only-v2-20261018",
+        "task": "terminal-bench/db-wal-recovery",
+        "trial": "db-wal-recovery__HCpq6qb",
+        "classification": "setup_passed",
+        "model_started": false,
+        "official_assertions_started": false,
+        "result_path": "jobs/dev34-baseline-main-source/jobs/flash1017-install-only-v2-20261018/db-wal-recovery__HCpq6qb/result.json",
+        "result_sha256": "247dd24bab0aeddab02aed1fabddbbda43c9f5e32570fc067915819cc5294a59",
+        "bootstrap_sha256": "51366101df4e7f6425f8fe241b10713287e6f89274173cba25c6ffe708d61dc3",
+        "manifest_sha256": "72aa6516e6c4cd285af8b0b7c44cabc2545f0194baa4b6e20286daee4bfce240",
+        "input_config_sha256": "6fc60e637d55c82e22218277b63f76fdb3402ed4c45a8cd710a9a299ab78b549"
+      },
+      {
+        "job": "flash1017-install-only-v2-20261018",
+        "task": "terminal-bench/kv-store-grpc",
+        "trial": "kv-store-grpc__nqHGLa5",
+        "classification": "setup_passed",
+        "model_started": false,
+        "official_assertions_started": false,
+        "result_path": "jobs/dev34-baseline-main-source/jobs/flash1017-install-only-v2-20261018/kv-store-grpc__nqHGLa5/result.json",
+        "result_sha256": "c540edec6d547acd17f2261ba522a329f7148914d4f42a1a186e14dd4a2cbf05",
+        "bootstrap_sha256": "9abae99e4800f37de01cc8c2eb9d13d0437c60d0c1c0724f648d551c7f382b77",
+        "manifest_sha256": "72aa6516e6c4cd285af8b0b7c44cabc2545f0194baa4b6e20286daee4bfce240",
+        "input_config_sha256": "6fc60e637d55c82e22218277b63f76fdb3402ed4c45a8cd710a9a299ab78b549"
+      },
+      {
+        "job": "flash1017-install-only-v2-20261018",
+        "task": "terminal-bench/query-optimize",
+        "trial": "query-optimize__scB4hts",
+        "classification": "setup_passed",
+        "model_started": false,
+        "official_assertions_started": false,
+        "result_path": "jobs/dev34-baseline-main-source/jobs/flash1017-install-only-v2-20261018/query-optimize__scB4hts/result.json",
+        "result_sha256": "809aacfb8410e2d978f39767c2b7d7a151205293e92f1d6147bd1c24febc739f",
+        "bootstrap_sha256": "4a93e54276aa8f2cc86d505dad8f483806c983d5c140d69bb855ed60c1cd32b6",
+        "manifest_sha256": "72aa6516e6c4cd285af8b0b7c44cabc2545f0194baa4b6e20286daee4bfce240",
+        "input_config_sha256": "6fc60e637d55c82e22218277b63f76fdb3402ed4c45a8cd710a9a299ab78b549"
+      },
+      {
+        "job": "flash1017-install-only-v2-20261018",
+        "task": "terminal-bench/torch-pipeline-parallelism",
+        "trial": "torch-pipeline-parallelism__xtbqWbc",
+        "classification": "setup_passed",
+        "model_started": false,
+        "official_assertions_started": false,
+        "result_path": "jobs/dev34-baseline-main-source/jobs/flash1017-install-only-v2-20261018/torch-pipeline-parallelism__xtbqWbc/result.json",
+        "result_sha256": "778b40ce04610552b7b0b34fd7f2b741b945c982680aabba0f060dbe7caba584",
+        "bootstrap_sha256": "fcebfd336f1b466235db7c3fe8ba5ab12f89adf5748f461affa9f8ab75ebc7a6",
+        "manifest_sha256": "72aa6516e6c4cd285af8b0b7c44cabc2545f0194baa4b6e20286daee4bfce240",
+        "input_config_sha256": "6fc60e637d55c82e22218277b63f76fdb3402ed4c45a8cd710a9a299ab78b549"
+      },
+      {
+        "job": "flash1017-install-only-v2-20261018",
+        "task": "terminal-bench/regex-chess",
+        "trial": "regex-chess__ZBNBNv9",
+        "classification": "setup_passed",
+        "model_started": false,
+        "official_assertions_started": false,
+        "result_path": "jobs/dev34-baseline-main-source/jobs/flash1017-install-only-v2-20261018/regex-chess__ZBNBNv9/result.json",
+        "result_sha256": "34bafd3c020a4c9cfc8ed473cb821ad7f767ad29e14270e22157dc33d893c851",
+        "bootstrap_sha256": "9abae99e4800f37de01cc8c2eb9d13d0437c60d0c1c0724f648d551c7f382b77",
+        "manifest_sha256": "72aa6516e6c4cd285af8b0b7c44cabc2545f0194baa4b6e20286daee4bfce240",
+        "input_config_sha256": "6fc60e637d55c82e22218277b63f76fdb3402ed4c45a8cd710a9a299ab78b549"
+      }
+    ],
+    "limitation": "Successful agent installation does not certify unregistered verifier dependencies. All six profiles were explicitly not_configured. No preflight ran a model or official task assertions."
+  },
+  "final_image_integrity": {
+    "observed_at": "2026-10-18T14:10:12.979320+00:00",
+    "original_task_images_verified": 6,
+    "image_ids_and_registry_digests_unchanged": true,
+    "remaining_scored_trial_containers": []
+  },
+  "post_freeze_audit": {
+    "completed_at": "2026-10-18T14:12:12.926370+00:00",
+    "method": "Review all six frozen trajectories for benchmark references, local verifier paths, external hosts and exact provider credentials; inspect official stdout for every trial and the truncation metadata for regex chess.",
+    "official_assertions_executed_tasks": 4,
+    "setup_errors_in_model_phase": 0,
+    "verifier_errors": 2,
+    "verifier_dependency_download_failure_tasks": [
+      "terminal-bench/adaptive-rejection-sampler",
+      "terminal-bench/torch-pipeline-parallelism"
+    ],
+    "benchmark_material_exposure_tasks": [],
+    "hidden_solution_or_test_retrieval_observed": false,
+    "provider_credentials_observed_in_trajectories": false,
+    "public_network_access": "The sampler probed the public R project host while installing R. gRPC /proc reads inspected process command lines, not environment files. No external benchmark-answer retrieval was observed.",
+    "limitation": "Retrospective review cannot prove absence of all exposure or training contamination. These are previously exposed development tasks."
+  },
+  "comparison_note": "Three previous failures passed this time; regex chess was a valid scored failure; sampler and pipeline are unscored because official dependencies failed before pytest. Raw zeros for those infrastructure errors are preserved but are not counted as model failures. These single selected-task trials do not isolate a checkpoint effect or update the entire 34-task pool score.",
+  "accounting_note": "Costs are the unchanged agent client token-based estimates, not independently verified provider billing."
+}
diff --git a/benchmarks/harbor/src/harbor_adapter/_verifier_runner.py b/benchmarks/harbor/src/harbor_adapter/_verifier_runner.py
index da4f84c..3b2f830 100644
--- a/benchmarks/harbor/src/harbor_adapter/_verifier_runner.py
+++ b/benchmarks/harbor/src/harbor_adapter/_verifier_runner.py
@@ -1,4 +1,4 @@
-"""Run a Linux verifier inside its container with bounded timeout retries.
+"""Run a Linux verifier with bounded, explicitly selected recovery rules.
 
 This file is sent to the container as Python source and uses only the standard
 library. It never starts an agent or changes the task's test script.
@@ -7,6 +7,7 @@
 import json
 import os
 from pathlib import Path
+import re
 import shutil
 import signal
 import subprocess
@@ -51,14 +52,41 @@ def _archive(logs, attempt):
             shutil.copy2(source, target / source.name)
 
 
-def run(command, timeout_sec, max_attempts, logs, control):
-    """Return execution status; only our own elapsed deadline is retryable."""
+def dependency_download_failed(logs):
+    """Recognize uv network failures before tests, never a completed test score."""
+    if (logs / "ctrf.json").exists():
+        return False
+    for name in ("reward.txt", "reward.json"):
+        path = logs / name
+        if path.exists():
+            try:
+                value = (float(path.read_text()) if name.endswith(".txt")
+                         else json.loads(path.read_text()).get("reward"))
+            except (ValueError, TypeError, AttributeError):
+                return False
+            if value != 0:
+                return False
+    path = logs / "test-stdout.txt"
+    text = path.read_text(errors="replace") if path.exists() else ""
+    if re.search(
+        r"test session starts|collected \d+|FAILED |ERROR collecting|"
+        r"\b(?:passed|failed|errors?|skipped)\b.*\bin \d", text,
+    ):
+        return False
+    return bool(
+        re.search(r"Failed to (?:download|fetch) `[^`]+`", text)
+        and re.search(r"network timeout|operation timed out|Request failed after", text)
+    )
+
+
+def run(command, timeout_sec, max_attempts, logs, control, retry_dependency_errors=False):
+    """Retry deadlines and, when enabled, pre-test dependency download failures."""
     logs = Path(logs)
     control = Path(control)
     cancelled = control.with_suffix(".cancelled")
     attempts = []
     summary = {"timeout_sec": timeout_sec, "max_attempts": max_attempts,
-               "attempts": attempts}
+               "retry_dependency_errors": retry_dependency_errors, "attempts": attempts}
 
     def interrupted(signum, frame):
         raise InterruptedError("Verifier execution cancelled")
@@ -92,6 +120,8 @@ def interrupted(signum, frame):
                 try:
                     record["return_code"] = process.wait(timeout=timeout_sec)
                     record["status"] = "completed"
+                    if retry_dependency_errors and dependency_download_failed(logs):
+                        record["status"] = "dependency_error"
                 except subprocess.TimeoutExpired:
                     _stop_process(process)
                     record["status"] = "timeout"
@@ -139,5 +169,13 @@ def cancel(control):
     if sys.argv[1] == "cancel":
         cancel(sys.argv[2])
     else:
-        command, timeout, count, logs, control = sys.argv[1:]
-        print(json.dumps(run(command, float(timeout), int(count), logs, control)))
+        arguments = sys.argv[1:]
+        if len(arguments) == 5:
+            command, timeout, count, logs, control = arguments
+            retry_dependencies = False
+        else:
+            command, timeout, count, retry, logs, control = arguments
+            retry_dependencies = retry == "1"
+        print(json.dumps(run(
+            command, float(timeout), int(count), logs, control, retry_dependencies,
+        )))
diff --git a/benchmarks/harbor/src/harbor_adapter/verifier.py b/benchmarks/harbor/src/harbor_adapter/verifier.py
index 7fd6e3f..cd5fc3b 100644
--- a/benchmarks/harbor/src/harbor_adapter/verifier.py
+++ b/benchmarks/harbor/src/harbor_adapter/verifier.py
@@ -16,15 +16,21 @@
 _RUNNER = Path(__file__).with_name("_verifier_runner.py").read_text()
 
 
+class VerifierDependencyError(RuntimeError):
+    """Authorized dependency-download recovery exhausted its attempt budget."""
+
+
 class _RetryEnvironment:
     """Delegate Harbor operations, wrapping only the test-script execution."""
 
-    def __init__(self, environment, command, timeout_sec, max_attempts, logger):
+    def __init__(self, environment, command, timeout_sec, max_attempts, logger,
+                 retry_dependency_errors=False):
         self._environment = environment
         self._command = command
         self._timeout_sec = timeout_sec
         self._max_attempts = max_attempts
         self._logger = logger
+        self._retry_dependency_errors = retry_dependency_errors
 
     def __getattr__(self, name):
         return getattr(self._environment, name)
@@ -35,7 +41,8 @@ async def exec(self, *, command, **kwargs):
         control = f"/tmp/nanopy-verifier-{uuid.uuid4().hex}.pid"
         invocation = shlex.join([
             "python3", "-c", _RUNNER, command, str(self._timeout_sec),
-            str(self._max_attempts), "/logs/verifier", control,
+            str(self._max_attempts), str(int(self._retry_dependency_errors)),
+            "/logs/verifier", control,
         ])
         try:
             result = await self._environment.exec(command=invocation, **kwargs)
@@ -55,6 +62,10 @@ async def exec(self, *, command, **kwargs):
         if result.return_code != 0:
             raise RuntimeError(f"Verifier retry runner failed: {result.stderr}")
         record = json.loads(result.stdout)
+        if record["status"] == "dependency_error":
+            raise VerifierDependencyError(
+                f"Verifier dependency downloads failed after {self._max_attempts} attempts"
+            )
         if record["status"] == "timeout":
             raise TimeoutError(
                 f"Verifier timed out after {self._max_attempts} attempts "
@@ -71,13 +82,16 @@ class RetryingVerifier(Verifier):
     4``. Each attempt retains the task's native verifier timeout.
     """
 
-    def __init__(self, *, max_attempts=3, **kwargs):
+    def __init__(self, *, max_attempts=3, retry_dependency_errors=False, **kwargs):
         if type(max_attempts) is not int or not 1 <= max_attempts <= 3:
             raise ValueError("max_attempts must be an integer between 1 and 3")
+        if type(retry_dependency_errors) is not bool:
+            raise ValueError("retry_dependency_errors must be a boolean")
         super().__init__(**kwargs)
         if self.environment.os != TaskOS.LINUX or self.step_name is not None:
             raise ValueError("RetryingVerifier supports single-step Linux tasks")
         self._max_attempts = max_attempts
+        self._retry_dependency_errors = retry_dependency_errors
         self._attempt_timeout = self.task.config.verifier.timeout_sec
         if not math.isfinite(self._attempt_timeout) or self._attempt_timeout <= 0:
             raise ValueError("The task's verifier timeout must be finite and positive")
@@ -110,6 +124,7 @@ async def verify(self):
         environment = self.environment
         self.environment = _RetryEnvironment(
             environment, command, self._attempt_timeout, self._max_attempts, self.logger,
+            self._retry_dependency_errors,
         )
         try:
             return await super().verify()
diff --git a/benchmarks/harbor/tests/test_verifier.py b/benchmarks/harbor/tests/test_verifier.py
index 075c677..db7620f 100644
--- a/benchmarks/harbor/tests/test_verifier.py
+++ b/benchmarks/harbor/tests/test_verifier.py
@@ -81,7 +81,8 @@ async def exec(self, *, command, env=None, **kwargs):
         return ExecResult(return_code=process.returncode, stdout=stdout.decode(), stderr=stderr.decode())
 
 
-def make_verifier(tmp_path, script, *, mounted=True, max_attempts=3, include_logs=None, exclude_logs=None):
+def make_verifier(tmp_path, script, *, mounted=True, max_attempts=3, include_logs=None,
+                  exclude_logs=None, retry_dependency_errors=False):
     task_dir = tmp_path / "task"
     (task_dir / "tests").mkdir(parents=True)
     (task_dir / "instruction.md").write_text("Write the solution.")
@@ -94,7 +95,8 @@ def make_verifier(tmp_path, script, *, mounted=True, max_attempts=3, include_log
         task={"path": task_dir}, verifier_timeout_multiplier=4,
         verifier=VerifierConfig(
             import_path="harbor_adapter:RetryingVerifier",
-            kwargs={"max_attempts": max_attempts},
+            kwargs={"max_attempts": max_attempts,
+                    "retry_dependency_errors": retry_dependency_errors},
             include_logs=include_logs or [], exclude_logs=exclude_logs or [],
         ),
     )
@@ -163,6 +165,75 @@ def test_any_valid_score_stops_after_first_attempt(tmp_path, reward):
     assert len(summary(verifier)["attempts"]) == 1
 
 
+@pytest.mark.parametrize("mounted", [True, False])
+def test_authorized_download_retries_keep_output_and_stop_on_scored_zero(tmp_path, mounted):
+    verifier, _, _ = make_verifier(tmp_path, '''
+count=$(cat "$STATE" 2>/dev/null || echo 0)
+count=$((count + 1))
+echo "$count" > "$STATE"
+if [ "$count" -lt 3 ]; then
+    echo 'Failed to download `numpy==2.3.3`'
+    echo 'Failed to download distribution due to network timeout.'
+else
+    echo '{"results":{"summary":{"tests":1,"failed":1}}}' > "$LOGS/ctrf.json"
+    echo '1 failed in 0.01s'
+fi
+echo 0 > "$LOGS/reward.txt"
+''', mounted=mounted, retry_dependency_errors=True)
+    assert asyncio.run(verifier.verify()).rewards == {"reward": 0}
+    assert [r["status"] for r in summary(verifier)["attempts"]] == [
+        "dependency_error", "dependency_error", "completed",
+    ]
+    assert (verifier.trial_paths.verifier_dir / "attempts/1/reward.txt").read_text() == "0\n"
+
+
+def test_exhausted_download_errors_are_unscored(tmp_path):
+    from harbor_adapter.verifier import VerifierDependencyError
+
+    verifier, _, _ = make_verifier(tmp_path, '''
+echo 'Failed to download `nvidia-cusparse-cu12==12.5.4.2`'
+echo 'Request failed after 3 retries: operation timed out'
+echo 0 > "$LOGS/reward.txt"
+''', retry_dependency_errors=True)
+    with pytest.raises(VerifierDependencyError, match="3 attempts"):
+        asyncio.run(verifier.verify())
+    assert len(summary(verifier)["attempts"]) == 3
+    assert not verifier.trial_paths.reward_text_path.exists()
+    assert (verifier.trial_paths.verifier_dir / "attempts/3/reward.txt").exists()
+
+
+@pytest.mark.parametrize("guard", [
+    'echo "test session starts"',
+    'echo "1 failed in 0.1s"',
+    'echo "{}" > "$LOGS/ctrf.json"',
+    'echo 1 > "$LOGS/reward.txt"',
+])
+def test_download_text_never_retries_started_tests_or_success(tmp_path, guard):
+    verifier, _, _ = make_verifier(tmp_path, '''
+echo 'Failed to download `numpy==2.3.3`'
+echo 'network timeout'
+echo 0 > "$LOGS/reward.txt"
+''' + guard + "\n", retry_dependency_errors=True)
+    asyncio.run(verifier.verify())
+    assert len(summary(verifier)["attempts"]) == 1
+
+
+def test_download_recovery_is_disabled_by_default(tmp_path):
+    verifier, _, _ = make_verifier(tmp_path, '''
+echo 'Failed to download `numpy==2.3.3`'
+echo 'network timeout'
+echo 0 > "$LOGS/reward.txt"
+''')
+    assert asyncio.run(verifier.verify()).rewards == {"reward": 0}
+    assert len(summary(verifier)["attempts"]) == 1
+
+
+@pytest.mark.parametrize("value", [1, "true", None])
+def test_invalid_dependency_retry_flags_are_rejected(value):
+    with pytest.raises(ValueError, match="retry_dependency_errors"):
+        RetryingVerifier(retry_dependency_errors=value)
+
+
 @pytest.mark.parametrize("mounted", [True, False])
 def test_three_timeouts_leave_no_canonical_reward_and_keep_all_logs(tmp_path, mounted):
     verifier, environment, _ = make_verifier(tmp_path, '''