Engine Review

• output/engine_audit_c{XX}.json — ailment patterns (Phase 38.1)
• output/engine_anomalies_c{XX}.json — anomalies with triage paths (Phase 38.7)
• output/baseline_briefs_c{XX}.json — auto-generated event briefs (Phase 38.8)

If any file is missing or invalid JSON, stop. Report the failure to the user before drafting anything — a missing detector output means the framing would be guessing.

Step 2 — Read the three JSON files

Read all three. Don't re-scan sheets. The auditor has already done the deterministic work — including, after Phase 38.2/3/4 (S146 spine step 5), the mitigator check, remedy recommendation, and Tribune framing per pattern.

Audit JSON patterns[] — each pattern carries:

• Detection fields (Phase 38.1): type (stuck-initiative / repeating-event / math-imbalance / cascade-failure / writeback-drift / production-imbalance / improvement / incoherence / anomaly), severity, cyclesInState, affectedEntities (citizens / neighborhoods / initiatives / councilSeats), evidence (sheet, rows, fields), description, detectorVersion.
• Mitigator state (Phase 38.2): mitigatorState.exists, mitigators[] (per-initiative effectsFiring + effectEvidence), gap (no-mitigator | mitigator-stuck | mitigator-firing-but-insufficient | remedy-working), recommendedAction, ailmentCategory.
• Remedy path (Phase 38.3): remedyPath.worldSide[] (advance-initiative / propose-new-initiative / character-intervention / council-vote / mayoral-pressure, each with type/target/action/rationale/expectedEngineEffect), and remedyPath.techSide.bugReport (only populated when writeback chain is structurally broken).
• Tribune framing (Phase 38.4): tribuneFraming.storyHandles (per desk: civic / business / culture / sports / letters), tribuneFraming.threeLayerCoverage (engine / simulation / user-actions one-liners pre-written), suggestedFrontPage boolean, capabilityHooks (literal phrases Phase 39.1 grades against).
• Measurement (Phase 38.5): measurement.available. When true: priorCycle, expectedField, expected, observed, delta, verdict (remedy-firing-as-expected | remedy-firing-insufficient | remedy-not-firing | remedy-overshot), priorRemedyType. When false: reason (no-prior-audit | no-prior-match | prior-had-no-expectation). Top-level measurementHistory[] aggregates these across patterns for cross-cycle learning.

Anomalies JSON anomalies[] — triagePath (cover-as-story | route-to-engine-debug | suppress-until-verified), confidence, historicalContext. On first run with no prior audit, anomalies[] may be empty — expected, not a failure.

Baseline-briefs JSON briefs[] — id, eventClass, subjectIds, neighborhood, cycle, facts, threeLayerHandle, tier (default C), promotionHints.

Step 3 — Frame each ailment as a 7-field brief

For every pattern (excluding type: 'improvement' and type: 'anomaly' — those go to their own sections), produce a seven-field markdown block. Most fields now have structured source data from the auditor's enrichers — your job is to translate the structured fields into Tribune voice, not to discover or invent.

If tribuneFraming.suggestedFrontPage: true, mark the brief with FRONT PAGE CANDIDATE at the top — sift uses this to seed front-page scoring.

If tribuneFraming.capabilityHooks is non-empty, list the hooks at the bottom of the brief — they are the literal phrases Phase 39.1's assertHighestSeverityAilmentCoveredOnFrontPage grades against. Sift can pass them through to the relevant reporter as required coverage tokens.

Watch for first-run startup artifacts

The first audit run after S146 derived cyclesInState from each initiative's LastUpdated date string. That can produce surprisingly large numbers (Temescal at cyclesInState=88 from "3/25/2026"). The number self-corrects once a second audit JSON exists to diff against. Don't frame these inflated counts as "stuck for 88 cycles" in narrative copy. Read cyclesInState qualitatively on first run ("stuck, with the design phase predating the current build of the auditor"); use the precise number from the second cycle onward.

When the structured fields are empty

Phase 38.2 gap: 'remedy-working' doesn't surface until a second audit cycle exists with measurable deltas (per engine terminal's §17 acceptance note — "inherent to the cross-cycle design"). On the first run after any pattern, effectEvidence.verdict will commonly read no-history. That's expected — write the brief as "early signal, will measure next cycle" rather than treating it as a failure of the mitigator check.

Step 4 — Anomaly triage section

For each entry in anomalies[], write one paragraph:

• The anomaly (what changed, by how much)
• The triage call: cover as story / route to engine debug / suppress until verified
• One-line reasoning, citing confidence and historicalContext

If anomalies[] is empty, write a one-line "No anomalies flagged this cycle." Don't fabricate.

Note: routing an anomaly to engine-debug means writing a separate brief at output/engine_anomalies_c{XX}_followup.md flagged for the engine terminal to investigate. The anomaly is suppressed from the edition until cleared.

Step 5 — Baseline brief pass-through

Don't re-write the baseline briefs. The auditor already structured them for sift. In the engine-review markdown, include:

• Total brief count, broken down by byEventClass from the JSON's summary
• Count with promotionHints (these are sift's promotion candidates)
• One-line note on which neighborhoods or active ailments the briefs cluster around
• Reference: output/baseline_briefs_c{XX}.json

/sift reads the JSON directly when deciding promote / publish-as-baseline / suppress per Phase 39.9 tiered review.

Known limitation — EventType taxonomy

As of S146, most events in WorldEvents_V3_Ledger resolve to eventType: misc-event, so subjectIds on most baseline briefs is []. This blocks citizen-attributed promotion (e.g., a death brief that should hint at Beverly Hayes by POP-ID can't, because the event isn't typed as citizen-death). Sift can still promote on neighborhood + ailment overlap from promotionHints, but the citizen-specific path is degraded until the engine adds a richer EventType breakdown. Tracked in ROLLOUT_PLAN as a follow-up item.

Step 6 — Improvements section

For every type: 'improvement' in patterns[], write one short paragraph: what's working and why. Don't bury good news. Phase 38.4 (S146 spine step 5) now also threads improvement-side handles into tribuneFraming.storyHandles when a positive trend has a named cause — surface those as story candidates for sift, same format as ailment briefs but tagged IMPROVEMENT at the top.

Step 7 — Measurement check (cycles after the first)

The audit JSON carries measurement state directly: every pattern has a measurement field; the JSON has a top-level measurementHistory[] rollup. Don't read prior engine_review_*.md files — the structured fields already record what fired and what didn't.

Render in three parts:

1. Per-pattern table. One row per pattern with measurement.available === true: pattern type, affected entity, prior remedy type, expected, observed, verdict. If every pattern is available: false because no prior audit exists yet (reason: 'no-prior-audit'), write the single line First review — no prior to compare. Measurement loop will activate on next cycle. and skip the table.

2. Remedy-type track record. Group measurementHistory[] by priorRemedyType, count verdicts (remedy-firing-as-expected, remedy-firing-insufficient, remedy-not-firing, remedy-overshot). One row per type. This is the city's multi-cycle learning signal — what kind of intervention has actually moved the world.

3. Win callout. If any pattern this cycle reads verdict: 'remedy-firing-as-expected' AND its prior-cycle entry in measurementHistory[] was remedy-not-firing, name it in voice: the gap closed — that's a story candidate, not just a data point. One line under the table.

For available: false rows other than the first-run case, gate display on measurement.reason: no-prior-match → render as —; prior-had-no-expectation → render as no prior expectation. Don't omit them silently.

Output File

Write to output/engine_review_c{XX}.md:

# Engine Review — Cycle {XX}

**Cycle:** {XX} | **Date:** {timestamp}
**Auditor version:** {from JSON detectorVersions.engineAuditor}
**Source files:**
- `output/engine_audit_c{XX}.json` — {N} patterns
- `output/engine_anomalies_c{XX}.json` — {N} anomalies
- `output/baseline_briefs_c{XX}.json` — {N} briefs

## Ailments

### 1. [In-world symptom headline]
- **Tech diagnosis:** [grounded in evidence.fields]
- **Existing mitigators:** [from affectedEntities.initiatives]
- **Why working/not:** [gap analysis]
- **Remedy path:** [world-side preferred, tech-side fallback]
- **Tribune framing:** [three-layer story handles]
- **Measure next cycle:** [specific fields/milestones]

### 2. [next ailment]
...

## Anomalies

### [triage-by-triage, or "No anomalies flagged this cycle"]

## Improvements

### [what's working and why]

## Baseline Briefs (sift input)

- Total: {N} briefs ({byEventClass breakdown})
- With promotion hints: {N}
- Cluster note: {neighborhood / ailment overlaps}
- Source: `output/baseline_briefs_c{XX}.json`

## Measurement Check (from previous review)

| Pattern | Affected | Prior remedy | Expected | Observed | Verdict |
|---|---|---|---|---|---|
| [type] | [neighborhood / initiative] | [priorRemedyType] | [expected] | [observed] | [verdict] |

(First-run case: `First review — no prior to compare. Measurement loop will activate on next cycle.`)

[Win callout, if any pattern flipped from `remedy-not-firing` last cycle to `remedy-firing-as-expected` this cycle.]

### Remedy-type track record

| Remedy type | Firing-as-expected | Firing-insufficient | Not-firing | Overshot |
|---|---|---|---|---|
| [priorRemedyType] | [N] | [N] | [N] | [N] |

## Summary

- Ailments: {count} ({severity breakdown})
- Anomalies: {count} ({triage breakdown})
- Improvements: {count}
- Baseline briefs: {count} ({with-promotion-hints count})
- Measurements: {firing-as-expected count} / {total measured} firing as expected; {not-firing count} not firing

Where This Sits

Step 4 in the run-cycle chain. After pre-flight, pre-mortem, and cycle execution. Before build-world-summary. World summary reads this file and incorporates the framing; sift reads the same file plus the baseline-briefs JSON for editorial planning.

Why this rewrite (S146 — two passes)

First pass (after Phase 38.1 + 38.7 + 38.8): scanned 11 sheets by hand → consume audit + anomalies + baseline-briefs JSON. Detection moved out of the skill into scripts/engineAuditor.js. Skill became the framer.

Second pass (after Phase 38.2 + 38.3 + 38.4, this rewrite): even the framing logic mostly moves into structured fields. The auditor now writes mitigatorState, remedyPath, and tribuneFraming per pattern — which is what the skill was previously synthesizing by hand. The skill's remaining job: translate structured engine fields into Tribune voice, surface front-page candidates, and gate anomalies. Most assertions about "what to write" come from tribuneFraming.threeLayerCoverage directly.

Determinism in code, judgment in the skill — and the judgment surface keeps shrinking as the auditor learns. Sources: docs/engine/PHASE_38_PLAN.md §6 (first rewrite) and §16.5 (this rewrite).