Skip to content

Add c1/c4 AND-clause audit doc (inventory + diagnosis plan)#170

Merged
LuminLynx merged 2 commits into
mainfrom
claude/rubric-audit-c1-c4
May 21, 2026
Merged

Add c1/c4 AND-clause audit doc (inventory + diagnosis plan)#170
LuminLynx merged 2 commits into
mainfrom
claude/rubric-audit-c1-c4

Conversation

@LuminLynx
Copy link
Copy Markdown
Owner

@LuminLynx LuminLynx commented May 21, 2026

Summary

The deferred c1 / c3 follow-up to the criterion-2 AND-clause sweep (docs/RUBRIC_AUDIT.md:13, :91). Adds docs/RUBRIC_AUDIT_C1_C4.md — a doc-only PR, no rubric or regression changes.

Inventories the framing criterion (c1) and the regime/mapping criterion (now c4 in split units, c3 in the two 3-criterion units) across all 13 units, scores AND-clause severity on the same rubric as the c2 sweep, and defines the operator-local grader-diagnosis step that gates any actual split.

Key findings

  • Regime criterion — clean two-tier split: units 1–4 are clean single distinguishes; units 5–13 all append a meta-recognition (hybrid default / layering / "common PM error") to the mapping. Same lenient-bundling shape the c2 sweep targeted; the "common PM error" clause (units 10–13) is the most droppable concept in the curriculum.
  • c1: substantive AND in 12 of 13 units (tokenization clean; vector-search-rag mildest), with a HIGH framing-reframe group (4, 5, 7, 9, 11, 12).

No grades or expected values change. The rubrics table is version-scoped, so this doc commits no schema or data changes; it only records the inventory and the plan.

Note: the second commit corrects a stale-data error in the first draft, which had claimed units 8/9 still carried a bundled c2. They were already split and merged (#167/#168) before this audit began; the c2 sweep is complete (all of positions 2–12 are 4-criterion). The c1/regime analysis for 8/9 is unaffected.

Test plan

  • Doc-only change — git diff touches a single new file under docs/
  • Operator step (local, not in CI): run differential pairs through live Sonnet 4.6 per the doc's step 4 before any split proceeds

https://claude.ai/code/session_019xEvNkByf5ic4kbMZFdKDR

claude added 2 commits May 21, 2026 21:11
The deferred c1/c3 follow-up to the criterion-2 sweep. Inventories the
framing criterion (c1) and the regime/mapping criterion (now c4 in split
units) across all 13 units, scores AND-clause severity, and lays out the
operator-local grader-diagnosis step that gates any split.

Two-tier finding on the regime criterion: units 1-4 are clean single
distinguishes; units 5-13 all append a meta-recognition (hybrid / layering
/ "common PM error") to the mapping — the same lenient-bundling shape the
c2 sweep targeted. c1 carries a substantive AND in 12 of 13 units, with a
HIGH framing-reframe group (4,5,7,9,11,12).

Also surfaces a c2-rollout gap found during inventory: units 8 and 9 were
never split and still carry the bundled c2 (flagged, not actioned).

https://claude.ai/code/session_019xEvNkByf5ic4kbMZFdKDR
The first draft claimed units 8 and 9 still carried a bundled c2. That was
read off a stale working tree from before a git pull — PRs #167 and #168
had already split both (rubric + regression sets to 4 criteria) and merged.
Remove the false "incomplete sweep" section, fix the numbering note (only
tokenization and the reverted multimodal remain 3-criterion), and correct
the regime-criterion index for units 8/9 from c3 to c4.

The c1 and regime-criterion analysis for 8/9 is unchanged — the split left
their c1 text alone and only renumbered the regime criterion.

https://claude.ai/code/session_019xEvNkByf5ic4kbMZFdKDR
@LuminLynx LuminLynx merged commit 546aec4 into main May 21, 2026
2 checks passed
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.

2 participants