Proposal: community rule library with local empirical gating (privacy-preserving federated skill transfer)

Motivation

Skill efficacy in this pipeline scales roughly with data volume. But the obvious fix, pooling users' session data, has a near-fatal problem: transcripts are among the most personal artifacts in computing (code, prompts, mistakes), and regex-level secret scrubbing (redact_secrets-style) is nowhere near sufficient for public sharing.

Key observation

What transfers between users at the skill level is not the transcript or the task it's the distilled rule. "Write commit subjects in imperative mood" is useful to thousands of people; the session that taught it is useful to no one but its author.

There is, however, a second tier where transcripts do have transferable value: the meta level. As people experiment with modified versions of the pipeline itself (miner variants, replay backends, gate policies the kind of evidence-chain iteration discussed in #151), a session corpus that demonstrably trained a successful directive becomes a reusable benchmark: anyone iterating on a pipeline variant can rerun it against the same corpus and compare what their variant mines, replays, and gates. That's transcript sharing with a fundamentally different consumer (pipeline developers, not skill consumers) and it would need to be strictly opt-in with real anonymization but it's worth naming as a distinct, longer-term tier of this proposal rather than conflating it with rule sharing.

For the rule-sharing tier, the gate architecture already handles untrusted candidates: every user's gate validates any proposed edit against their own held-out tasks before accepting it.

Proposal: community rule library + local empirical gating

• Define an import/export format for candidate edits (rule text + minimal metadata: category, provenance-free rationale, observed effect size).
• Users can publish distilled rules tiny, reviewable, low-risk artifacts to a shared library.
• Importing a rule places it in a probation state: it must pass the importer's own held-out gate before being accepted into their skill, exactly like a locally-mined candidate.

This is federated skill transfer with privacy built in: the shareable unit is the candidate edit, and trust comes from local empirical validation rather than from the publisher.

Bigger picture

Together with #154 (intent-level mining) and #155 (agentic replay), this is aimed at SkillOpt growing into an on-the-job-training system for specific job roles with the community library as the mechanism for learning from everyone's experience combined without sharing anyone's data.

Start by tracing the existing candidate-edit, miner, replay, and held-out gate paths; the issue does not name specific files or tests. Define the import/export boundary and probation flow around those entry points, then verify that imported rules are locally gated before acceptance and that no session transcripts are required for sharing.