Promotion tuner

Since 19 October one comparison that reads step promotes. This skill asks whether a tree's rule supports the recorded promotions and how a candidate rule reads the same solves. The comparison and the limits on claims belong to AGENTS.md, sections 7 and 9. A replay starts no solve, calls no model and writes nothing of a campaign's.

The command

bun .claude/skills/promotion-tuner/scripts/replay.ts --tree <worktree> \
  [--logs <dir>] [--since <date>] [--name <campaign>]... [--margin <n>] [--level <n>] [--totals]

• --tree names the worktree whose kernel reads the record. The totals name its commit and whether src/ has uncommitted changes.
• The logs are <logs>/<worktree>/.ana/d2/<name>/log.jsonl. By default --logs is the parent of the repository's main checkout. The largest copy of each campaign name is read. Seeds with seed.json, names beginning sim-, rh- or test-, and rehearsals with a faux/ seat or solver are left out (campaignLogs, rehearsed in scripts/tree.ts).
• --since selects campaigns whose last event is at or after the date, five days back by default. Repeated --name selects named campaigns regardless of date.
• --level replaces the recorded level for promotions and A/A controls.
• --margin is for trees whose pairs still carry an objective margin. The current tree has none and ignores it; the totals count affected readings where an objective has no margin.
• --totals prints only the totals.

Live logs grow. Read the baseline and candidate close together and record the minute of each.

The script imports the selected tree's read, lastReports, pairsReported, pairsOf, pairings, flipP, stepOf, billedBetween and BILLED. Current lastReports excludes old ladder reports. Old event fields remain readable, as AGENTS.md section 5, habit 4 requires.

The rule is feature-detected (loadTree): promotes(reports, level) where exported, otherwise readTwice(reports, level, trigger) on trees that retain it, otherwise the latest report's stepOf reading. Until 19 October a trigger bought a repeat and the two readings were joined. attemptsOf with triggered groups those older attempts; without both exports, each reading is one attempt. The recorded trigger is used only by trees that read it, with 0.25 as its fallback; the level falls back to 0.02 where the comparison's start records none.

What to read

For each pair of harnesses, old digest first, the script prints:

• The first point at which the selected rule promotes, the readings it rests on, their W-L-T tallies and p values. For an older two-stage rule it also prints joined: each named task counted once, with its differences summed.
• rightful for a recorded promotion the rule supports with neither an A/A step nor a later reading that lost more tasks than it won; in doubt with those reasons; or added by this rule for a pair the campaign did not promote. Inspect the A/A and later readings beside an added pair too.
• Each unsuccessful attempt and its shortfall. The current rule reports a scored set, frontier and retained tasks read in one test, short of its level; a tree of 12 to 20 October may also report retained tasks that regressed. Older two-stage trees report whether a repeat was bought, whether it was taken, and whether the joined reading cleared the level. NOT PROMOTED by this rule marks a recorded promotion the selected rule rejects.
• The untied frontier tasks won in every pairing, won in some, and lost. This count changes no decision.
• Each promoted epoch and the open epoch: hours and billed dollars, split into Examiner, comparisons and Builder time (scripts/time.ts).

The totals count recorded promotions, pairs the rule promotes, readings, A/A controls, readings that cleared the level alone and their later readings, and the median and range of promoted epochs' hours and dollars. A report without a frontier cannot be replayed as a promotion.

Read the two reconciliation counts first: frontier readings and A/A controls that read otherwise than stored. At the tree and terms that wrote a report both should be zero. A difference must be explained by the candidate change or a replay defect. Under --margin the frontier count says how many readings the margin moved; the A/A reconciliation is printed only at the recorded margin.

An A/A is the reference against itself on mirrored halves of its solves (pairsOf with halves: "mirrored"). It is available at four or more solves per side, with an even count. The script tests either direction at half the level, as controlled in src/builder/decide.ts does for the epoch's control. An A/A step is false by construction.

rightful means the recorded solves support the promotion under the selected rule. It cannot detect a lucky draw; the A/A, the false-promotion estimate and a later fresh reading provide other evidence (AGENTS.md, factor 2). Name the rule when only one rule supports a promotion.

Tuning the rule

The level is the operator's condition. A change to the measuring rule is no harness gain.

1. Run the baseline on main's tree, both whole and with --totals, into the scratchpad.
2. Sweep --level first where that answers the question. --margin can compare older trees that still read it. A code change gets its own worktree from origin/main, following stack-pr, and changes one thing: for example stepOf or the pairing of valid answers. A candidate can export promotes(reports, level) for another rule.
3. Run the same table on the candidate close to the baseline reading.
4. Accept a candidate only where promotions rise or hours per promotion fall, no recorded A/A reads a step at its terms, and no added promotion is contradicted by a later reading of its pair. Count added promotions whose pair has no later reading as untested.
5. Deliver both tables in the pull request as counts, with both reconciliation counts explained.
6. Confirm live only on the operator's word, with a fresh campaign from a prompt at the candidate's commit and a prediction recorded before the first solve (AGENTS.md, section 10, item 9). Continuing a campaign at another commit measures no climb.

Limits of the replay:

• An earlier promotion would change the reference, the Examiner's tasks and later comparisons. The replay reads pairs on the recorded path, not a campaign under the candidate.
• A reading never bought has no solves to replay.
• Reports before 14 October may name no task: their stored differences stand and a margin cannot move them. Reports before 12 October hold no frontier and are left unread.
• In a led campaign, Examiner sessions and comparisons run inside the Builder conversation. Builder alone subtracts their union; neither is time outside all three. A Builder session across a promotion bills where it ends. billedBetween also meters unfinished work, so the script prints the tree's sum less the parts to expose any difference.

The wall and the result

Keep the log on the exam side (AGENTS.md, sections 5 and 8). The script prints counts, session ids, short digests and campaign names, never task contents, answers, measures or checks. Only those aggregate results belong in a pull request or this skill.

Report, in order:

1. Both trees, their commits and the minute each was read.
2. Both reconciliation counts on each tree.
3. Recorded promotions: rightful, in doubt with reasons, and rejected by the selected rule.
4. Candidate totals beside the baseline, with each condition of tuning step 4 and its count.
5. Campaigns left out, reports read from stored differences, and live counts that may move.