[wangjohn]

What and why

WP-10b of plans/retry-wp1-wp10.md: a second attempt at bundled generic contrast. 60 original drafts (6 registers × 10, 15 per length band) in src/styleprofile/data/generic-contrast/; build --generic-contrast / generic_contrast=True; the contrast anchor (median, min, AUC) is computed leaving out a whole register at a time. Report 9.1 (optional settings.generic_contrast).

Commit order: design section, data sets and bars in docs/method.md (b0c06dd) → code + corpus → corpus/code frozen (7bceb21) → test set read once → results → follow-ups dc45b26 and 949b2fd (below).

Draft: the one-shot test misses one bar (length alone). Nothing about the corpus, anchor or bars was changed after seeing the test set. Because of the miss, generic-contrast likeness is shown as an indicative number with no verdict words (the plan's suggested alternative):

• score: generic LLM-likeness (indicative): 13.08; no verdict column or "most like" chunk list.
• score -q: … generic LLM likeness (indicative) 13.08.
• --fail-likeness refuses a generic profile.
• Library: ScoreResult.indicative / DocumentResult.indicative are True for generic contrast (docs/library.md).
• Reports still store the verdicts, so the evaluation is reproducible.
• The contrast label is generic LLM (changed from the plan's "generic LLM drafts", which read "…drafts drafts").

Evidence

Bar	Result
Dev: 5/5 examples/llm-drafts lean LLM or beyond	5/5	met
Test: ≥13/16 lean LLM or beyond	14/16	met
Test: AUC vs writer held-out ≥ 0.90	1.000	met
Test: length alone ≤ 0.70	0.850 (0.545 vs the 7 essays only)	not met
Writer: ≥95% "like the reference"	45/47 (95.7%)	met

Full per-draft table in docs/method.md. The plan's corpus length check against medium (AUC ≤ 0.70) is unreachable by construction (every medium doc has 711–1,363 words; bound ≥0.77, measured 0.81), which also explains the test's length miss.

Profiles built without generic contrast are byte-identical to main's apart from minor_version. make check (922 passed), make demo, make dist-check, make snapshots (no diff) and the bench gate pass.

Tests

tests/test_generic_contrast.py (corpus shape/bands/size/licence, labelling, --contrast combination, CLI flag, group folds, indicative display, quiet line, library flag, --fail-likeness refusal); schema test; scripts/check_dist.py checks the 60 drafts in the wheel and an installed smoke build.

Decisions

• Register folders are separate folder inputs, so pooling never mixes registers; group = generic:<register>, else the document.
• Indicative-only display chosen over a verdict feature; label changed to generic LLM on review.

Follow-ups

• A corrected, length-matched test with fresh held-out drafts would be needed before verdict words are justified.
• The chunk "flagged" count still includes chunks flagged by generic likeness (the words are hidden).

🤖 Generated with Claude Code

https://claude.ai/code/session_01QyXhSxvDMHzrqSoTVM6Vfn