fix(ml): sample the whole corpus and drop PROPN from distinctive_vocab (WO-18)

[sergi-torres]

Base is fix/wo-08-comparison-corpora, not main. Stacked PR. When #113 merges, GitHub retargets this one to main and a conflict will appear in docs/decision_log.md — see the bottom.

What

Two fixes and their honest documentation:

• (a) the corpus was truncated to 21%: a prefix was taken instead of sampling across the author's documents. A bisection-order sample now spends the 800k-character budget across all of them.
• (b) PROPN tokens are filtered out. Character names are statistically distinctive but they are not style: havisham and pemberley say nothing about how someone writes.

Why

Closes #100

⚠️ This branch does NOT meet its own Definition of Done, deliberately

The DoD required a top-10 overlap ≤3. The measured result is 5, which is worse than what #89 left. This was accepted by the project owner on 2026-07-28 and recorded with its reasoning in docs/decision_log.md.

The reason is arithmetic, not implementation. With three documents, countenance (29 occurrences in Dickens) and say (1,866) both have df = 3/3 and therefore identical idf = 1.000. TF-IDF collapses into raw frequency and countenance can never win. Proper nouns were the only thing discriminating; removing them (correct) raises the overlap.

Rejected alternatives: re-scoring with log-odds or IDF against general English, because it changes an algorithm that docs/style_features.md §4.1 declares closed; and reverting the PROPN filter, which buys the metric at the price of the defect the issue exists to remove.

Accepted consequence: the distinctive vocabulary shows common words. All three authors share know, little, make, say, time.

How IBM Bob helped

• Custom Mode used: none
• Bob session export: n/a

Claude Code agents. This branch went through an independent verifier in a cold context that never saw the diff and received only the Definition of Done. It found two things the executor had not covered:

1. The documentation asserted something false. style_features.md:305 still promised countenance/physiognomy/presently in the output, and :336 called that example "the standard the section sets for itself". None of the three appears even in the top-30. Both corrected, with a new Measured output block recording what the pipeline actually returns.
2. Half (b) had no test at all. The verifier proved it by deleting the entire PROPN filter — the suite stayed green. test_lemmatize_corpus_drops_proper_nouns now covers it.

Screenshots / output

Real top-10, measured with the seed's own code path over the full corpus/:

austen : say know think make come time good great look little
dickens: say know look come make man time little think hand
poe    : say make great time long man know day eye little

Positive control for the new test, in a disposable copy outside the repo:

PASS A — filter present:  1 passed in 14.37s
PASS B — filter removed:  1 failed
  AssertionError: proper nouns reached the TF-IDF input: ['havisham', 'pemberley', 'wemmick']

Suite: 219 passed, 3 skipped. Cost: 211-231 s against #89's 216 s (between -2% and +7%; the threshold was +50%).

Checklist

• Tests pass locally (make test)
• Lint passes (make lint)
• If schema changed (StyleProfile / Passport) → bumped version + updated docs/
• If new env var → added to .env.example
• If new/changed endpoint → updated docs/api_contract.yaml
• Documentation in the relevant README.md updated
• At least one teammate reviewed

Expected conflict in docs/decision_log.md. It is append-only and has four writers. Resolve by keeping all rows in chronological order — never by deleting someone else's.

Merging this requires a re-seed: every existing style_profiles row is invalidated.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
author-ai	Ready	Preview, Comment	Jul 28, 2026 4:43pm