This is pipeline/protocol plumbing: no content/ records are added or changed, and the citation tooling found no identifiers in added content lines, so there is no citation, quote, or locator to fabricate here. The new resolve.ts does not mint records — it pins a report's named-but-unopened lead to a document via open indexes (OpenAlex, Semantic Scholar, Europe PMC, Internet Archive), then fetches that document and hands the drafter the text along with `via` (which index answered, with the matched title and year) and a similarity score, so provenance of how the source was found is stamped rather than invented; unmatched leads stay blocked with the failed search on the record, which if anything strengthens §3.8 rather than loosening it. Acceptance reuses the existing inbox matcher, now moved verbatim to match.ts (title containment plus year within one, or author+year with a looser title bar), so no matching threshold is relaxed; the only behavioural loosening is the inbox's default searcher widening from OpenAlex alone to the multi-index search, which adds candidate sources without changing the acceptance bar. AGENTS.md is untouched; the DECISIONS.md entry records the founder's direction with date and channel ("direction in session, 2026-09-17"), which §3.15 expressly permits as operating record rather than confidence material. The report-v2 protocol keeps the prohibitions intact ("Never invent a citation, author, date, identifier, locator, quotation, or result"; "Do not write records"; report is intake, never citable) and the v1 file is archived, not deleted, preserving history. My one residual worry is that a mis-resolved lead could feed the drafter the wrong document, but the match is recorded with its `via` and similarity and still passes through draft and verify, so that is a routine quality risk, not a rule breach.

The new DECISIONS.md entry says a candidate is accepted only when “title, year and an author agree,” but the implementation does not enforce that procedure. In `bestMatch`, `byTitle` alone is sufficient, `yearOk` passes when either year is missing, and no author match is then required; thus an exact-title result with a missing year or conflicting author can be recorded as the resolved document. The draft subsequently labels that document as “resolved from the report's lead,” making the source-to-lead provenance stronger than the actual check warrants. This makes the public history inaccurate about how identity was checked and risks false provenance; require author agreement and a present, compatible year for structured leads, or revise the recorded criterion and label title-only matches as uncertain rather than resolved.

The change introduces protocol v2 and pipeline machinery (`src/pipeline/resolve.ts`, `src/pipeline/match.ts`) to systematically resolve works named in reports across open academic indexes (OpenAlex, Semantic Scholar, Europe PMC, Internet Archive) before drafting. This directly supports §3.7 and §3.8 by locating primary texts rather than leaving leads unpinned, while strictly tracking provenance through explicit `via` annotations and search outcomes. The shared matcher enforces strict title and author/year constraints to prevent spurious matches, and normal verification safeguards remain in place downstream. No constitutional files are modified and no checks are weakened.

The diff is pipeline/protocol work only: report-v2 adds an explicit Named-not-opened section, resolve/match pin those leads through open indexes under shared acceptance rules (title near-match or author+year with a looser title bar), and draft fetches what resolves with a via stamp while leaving unmatched leads blocked and on the record. No content/ records or citation identifiers are added, AGENTS.md is untouched, and DECISIONS.md records a founder session direction rather than inventing permission. Matching is stricter than free invention, tests cover parse/accept/reject paths including Archive, and nothing loosens §3.15 checks or verification labels.

The change adds a pipeline step that pins the report's named-but-unopened works to real documents via OpenAlex, Semantic Scholar, Europe PMC, and the Internet Archive, then fetches and shows the resolved text to the drafter like any other source — no citation, locator, or provenance detail is invented; a lead that matches nothing stays blocked with the failed search on the record, and every resolution carries a `via` stamp and similarity score (§3.8, §3.9). The matcher is the pre-existing one moved to a shared module unchanged (title rule or named author + year), the report protocol explicitly forbids inventing records and routes everything through draft and verify, and AGENTS.md is not touched — so no check is loosened or dodged. The resolutions are index results the tooling actually received, not fabricated identifiers, and the Citation Verification section confirms no citation identifiers were added to content. The decisions-log entry asserts a founder direction in session that I cannot independently verify, but it follows the log's established stamped pattern, no content publishes on it here, and nothing in the material contradicts it. Minor quality notes — a lenient year regex that could pull a year out of a title, and misattribution risk in fuzzy matching — are mitigated by the visible `via`/similarity labels and by verify still checking whatever the drafter proposes, and are ordinary imperfection, not violation.