task: record the five success signals over the first ten pull requests

[serina-mcfall]

Parent PRD

#109

Objective

A recorded, published measurement of the agent's first ten reviews against the five signals #109 names — including the adjudicator metric in isolation, which #109 reports no published system provides.

Definition of done

• For each of the first ten PRs, a comment on enh: agent workflow — PR review agent as a first-pass reviewer #109 records: precision, recall, cost in tokens and wall-clock, verification yield, and adjudicator accuracy in isolation.
• Adjudicator accuracy is recorded per finding it confirmed or refuted, against what the human concluded — not inferred from the end-to-end outcome. An end-to-end number cannot distinguish a good reviewer with a weak judge from the reverse.
• Recall is captured by asking the human reviewer what they found that the agent missed, not by assuming the agent's misses are unknowable.
• A verification yield at or near zero is called out rather than reported as success — it means the adjudication pass is not doing anything.
• The failures are published in the same place as the successes. A review agent whose reported quality only ever improves is not being measured.
• The record states its own sample size and does not describe ten PRs as a rate.

Impacted components

• enh: agent workflow — PR review agent as a first-pass reviewer #109 — the comments are the artifact
• whatever the publish task emits, which must carry enough per-finding detail to score afterwards

Out of scope

Proposing Phase 2. That needs its own ADR amending launchpad/AGENTS.md §5 rule 1, and its own evidence bar — this task produces the evidence, it does not argue from it.

Tuning the agent mid-measurement. A number produced while the thing being measured is changing is not a measurement.