contributor-trust

name	contributor-trust
description	Aggregate the contributor trust ledger into a shrunk-and-decayed reliability score that weights review priority and gate intensity — never a reason to skip a gate. Band A, no LLM: classification happens upstream at close, this is arithmetic over an existing event log.

contributor-trust

When to load: computing the contributor reliability score that triage-scoreboard consumes as a mechanical dimension. Implements the trust-ledger half of contributor recognition. Band A — deterministic, read-only, no LLM. Steward role. Loaded on demand during a scoreboard refresh, not wired as its own scheduled job: the scoreboard recomputes mechanical dimensions fresh each run, so an on-demand score is always current.

Precondition: a ledger path is configured. Read trust.ledger_path from config.yaml; blank → there is no ledger to aggregate, stop (silent no-op).

Config: the trust: block — ledger_path, prior, shrinkage, half_life_days. This skill never writes events: it reads the ledger and publishes the trust map. release-pipeline writes an event at close.

The ledger

An append-only record of one event per PR's terminal outcome. A bounce that later converges is one good-rework event — not a negative followed by a positive. Intermediate bounces are not events at all; only the terminal state is written.

A terminal outcome can still be corrected — a clean-ship later found to be a regression — by appending a new event for the same pr, never by rewriting the old one. The log stays append-only; the last-appended event for a PR wins at read time (see Step 2). "One event per PR" is therefore an effective-state rule, not a promise that the raw log holds a single line per PR.

Each event: pr (repo#number) · author (handle) · outcome (below) · at (UTC date).

Class	Outcomes	Scoring
positive	clean-ship, shipped-with-minor-fixes, good-rework	value 1
negative	not-fixed, went-stale, regression, critical-flaw	value 0
unscored	neutral, superseded, shelved, wip, concept-call	excluded — recorded, never judged

Steps

1. Read config.yaml. Load trust.ledger_path, prior, shrinkage (k), half_life_days. Blank ledger_path → silent no-op.

2. Read the ledger, collapse to one event per pr, then group by author. The last-appended event for a PR wins (last-write-wins over the append-only log), so a later correction supersedes an earlier entry and no PR is ever counted twice. Collapse by append order, not by at — at is the event's terminal date the decay in Step 4 needs, and a correction may legitimately carry a recent at. Malformed or unknown-outcome events are surfaced, never guessed — an event you can't classify is not an event you may score.

3. Drop unscored events from the computation. They stay in the ledger; they never move a score and never contribute to effective_n.

4. Value and decay each scored event. v = 1 (positive) or 0 (negative). w = 0.5 ** (age_days / half_life_days) — an event's weight halves every half_life_days.

5. Shrink toward the prior:

   score = (Σ wᵢ·vᵢ + k·prior) / (Σ wᵢ + k)
   

   Both mandated properties fall out of this one expression rather than being bolted on: a contributor with zero scored events scores exactly prior (the Σw = 0 case reduces to k·prior / k), so the first-timer needs no special branch; and a contributor with one event barely leaves the prior.

6. Publish the trust map, keyed by handle: {score, events_scored, effective_n, last_event_at}. Publish effective_n (= Σ wᵢ, the decayed event count) alongside every score — it is what tells a consumer how much the score is worth. A 0.9 backed by an effective_n of 0.2 is almost entirely prior.

7. Never write. This skill does not append, edit, or reconcile the ledger. Same ledger + same clock → same map.

Pitfalls

• The score never gates. It weights review priority and gate intensity. It is an input to triage, never an excuse to skip the authoritative gate — every change passes the gates regardless of who wrote it. A trust score that waves a change through has become the thing this system exists to prevent.
• The trust map is internal triage state — don't render it publicly. It carries negative signal and is an input to your review intensity, not a public leaderboard. Publishing per-contributor reliability scores would chill exactly the experimentation contributor recognition exists to reward. Keep it beside the scoreboard as internal state, out of any rendered or public surface.
• Today the score is upward-biased — say so, don't compute past it. The only declared ledger writer is release-pipeline, which runs only on the ship path (it merges an all-clear PR). Nothing currently writes went-stale, not-fixed, regression, or critical-flaw. So the ledger accrues only positives at every volume — a high score is inflated whether effective_n is low or high, because effective_n measures how much positive evidence exists, not freedom from bias. A high score can mean "genuinely reliable" or "failures were never recorded," and nothing in the ledger distinguishes them. Treat every score as low-confidence until a reconcile job records non-ship outcomes.
• Don't over-judge one data point. This is enforced by shrinkage, not by a special case. If you find yourself adding an if events == 1 branch, the prior is doing the job already.
• Unscored ≠ negative. Excluding an event and penalizing it are different things. A superseded or shelved PR must not cost the author — that's how you teach contributors that experimenting is expensive.
• One terminal event per PR (write side). Write a bounce-then-converge as the single good-rework outcome it is — don't emit a negative and then a positive. Read-side last-write-wins (Step 2) is a backstop against accidental duplicates and the mechanism for deliberate corrections; it is not permission to append sloppily on the write side.
• Guessing an outcome. Classification is upstream judgment made at close. This skill classifies nothing — it aggregates. If an event's outcome is missing or unrecognized, surface it; never infer one from the PR.

Verification

• Same ledger + same clock → the same trust map (deterministic; no model in the loop).
• A contributor with zero scored events scores exactly prior.
• A contributor with one recent positive event lands near prior, not at 1.0 (shrinkage holds: with k=3.0, prior=0.5, a single fresh positive gives (1 + 1.5) / 4 = 0.625).
• Unscored events move no score and add nothing to effective_n.
• A corrected outcome supersedes the earlier one: appending a regression event for a PR that already has a clean-ship scores that PR once — as the regression — because Step 2 keeps only the last-appended event per pr. Two events for one PR never double-count.
• Ageing an event several half-lives drives its weight toward zero (5 half-lives → 0.031), so an all-stale ledger reverts to prior.
• The ledger is unchanged after a run (read-only).
• Blank trust.ledger_path → silent no-op.
• Every published score carries its effective_n.

Related: contributor recognition · the triage scoreboard · the reconcile pattern.