release: 2.23.0 - findings.json for real audits, suite 6 wired into the live eval - #136
Merged
Merged
Conversation
…to the live eval otchet.md now asks every completed audit to also write findings.json - same seven-plus-confidence fields as the prose, one object per finding, so a result can be read by a dashboard or compared month to month without re-parsing Bulgarian text. This is a material guidance change (minor bump). eval_skill.py gains --lifecycle: suite 6's five-month timeline fixture (I11) sent to a live model for the first time, graded by person-and-month rather than by row via a new (location_lifecycle, grade_lifecycle) pair, with lifecycle_test.py's own reconcile() as ground truth. LIFECYCLE_KEYWORDS is uncalibrated against a paid transcript, same caveat KOMPLEKT_KEYWORDS already carries - triage misses before believing them. lifecycle_test.py adds a free self-check that a hand-written correct sentence for each of the 6 I11 break ids grades as identified, before any paid seed is spent proving it. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…pen twice
location_lifecycle()'s bare-digit fallback matched any 1-2 digit number found
anywhere in a finding's kade text - a row reference ("ред 8"), a statute
article - and silently mislocated the finding instead of failing closed, the
way location()'s own row parsing already refuses to do. Now the bare-digit
fallback only applies when the number sits immediately adjacent to the
identifier (as reconcile()'s own ground-truth "ИДЕНТ ММ" string, and
lifecycle_test.py's self-check, actually have it); anywhere else in the
sentence it is dropped rather than trusted.
print_graded()'s amount-mismatches sub-loop also still used the pre-lifecycle
`where == "file"` check three lines below the isinstance(where, int) fix
applied to the main loop in the same function, so a lifecycle amount
mismatch printed "[row СЛ-003 04]" instead of "[СЛ-003 04]".
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
This was referenced Sep 9, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
otchet.md: every completed audit also writesfindings.json— same seven-plus-confidence fields as the prose report, one object per finding,nullfor findings with no numeric/named location. Material guidance change → minor bump.eval_skill.py: new--lifecyclemode sends suite 6's five-month timeline fixture (I11 — raise without annex, pay after termination, severance without termination, sick-days restart, class raised early/not raised) to a live model for the first time, graded by person-and-month (location_lifecycle/grade_lifecycle) rather than by row, since the same row number is a different person in each monthly sheet. Ground truth islifecycle_test.py's ownreconcile().lifecycle_test.py: free self-check proves a hand-written correct sentence for each of the 6I11_*ids grades asidentifiedbefore any paid seed runs.CLAUDE.md: documents both additions.Caveat carried into this PR
LIFECYCLE_KEYWORDShas never been calibrated against a paid transcript (same limitationKOMPLEKT_KEYWORDSalready states). A first real--lifecyclerun (paid, ~$2-5/seed on Sonnet) should happen before trusting it for grading; that run is not part of this PR perCLAUDE.md's rule to never runeval_skill.pyunprompted.Test plan
test/rates_test.pytest/skill_test.py(manifests/version consistency)test/checks_test.pytest/preflight_test.pytest/k_checker_test.pytest/audit_test.pytest/komplekt_test.pytest/lifecycle_test.py(including the new eval-grading self-check)test/eval_skill.py --selftest(free, no paid session)test/eval_skill.py --lifecycle --dry --seeds-list 1(prompt/CLI plumbing sanity)🤖 Generated with Claude Code