Skip to content

ledger(wildct): amendment R7, H25w keeps its registered CPU Fast Draft - #333

Merged
88plug merged 2 commits into
mainfrom
claude/wildct-r7-h25w-draft
Nov 7, 2026
Merged

88plug merged 2 commits into
mainfrom
claude/wildct-r7-h25w-draft

Conversation

@88plug

@88plug 88plug commented Nov 7, 2026

Copy link
Copy Markdown
Owner

Wildct amendment R7, mechanics only, before any H25w hold. It changes no hypothesis, prediction, arm, model, statistic, threshold, cap, disk need, hold count or order. Nothing on wildct changed.

What was found

The pre-registration's H25w line says the agent manifest "keeps its default CPU Fast Draft in both arms, as H25 run 4 did". It stages the draft (OpenVINO/Qwen3-0.6B-int4-ov @ f864c610) for that, and its command sets no draft variable.

When the line was written, the launcher drafted by default for any CPU target serving an OpenVINO Qwen3 model (#116). The model optimizer (#163) merged three minutes before the pre-registration (#181). It narrowed the default: the launcher drafts on the CPU by default only where an H34-family row is SUPPORTED for the host's CPU class. No such row names the i7-13700 (H34w is PENDING). So on wildct, release 0.1.89's agent manifest writes no draft, and the registered command would serve both arms without one.

Read from main's launcher at 28b1e70 (equal to 0.1.89's apart from VERSION), with K3D_GPU_SYSROOT naming the i7-13700 and run.sh's backend and device:

  • With no draft variable, the manifest has no draft lines. The launcher prints "agent: no CPU Fast Draft: no EXPERIMENTS.md H34-family row measured 13th Gen Intel(R) Core(TM) i7-13700 (K3D_GPU_AGENT_DRAFT=CPU forces it)".
  • With K3D_GPU_AGENT_DRAFT=CPU, it writes --draft_source_model=OpenVINO/Qwen3-0.6B-int4-ov and --draft_device=CPU. The manifest is then byte-identical to the default one rendered for the i9-13900T, where H25 run 4 ran.

H25x served both arms with the same draft (bench/xhost/results/2026-10-20/H25x/hold/out/on.yaml).

Change

  • bench/wildct/row.sh: h25w adds K3D_GPU_AGENT_DRAFT=CPU. run.sh renders one manifest and derives OFF from it by the one --enable_prefix_caching=false line, so both arms get the draft alike.
  • tests/wildct_hold_test.sh: row.sh's H25w environment carries the variable.
  • EXPERIMENTS.md: the amendment. The H25w row is not edited here; its run PR links the amendment.

H34w is not affected: h34.sh writes its own graphs.

Checks

  • tests/wildct_hold_test.sh: 41 passed. Against main's row.sh, its H25w line fails (no K3D_GPU_AGENT_DRAFT=CPU in the printout).
  • python3 bench/wildct/test_wildct.py, tests/ledger_freshness_test.py, tests/ledger_drift_check_test.py, tests/known_issues_test.py: OK.
  • python3 tests/ledger_drift_check.py: 43 files agree with 82 ledger rows. tests/claims_check.sh: 28 files clean.
  • shellcheck v0.11.0: bench/wildct/row.sh clean.

Touches no shipped path, so it cuts no release.

The H25w line registers both arms with the agent manifest's default CPU
Fast Draft, as H25 run 4 had it, and stages the draft for that. Since the
model optimizer (#163) the launcher drafts on the CPU by default only where
an H34-family row is SUPPORTED for the host's CPU class, and none names the
i7-13700, so the registered command would serve both arms without the
draft. row.sh h25w now sets K3D_GPU_AGENT_DRAFT=CPU, which renders the same
manifest the i9-13900T default does. Both arms alike; no hypothesis,
prediction, arm, model, statistic, threshold, cap or order changes.
@88plug
88plug merged commit ce4ff16 into main Nov 7, 2026
82 checks passed
@88plug
88plug deleted the claude/wildct-r7-h25w-draft branch November 7, 2026 05:17
88plug added a commit that referenced this pull request Nov 7, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant