[appleweiping]

Scope

Add an original bounded trainable keyphrase-centroid expertise workflow. It consumes the existing keyphrase extractor output and caller-authored disjoint train/validation triplets, learns mean term-vector pairwise logistic rankings, selects the first best validation-MAP epoch, and emits match-ready holdout affinities and normalized inputs for the existing constrained assignment CLI.

Evidence

• Independent one-step BCE gradient and ranking-MAP hand oracles, train-only vocabulary and updates, disjoint validation/holdout submission IDs, hard-conflict checks, deterministic replay and exact-source hashes.
• Independent review found that replaceable in-memory fields could diverge from captured raw labels/features. The second signed commit reloads and compares every normalized input before training or scoring and adds forged-label/feature no-publication regressions.
• Windows full suite: 959 passed, 1 symlink skip; Linux full suite: 960 passed. Pure-source branch coverage: 2435/2696 = 90.32%.
• Ruff, mypy, Bandit, dependency audit, build, isolated wheel/sdist CLI installs and synthetic assignment smoke passed locally.
• Both commits have verified ED25519 signatures and DCO sign-offs.

Limits

This is a local dependency-free expertise model, not the frozen OpenReview Expertise PyTorch pipeline, SPECTER, TPMS, live OpenReview write-back, official gold results or whole-repository parity. Source digests detect byte changes; they are not authenticity signatures or proof that caller labels are trustworthy.

Assisted-by: OpenAI Codex