Skip to content

Phase 1: golden fixture harness — the Definition of Working - #6

Merged
mrlucas679 merged 1 commit into
mainfrom
claude/festive-cannon
Jul 5, 2026
Merged

Phase 1: golden fixture harness — the Definition of Working#6
mrlucas679 merged 1 commit into
mainfrom
claude/festive-cannon

Conversation

@mrlucas679

Copy link
Copy Markdown
Owner

Executes MASTER_PLAN.md Phase 1:

  • 24 golden translation fixtures (medical/emergency/modal/aspect/question/rights/regression/hostile/exploratory) + 6 intent-level reconstruction fixtures — every expectation verified against real pipeline output, known gaps labeled instead of faked
  • Word-order-aware scorer: required/forbidden signs, marker preservation, [before, after] order pairs (the SASL axis bag-of-signs metrics miss), avatar-library compliance with signs_library.js as ground truth
  • CI gate: all critical cases must pass 100%; rules-only baseline printed in logs

Four real grammar bugs caught by the harness on its first run, all fixed with regressions:

  1. 'need' → 'NE' (naive -ed stripping)
  2. False FINISH aspect from -s/-ing forms
  3. 'walking' → 'WALKE' (silent-e heuristic)
  4. RIGHTS → RIGHT (library signs re-stemmed into non-signs)

Vocabulary hole found: NEED — the top clinic verb — was absent from the avatar library; added as review-flagged placeholder (conf 2). The measured 16-sign transformer/avatar gap is recorded in the plan for Phase 3 real-data recording.

Baseline: 24/24 golden, 16/16 critical, mean recall 1.0. Suite: 137 passed.

🤖 Generated with Claude Code

The Definition of Working (MASTER_PLAN Phase 1):
- tests/golden/translation_cases.json — 24 fixtures (medical, emergency,
  modal, aspect, question, rights, regression, hostile, exploratory),
  every expectation verified against real pipeline output
- tests/golden/sign_reconstruction_cases.json — 6 intent-level fixtures
- tests/golden/scoring.py — scorer with required/forbidden signs, marker
  preservation, WORD-ORDER pairs (the SASL axis a bag-of-signs metric
  misses), and avatar-library compliance (signs_library.js is ground
  truth, not the transformer JSON)
- pytest gates: every critical case must pass; baseline printed in CI

Bugs the harness caught on its first run, all fixed with regressions:
1. 'need' → 'NE' (naive -ed stripping; min stem length + eed guard)
2. false FINISH aspect from -s/-ing forms ('My chest hurts' was past)
3. 'walking' → 'WALKE' (silent-e heuristic; library-aware + irregular map)
4. RIGHTS → RIGHT (library signs must never be re-stemmed)

Vocabulary: NEED added to signs_library.js as a review-flagged placeholder
(conf 2, MUST-family motion) — the top clinic verb was unplayable. The
remaining 16-sign transformer/avatar gap is recorded in MASTER_PLAN for
Phase 3 real-data recording.

Baseline: 24/24 golden cases, 16/16 critical, mean recall 1.0.
Suite: 137 passed.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
@mrlucas679
mrlucas679 merged commit 39ead50 into main Jul 5, 2026
14 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.

1 participant