[ancsingabor]

The docs/guide/ pass is finished, so this is the layer-2 sweep. Every ADR was read against the code. Nothing is shortened — they stay the complete record, per CLAUDE.md. What changes is that each file now says on its Status: line whether it has moved since acceptance, and four claims that had gone false are corrected.

Real defects, found by checking each ADR against code

ADR	Claim	Reality
013	header: (implemented Milestone 1)	Not implemented. No app/[lang]/; proxy.ts has no locale logic at all; config/copy/index.ts still says "once app/[lang]/ routing lands". The config half (brand.defaultLocale, brand.locales, getCopy) did land
010	sources.source_kind is a required column	source_kind is the Postgres enum type; the column is sources.kind. And the tier is not required — 0005's constraint only forbids a tier on scientific/historical, and sourceSchema.authority_tier is .nullable().default(null)
005	gate check 3: "every claim-bearing sentence retains ≥1 citation"	ADR-018 establishes this is undecidable over free prose and replaces it. 005 had no pointer, so a reader of it alone takes the stronger guarantee as the one the code makes
002	"chunks references it"	Neither table references the other; chunk_units joins them, for the one case a units.chunk_id column could not model

ADR-010 is the one with teeth. The schema makes the category error that ADR exists to prevent unrepresentable in one direction only: ranking a physics paper by ecclesial authority cannot be expressed, but ingesting an encyclical with no tier at all can — it would pass the manifest schema, the database and every assertion in assert.ts, then reach context assembly as an unlabelled block. It has not bitten only because both sources are tiered by hand and there are two of them. The check is now owed in status.md; the honest place for it is the manifest schema, beside the resolved license ADR-003 already demands.

ADR-013's defect had a sibling in README.md, which said "Apologia serves /hu/… and /en/…". Fixed in the same pass.

Smaller corrections

• 004 listed embed among the stages that run — it is deliberately unwritten
• 015's integration-row cell still read "decided when the first one is written" although its own amendment decides it (Vitest)
• 023's "234 cases over ~1,550 lines" is now dated; it is 252 over ~3,450
• 009 said the honeypot field is retained. The helper is, and is tested — but no form renders it, because there is no question box yet

The alignment

Every ADR that has moved now carries its story on the Status: line, in the shape ADR-019 already used. docs/adr/README.md repeats it in the index's Status column and states the layering explicitly: the guide is the way in, these are the argument behind it, and where they disagree the ADR is right.

Thirteen of the twenty carry a marker. Seven have a dated ## Amendment or ## Open question; three were superseded by a later ADR (004, 005 by 023/018; 014 by 017); 011, 017 and 020 record a correction made elsewhere. The seven that stand exactly as accepted are 001, 006, 009, 016, 018, 021, 022 — largely the ones whose subject has not been built. In this repository a decision gets corrected when code meets it, so an unamended ADR is usually an untested one. The count ships with the one-liner that re-derives it.

ADR-023's two post-acceptance corrections (outcome 4, the restated bundle limit) were sitting at ### inside ## Trade-offs, which read as part of the original argument. Promoted to a dated ## Amendment.

One lead disproved

The honeypot was carried into this session as a field named by ADR-009 and ADR-015 and existing in no code. It exists — isHoneypotTripped() in lib/rate-limit.ts, with unit tests. ADR-015's row was correct all along; only ADR-009's wording needed the precision.

Checks

docs:lint green (20 ADRs, 302 links, 25 anchors) · npm test 350/350 · tsc --noEmit clean.

🤖 Generated with Claude Code