On-Page Card Reuse SearchCopy Anchor Link

"Has this page already been cut?" — answered from one D1 table, fed by two sources, and enriched for corpus cards with a parsed citation and a cached LLM annotation.

• Routes: GET/POST /api/evidence-reuse-check, POST /api/evidence-reuse-check/annotate, GET /api/evidence-reuse-check/dashboard, POST /api/admin/debate-cards/reuse-index (backfill)
• Packages: debate-card-parser (src/extractors/card-record.ts), debate-research-evidence (src/lib/parquet-card-reuse.ts, src/lib/card-reuse-annotation.ts, src/lib/shared-evidence-library.ts)
• App: lib/admin/debate-card-import.ts, lib/evidence-reuse-check/card-annotation.ts
• Extension: apps/debate-browser-ext/src/reuse/api.ts, entrypoints/popup/App.tsx, src/components/ReuseMatchCard.tsx
• User docs: On-Page Card Reuse Search

The indexCopy Anchor Link

evidence_reuse_index holds one row per "this URL has been cut" fact, keyed by a caller-assigned id, with normalized_url indexed for the lookup. normalizeSourceUrl lowercases and drops the scheme, a leading www., the query/fragment and trailing slashes. The route keeps its own copy of that function, so the two must stay identical.

Two writers:

1. Library entries. POST /api/evidence-reuse-check registers one card's sourceUrl under whatever id the caller assigns.
2. The Parquet card corpus. Every batch writeDebateCardBatch writes to debate_cards also goes through buildReuseIndexStatements. Each card's cite/fullcite is parsed by debate-card-parser's parseCardRecord. A card whose citation names a URL is upserted as card:<id> (argBlock = the tag, topic = the caselist). A card that no longer names one has its old card:<id> row deleted, so a re-import that fixes a cite never leaves a stale match. These statements run in the same db.batch as the card upserts, so a batch lands whole or not at all. Inserts are capped at 14 rows per statement (7 columns under D1's 100-parameter ceiling).

Cards imported before corpus indexing existed are backfilled by POST /api/admin/debate-cards/reuse-index with body { afterId, limit }. It reads only the cite columns in id order, applies the same statements, and returns { processed, indexed, nextAfterId, done }. Call it in a loop, feeding nextAfterId back in, until done. It uses the same auth as the importer: an admin session, or CARD_IMPORT_TOKEN as a bearer token.

URL extractionCopy Anchor Link

extractSourceUrl strips tags, decodes entities, and takes the first http(s):// or bare www. URL in the citation. A bare www. URL gets https:// added. Trailing punctuation is removed with cleanUrl, the same cleanup htmlToCards applies. The URL comes from fullcite first and falls back to cite.

The lookupCopy Anchor Link

GET /api/evidence-reuse-check?url=&source= returns up to 20 matches and appends a row to reuse_check_log. That log feeds the team dashboard, and a weekly cron purges rows older than 180 days. For each match whose id parses as card:<id>, attachCorpusCards loads the debate_cards row and adds:

• card — buildReuseCardDetails: tag, cite, full cite, parsed author and year, up to 6 highlighted quotes (each <mark> run; falls back to spoken when the markup has none), caselist, event, level, side, and duplicateCount.
• annotation — the saved LLM annotation, when one exists.

This enrichment is best-effort: a failure is logged, and the "already cut" answer still returns.

The annotationCopy Anchor Link

POST /api/evidence-reuse-check/annotate takes { cardId }, never card text, so the model only ever sees cards that are in the library.

• The content sent is buildCardReuseAnnotationContent: tag, full citation, the highlighted text, then the full card text, capped at MAX_ANALYSIS_CONTENT_CHARS (20,000).
• The model is claude-opus-5 at effort medium, held to CARD_REUSE_ANNOTATION_SCHEMA through output_config.format. The request sends fallbacks: "default" (beta server-side-fallback-2026-07-01), so a classifier-declined request is re-run on the recommended model. The stored model column records which model actually answered.
• A refusal or max_tokens stop, or an answer that fails parseCardReuseAnnotation, returns an error and saves nothing.
• Results are cached in card_ai_analyses beside the evidence search's analyses. The key is the SHA-256 of the content plus the SHA-256 of CARD_REUSE_ANNOTATION_VERSION and the prompt. Bump the version whenever the prompt or schema changes shape, so old answers regenerate instead of being served in the old shape.
• A saved annotation is served to anyone. Generating a new one requires a session: the extension sends its bearer token, and a signed-out caller gets a 401 with a sign-in message.

Shape: claim, supportScore (0–10, clamped), authorQuality (rating is one of strong, adequate, weak or unknown, plus qualifications and concerns), and flaws (each with flaw, severity of high, medium or low, and explanation).

The extensionCopy Anchor Link

The popup calls the lookup for the active tab. It renders each match with ReuseMatchCard: the tag, an author/year/caselist line, the first 3 quotes, and the annotation or a Find flaws & rate author button. When the reader is signed in, the first 3 unannotated corpus matches are annotated automatically (AUTO_ANNOTATE_LIMIT). A page cut dozens of times must not start dozens of model calls on its own.

The extension builds standalone with bun run build in its directory. wxt.config.ts aliases debate-api-client to its TypeScript source, and tsconfig.json has the matching path, because that package's dist/ only exists after its own build.

Known gapsCopy Anchor Link

• Cards whose citation has no URL, or only a DOI or a publication name, are never indexed. They cannot be matched to a page.
• Matching is exact on the normalized URL. The same article at a different path (AMP pages, syndicated copies, /amp, print views) does not match.
• The backfill is a manual loop against the admin route. No cron runs it.
• The web app's "Check this page" box (ReuseCheckScreen in debate-webview) does not show card or annotation yet. Only the extension popup renders them.
• Annotation cost is bounded only by the session requirement and the once-per-card cache. There is no per-user rate limit.