apply answerability-audit repair queue: ACCEPT variants, derivability fixes, rubric rewrites (46 masters)

- MISSING_ALTERNATIVES (63/64 applied): uncontracted/contracted forms, digit
  variants (05|5), synonyms and full-sentence variants, via ACCEPT lines and
  inline gap alternatives; skipped 'better/fine' for the Opposites drill (not
  opposites of "sick").
- NOT_DERIVABLE (16 fixed, 10 false alarms, 1 skipped): item-level fixes
  (visible word hints, missing premise, speaker clarifications "Joe (to Jim):",
  emphasis/cue corrections, un-gapped unguessable words, rounded-number
  accepts); converted the two book-10 paragraph clozes (Mary / The Mall) from
  CHECK: exact to CHECK: llm with logical-completion rubrics; cross-book
  recycled vocab (pilot, teller, barber, variety store, ...) left alone.
- UNCLEAR_TARGET (10): instruction/intro corrections (Food "Type the word for
  the food.", Explain-with-Too intro), cue/emphasis fixes.
- Rubric rewrites (57): applied audit-suggested grader rubrics (now effective
  with the pre-item RUBRIC parser fix); moved the two per-item rubrics of
  book-11 "Write the Main Idea" inside their items.

No clip-bearing line changed; all masters validate (0 errors, 0 new warnings)
and pass stage-master clip-identity checks. Full decision log:
reactor-module-tools module-convert/answerability-audit/APPLIED.md

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
2026-08-05 12:08:14 +08:00
parent 2d90d05bc5
commit 61234fe36b
46 changed files with 154 additions and 128 deletions

View File

@@ -871,7 +871,7 @@ RESPONSE: The teller needed to see some identification.
TEMPLATE: Jan ____{showed} the teller her ____{driver's license}.
RESPONSE: Jan showed the teller her driver's license.
TEMPLATE: Then she ____{wrote a check}.
TEMPLATE: Then she wrote a ____{check}.
RESPONSE: Then she wrote a check.
TEMPLATE: Jan and the teller ____{finished}.
@@ -1242,7 +1242,7 @@ INTRO: Now run the machine yourself. Each item gives you a starting point — bu
INSTRUCTION: Say a sentence or question built from the table.
INPUT: speak
CHECK: llm
RUBRIC: Accept any grammatically correct sentence or question that can be formed by combining the parts from the table and that begins as the item asks; the model answer is only one sample.
RUBRIC: Accept any grammatically correct sentence or question that begins as the item asks, especially ones using an indirect object; the model answer is only one sample. Do not restrict to the model's vocabulary — the grader cannot see the table.
EXAMPLE
TEMPLATE: Make a question starting with "Can I …".