Promotion audit

The working ledger for ADR-047: finding entries whose evidence has outgrown their status, a few at a time, so that any session can pick up where the last one stopped.

File	What it holds	Who writes it
queue.yaml	Every triaged candidate with its rubric score and state, the per-wave yield, and the handoff block	updated in every batch; bump version, append to history
decisions/<CODE>.yaml	One review: disposition, accepted and excluded scope, evidence, blockers, reviewed_revision, next action	written once, when the review is reached
batches/batch-NNN.yaml	What one batch reviewed, its outcomes, its yield, and what it learned	one per batch
NOTES.md	Working notes: patterns, traps, open threads	append freely
check.py	Fails if the queue and the decision files disagree	run before every commit
triage.py	Scores wave-1 candidates by the rubric, from metadata	run on a freshly fetched origin/main when re-triaging
triage_bundling.py	Scores wave-2 mathematics THEORYs for bundling (triage-003)	as above
triage_wave3.py	Scores Proposed THEORYs outside the queue for bundling and reading depth, routes formal results to re-derivation, and draws a control sample (triage-004, -005)	as above

Resuming

Read queue.yaml’s handoff, then NOTES.md. If in_review names a candidate, finish that one first. Then follow next_step. A continuation prompt that works on its own:

Continue the Nucleation promotion audit from record/promotion-audit/queue.yaml. Review the next five pending candidates in queue order, write one decision file per candidate as each is reached, update the queue and its hand-off, run check.py, commit, and stop. Do not change any entry’s status without a decision file that calls for it.

A batch, step by step

1. Set handoff.batch and mark the first candidate in_review. Commit.
2. Review it: read the entry in full and the records it rests on, at most two rounds of retrieval. Decide one of promote, narrow-and-promote, split, retain-proposed, needs-correction or defer. Every disposition but a promotion names a specific blocker.
3. Write decisions/<CODE>.yaml, set the candidate to decided with its disposition, and move in_review on. Commit after each decision when a session might end.
4. After at most five, write batches/batch-NNN.yaml, update the wave counts, bump the queue’s version, rewrite handoff, and run python3 check.py. Commit and stop.

Promotions are separate pull requests of at most three entries. Each cites the decision file it carries out. A decision is stale once the entry has changed since its reviewed_revision.

The guard

python3 -m pytest record/promotion-audit/test_check.py fires check.py against a consistent ledger and against each kind of drift. It was also fired once against the real ledger on 2026-10-19 (DP-4): with decisions/CLAIM-270.yaml moved away, it reported “CLAIM-270 is decided in the queue but has no decision file” and exited 1. With the file restored it passed.

Closed

ADR-057 replaces this procedure: the worklist is derived on each pass by record/status-review/worklist.py, and nothing here is updated. This directory stays as the record of the first audit, 2026-10-19 to 2026-10-21.