# Update runbook

Static v1 separates filing, scoring, computing, and deployment. An intake assistant files candidates. The maintainer scores them, freezes the ledger and snapshot, and publishes. An automated build step may compute and package; it does not deploy.

## Candidate to ledger row

1. **W-1, intake.** The intake assistant records the claim as made, its URL, publisher, publication date, and candidate occurrence date. Use the incident monitors, regulator and court records, operator disclosures, primary reporting, or a submitted claim.
2. **W-2, Gate 0 window.** Verify occurrence and primary-publication dates separately. Gate 0 uses occurrence. The ledger effective date is the primary publication date. Reference-window events are history unless Methodology §11.5 expressly permits an exception.
3. **W-3, Gate 1 incident screen.** Decide whether the fact is harm, a near-miss, or a documented failure in deployment or live evaluation. Do not assign severity to a screen failure. Route other facts through §12.2 precedence or record R-NULL.
4. **W-4, Gate 2 eligibility.** Run E1-E8. A failure produces a public non-scoring row naming the failed check. E1 requires a first-party document, official dataset, statute, or named-instrument series. Reporting can open a candidate but cannot alone confirm S2+.
5. **W-5, severity and fit.** For C6, assign the realization class and test Annex A-C6 predicates element by element. Record evidence for accepted and rejected elements. Apply the confirmation, independence, borderline, and one-fact/one-category rules. If S2 is not met, C6 is only an annotation and precedence continues.
6. **W-6, step and row.** The maintainer assigns R-NULL, R-UP-1, R-UP-2, or the applicable existing R0.1 rule. Draft the §12.4 fields, annex section, disposition, rationale, and source links. Every row needs a real fetched source supporting its description. Keep non-scoring rows. Never score a claim that the assembled evidence does not establish.
7. **W-7, freeze and publish.** Freeze accepted rows in `data/event-ledger-v1.1.json`, or the next versioned ledger. Published rows are immutable.
8. **W-8, corrections.** Append a timestamped correction for new primary evidence or factual error. Never rewrite a frozen row. Recompute affected category state and INC-CLOCK under Annex §A.6.1. Quarterly re-adjudication is the venue for judgment-only review; the operator decides contested calls.

For S2+, assemble occurrence and AI-implication primaries, independent corroboration, and primary evidence for every severity element. S3 needs a second independent primary confirming both occurrence and severity. Record what was fetched, the supporting line, and missing evidence.

## Compute and check

Keep Tier A frozen between quarterly reviews. Do not add a RESET row unless §11.7 authorizes one. From the repository root:

```bash
python3 compute/snapshot.py --snapshot-id v1.1 --as-of 2026-10-23 --ledger data/event-ledger-v1.1.json
python3 compute/snapshot.py --snapshot-id v1.1 --check
python3 compute/snapshot.py --snapshot-id v1.0 --check
python3 compute/test_snapshot.py
python3 scripts/publish_site.py
python3 scripts/test_publish_site.py
python3 scripts/publish_site.py --check
```

The compute writes `snapshots/v1.1.json` and the frontpage mockups. The publisher writes `site/snapshot.json`, the v1.0 archive, both ledgers, and public docs. A check must pass before handoff. Confirm the stamp, headline, band, M, INC-CLOCK, D-INC-CLEAR, all row dispositions, and source links against the inputs.

## Publishing boundary

The maintainer reviews the diff and performs the deploy. Compute and packaging are not publication. No automated build step runs Wrangler, Cloudflare commands, a commit, a push, or a deploy.