[wangjohn]

A passage can sound unlike a writer even when its aggregate style-distance label says “close.” This adds sentence-level rhetorical evidence and an optional personal profile that reuses existing passage ratings.

The four measurements cover hypothetical reader prompts, observation prefaces, short/long sentence pairs, and explicit lesson statements. Each returns counts, rates per 1,000 prose words, and exact evidence spans in normalized text. Baseline Delta/contrast features score the original draft, preserving its Markdown structure.

feedback-build learns separate voice and revision-effort associations from rated JSONL passages, the existing scores, and the new measurements. feedback-score reports those estimates with evidence. Profiles embed their base reference and preserve rating IDs, notes, grouping, and text hashes so later judgments can build on earlier feedback. Ordinary build/score behavior is unchanged; there are no new runtime dependencies.

Topic metadata is resolved consistently across every passage from a source before fitting; missing topic labels inherit the source’s supplied topic, conflicting assignments are rejected, and source/topic namespaces cannot collide. Grouped development diagnostics hold related passages out together and fit scaling on training groups only. Each target is disabled when it does not beat a mean-only baseline, and estimates are withheld below 75 words.

File I/O uses explicit UTF-8. Non-object JSON profiles produce a normal invalid-profile error. The experimental feedback format is version 2: earlier version-1 feedback profiles must be rebuilt from saved ratings and the base reference because baseline feature semantics changed. Existing judgments need not be repeated. Estimates remain experimental personal associations, not AI-authorship probabilities; feedback used to design features is not independent validation.

Validation:

• Full suite without spaCy: pytest -n 4 — 895 passed, 46 skipped.
• Focused feedback/CLI tests: 30 passed, including source identity with missing/conflicting topic metadata, namespace collisions, Markdown baseline preservation, simulated CP1252 file I/O, invalid JSON roots, and old-format rejection.
• Evidence offsets, false-positive controls, grouped holdouts, target gating, serialization, and CLI round trips remain covered.
• Ruff lint and formatting checks pass across the repository.
• Pyright passes using the isolated development environment.