Promotion: pre-freeze batch — stage enablement, claim scoring, sal-v2 groundwork

[modelmirror]

Promotion batch (draft — held until the batch completes)

Do not merge yet. This drafts the next staging→main promotion so the batch is visible as it assembles. The maintainer-dispatched promote workflow re-runs the gate, sync and freshness checks at merge time regardless. Merges as a merge commit (promotion ancestry rule).

Landed on staging

Stage enablement — cert, interim and merits are one pipeline now

• feat(events): carry the decision stage on the event surface #1011 — stage on the event surface (cert / interim / merits), stamped at minting, ranged-blob tolerant
• feat(leaderboard): stage-routed attribution and the stage stratum axis #1014 — stage-routed outcome attribution; the leaderboard's ranked board becomes explicitly cert-stage, non-cert cells aggregate into an unranked per-stage block, and skill_scored publishes the skill mean's true denominator
• feat(events): relabel application-docket baselines to motion/interim #1016 — application baselines relabeled to evt-motion-disposition / interim, with an idempotent migration
• feat(interim): open the interim predict gates, quota'd by the salience reserve #1020 — the interim predict path: scope narrowed to substantive applications, queue flipped, fan-out widened, interim_reserve_slots: 5 enforced inside N, interim outcomes recorded
• feat(prompts): the interim stage sections and an offline interim cell #1022 — stage-conditional predict/evaluate prompt contracts + an offline interim cell in the cascade smoke
• feat(pipeline): mint the open merits event at cert grant; retain granted dockets in rotation #1018 — a cert grant mints an open merits event and the granted docket keeps polling through judgment
• feat(corpus): merits-judgment extraction, backfill, and statpack stage section #1015 — merits judgment parser, corpus columns, backfill CLI, statpack merits section
• fix(merits): exclude cert-order grants from the merits population #1023 — the merits population excludes grants that decide in the cert order (GVR / summary reversal)
• feat(merits): the scoreable merits cell contract and its registered baseline #1025 — the scoreable merits cell contract: Prediction.judgment, P(disturbed) as the Brier axis, mandatory votes, merits-v1, the registered baseline, and Evaluation.base_rate_salience_version
• feat(prompts): the merits stage contract, and the merits fan-out #1027 — merits prompt sections, the merits fan-out, and an offline merits cell in the cascade smoke
• feat(statpack): add the interim-docket stage axis #1012 — the statpack's interim-docket stage axis

Claim scoring and its surface

• feat(evaluate): declare and score the cert mechanical claim set #1013 — the cert-v1 mechanical claim set, declared and harness-scored (moves every actor digest)
• feat(metrics): the claim-score board and judge-validation artifact #1017 — metrics/claim-scores.json and the judge-validation tau-b, suppressed below n=10
• chore(workflows): regenerate the claim-score board with the other metrics #1032 — the claim-score board regenerates with the other metrics, so it cannot go stale the moment the first claim block lands

Scoring the predictors, not just the predictions

• feat(evaluate): grade a candidate the evaluator cannot name #1029 — blind grading: the evaluator sees an alias, not a predictor id; the harness un-aliases before the cell is stamped
• feat(metrics): realized-Term skill beside prior-Term skill, and a corrected aggregation #1030 — realized-Term skill beside prior-Term skill, and skill-of-means replacing mean-of-ratios on both columns (the old aggregation ranked an always-deny constant above an honest forecaster)
• feat(semantic): the semantic-v0 scoring seam, built and explicitly provisional #1028 — the semantic-scoring seam, built and tested, explicitly alpha: the wiring is pre-registered as provisional, not the methodology

Security

• fix(retrieval): redact credential-shaped values at capture #1024 — credential-shaped values redacted at transcript capture, before truncation (moves digests; touches SECURITY.md)

Workflows (maintainer-merged)

• chore(workflows): wire blind grading, and the two corpus convergence sweeps #1031 — the blind-grading stage/un-alias steps, plus the two corpus convergence sweeps (application relabel, merits judgment backfill) running in the writer lane that owns the corpus pointer
• chore(workflows): regenerate the claim-score board with the other metrics #1032 — claim-scores in the metrics refresh

Ledger hygiene

• fix(ledger): reopen outcomes copied from a sibling case-baseline event #1034 — reopen-misattributed-outcomes: the repair for outcomes copied from a case-baseline sibling
• fix(events): a SCOTUS appeal entry collapses into the case baseline #1035 — a SCOTUS appeal entry collapses into the case baseline (the mint guard), plus remove-unmintable-events for the row already written that way

Both change no data by merging; their --apply + data/ commit is a maintainer step, because neither collect jail admits deleting an outcome.json.

Docs

• docs: describe the artifacts a predict cell produces #1026 — docs/predicted-artifacts.md: what one prediction consists of, every example machine-validated

Earlier in the window: #1000–#1006 (corpus fixes, version-pinned band pool, pre-registration docs, lookback pin, forecastable-events seam, sal-v2 intent).

Still to join

• sal-v2 (salience: sal-v1 conditions on signals a petition does not have on arrival — rework for prospective use #963) — gated on the run: re-walk OT2017-2025 to denial-completeness (historical refresh) #967 re-walk finishing, then a fresh gate replay, arrival-feature measurement on the denial-complete frame, and stats review before any constant is fixed

Issues this batch closes

#903, #904, #1007, #1008, #1009 already hand-closed on staging merge; #864, #907 and #963 close after promotion. #1010 and #1033 close when the two ledger applies land as a data/ commit on main — the tools are merged; the records are not yet removed.

Maintainer actions

1. The two corpus-lane sweeps now run themselves. chore(workflows): wire blind grading, and the two corpus convergence sweeps #1031 put the application relabel and the merits judgment backfill in run-seed's writer lane, so neither needs a dispatch — they converge on the schedule against the pointer the writers own (corpus-push is last-writer-wins, so a dev-checkout run would silently discard concurrent writes).
2. The two ledger applies (data: scotus/72460453's stay-motion event carries the petition's GVR as its outcome #1010, data: scotus/63195851's spurious appeal event duplicates the petition's cert grant #1033), as one data/ commit, in this order:

   uv run fedcourts remove-unmintable-events --apply       # scotus/63195851
   uv run fedcourts reopen-misattributed-outcomes --apply  # scotus/72460453
   # commit data/ to main, then corpus-push
   

   Removal first: it drops that case's whole event directory, which also clears the two triage lines the reopen sweep otherwise prints for it. Not lane-able — the change set deletes an outcome.json and rewrites an event.yaml, which assert_within_jail (adds only) and assert_cleanup_within_jail (deletes under predictions/ only) both refuse. Independent of the promotion; touches only data/.
3. After promotion: flip the evaluate handoff variable back on (held at 0 so the owed gradings do not re-run against a main that lacks the redaction fix), then fresh fedcourts process-digest --all on the promoted tree → paste into chore(metrics): freeze the proc-v1 process digests #991 → merge-commit chore(metrics): freeze the proc-v1 process digests #991 → bless (metrics: bless the settled process digest into the frozen set and record the freeze commit #907).