Read docs/editorial-style.md; the coin-summary register owns voice, title style, length, and closing structure. Follow Adding a Stablecoin §6b for addition-time coverage.

AI Summary Writer

Write StablecoinAiSummary entries in data/ai-summaries.json; shared/types/editorial.ts owns the schema. Read the base coin, all sidecars for that ID, and the field catalog in reference.md. For new/high-profile/refresh work, inspect the live detail page with available browser capability; see docs/process/agent-artifacts.md#harness-configuration.

Editorial Contract

• Interpret the central design trade-off and what Pharos evidence reveals; do not repeat classification badges or dashboard numbers.
• Pharos is the summary's only source: summaries carry no sources field. Every claim must be backed by the base coin, a sidecar, or live analytics shown on the page, each of which carries its own provenance. Evidence that exists only outside Pharos goes into the owning sidecar first (with its dated sources) or the claim is dropped; a summary is never the sole home of a fact.
• Prefer durable relative wording for market cap, APY, TVL, event counts, and stress bands. A retained static adoption/funding/holder/TVL/supply/market-cap fact needs a backing Pharos module, fact date, chain scope, and denominator.
• Never call addresses users/holders without a defined method, compare unlike financing and token-supply quantities, or invent certainty for missing evidence.
• Read shared/lib/methodology-versions/current-version.json for the current methodology and ReportCardGradeSchema in shared/types/report-card-grade.ts for grade vocabulary; shared/types/safety-score-v9-public.ts owns pillar keys. Missing reviewed evidence is NR, not safety.
• AI drafts set actual authoredBy, model, updatedAt, and factsAsOf; only a named reviewer may add reviewedBy/reviewedAt after approving the exact text.
• Live dashboard values the prose must cite (overall/pillar/peg grade, score, circulating USD) go in as registered claim tokens, never literals. Each token used ({{grade}}, {{score}}, {{supplyUsd}}) must occur exactly once in text, with one claimTokens entry carrying token, placeholder, an allowed source from AI_SUMMARY_CLAIM_REGISTRY (shared/types/editorial.ts), and a calendar-valid factsAsOf day. Multiple distinct tokens are allowed, but two grade claims cannot bind to different sources through the same {{grade}} token. shared/lib/__tests__/ai-summary-claims.test.ts validates the edited data file. The detail page resolves tokens from its live hero/report-card data; missing values or malformed/unregistered tokens render N/A, including the build-time static fallback. Only active coins carry tokens; pre-launch summaries render literally.

Queue And Workflow

The permanent QA producer reads four live endpoint families (report cards, stress signals, peg summary, stablecoins). Its detectors are heuristic and skip summaries without a current card; absence from the queue does not verify the prose. Findings never approve exact text or manufacture reviewer provenance.

1. Run npm run candidates:ai-summaries for queue work; prioritize high, then medium.
2. For each coin, inventory current prose claims against static metadata and current live analytics. Preserve sound text/title; correct stale facts and de-brittle volatile wording rather than rewriting for variety.
3. Draft the schema-valid entry and present it for review when approval is requested. Keep one main claim per sentence and close on the durable constraint.
4. After edits, run npm run typecheck and npx vitest run shared/lib/__tests__/ai-summary-claims.test.ts; the latter validates the actual edited data file, including claim tokens and calendar dates. Run the existing term-markup tests when changing markup behavior. For queue work, rerun npm run candidates:ai-summaries and confirm refreshed entries leave the selected severity bands.

Report changed IDs, backing Pharos evidence and dates, review state, remaining candidates, and validation.