Skip to content

test(eval): pin check error and baseline delta; fix harness docs #346

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
napetrov merged 1 commit into from
Nov 1, 2026
Merged

test(eval): pin check error and baseline delta; fix harness docs #346

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 4 additions & 2 deletions dokimi/docs/preparing-a-repository.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -199,8 +199,10 @@ Recurring review comments are review knowledge a repository already has.
|---|---|
| Claude Code 2.1.286 (measured 2026-10) | `CLAUDE.md`. `AGENTS.md` only when no `CLAUDE.md` is in its start-up chain. Subdirectory files lazily, when it reads a file there (a subdirectory `AGENTS.md` only if no `CLAUDE.md` is in the chain). |

Other harnesses (Codex, Copilot and others) are not listed because no loading
behavior was measured for them. Each needs its own `harness_profile.v1` and a
Codex (codex-cli 0.160) and Copilot CLI (1.0.91) were measured too: what each
loaded is in the `guidance.notes` of the built-in `codex` and `copilot` profiles
(`dokimi/repos/profiles/codex.yaml`, `dokimi/repos/profiles/copilot.yaml`). A
harness without a built-in profile needs its own `harness_profile.v1` and a
measurement with `dokimi repo reach`; the `text`-trace limit below applies.

Put a `CLAUDE.md` containing `@AGENTS.md` next to each `AGENTS.md`, including in
Expand Down
3 changes: 3 additions & 0 deletions dokimi/tests/test_arm_cell_errors.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -117,6 +117,9 @@ def test_delta_vs_baseline_is_paired_over_the_questions_both_arms_scored():
summary = runner._summarize(evaluations, "baseline")
assert summary["per_arm"]["skill"]["delta_vs_baseline"] == -10.0
assert summary["per_arm"]["skill"]["avg_aggregate"] == 55.0 # the unpaired mean is still reported
# The baseline has no delta against itself: the key is there, as null, not 0.0.
assert "delta_vs_baseline" in summary["per_arm"]["baseline"]
assert summary["per_arm"]["baseline"]["delta_vs_baseline"] is None


@pytest.mark.parametrize("reply", ["", " \n\t", None])
Expand Down
9 changes: 9 additions & 0 deletions dokimi/tests/test_reference_checks.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -474,6 +474,15 @@ def test_eval_reference_counts_a_judge_failure_as_an_error_for_every_arm():
assert (cell["passed"], cell["aggregate"], cell["error"]) == (None, None, "judge: TimeoutError")


def test_the_failed_judge_check_names_its_error_in_the_cell_and_the_graded_check_has_none():
data = {"baseline_arm": "baseline", "arms": ["baseline"], "answers": [
{"question_id": "q1", "question_text": "Capital?", "arms": {"baseline": {"answer": "Paris"}}}]}
result = score(data, [{"id": "q1", "question": "Capital?", "grading": _MIXED}], ask=_judge_raises)
match, judge = result.evaluations[0]["scores"]["baseline"]["checks"]
assert (judge["type"], judge["error"]) == ("judge", "judge: TimeoutError")
assert match == {"type": "match", "passed": True, "score": 1.0}


def test_a_reference_cell_carries_its_answer_error_as_an_arm_cell_does(tmp_path):
data = {"baseline_arm": "baseline", "arms": ["baseline", "skill"], "answers": [
{"question_id": "q1", "question_text": "Capital?",
Expand Down
Loading