The diff adds only scripts, a workflow job, tests, and a DECISIONS.md log entry — no AGENTS.md edit and no new citations under content/ (verification tooling found no identifiers to resolve). The promotion pipe strengthens rather than weakens the checks: it dedupes against the ledger by identifier and title similarity (§3.10), mechanically rejects any evidence draft whose double-quoted spans are not found verbatim in the fetched text (§3.8), enum-checks direction against the supports/undermines/qualifies/context vocabulary (§3.5), requires existing claim anchors (no dangling IDs), separates sourceStatement from editorInference (§3.9), stamps records with reviewState 'ai_extracted', verification 'ai_verified', runId/date/origin (§4), writes an append-only dispositions ledger and case history entry (§3.14), refuses to promote a proposal with no surviving verified evidence, and cannot publish — output lands in a labeled needs-approval PR with a human reviewer, never the auto-tier maintenance PR (§3.15). The bench-core.mjs 4-of-5 advancement rule with a one-seat concern veto is new, uncalled library code governing research-proposal advancement (whose freeze PR still faces the arbiter), not a relaxation of any existing publication gate; multi-model concurrence is not defined as unanimity in the constitution. Two imperfections I note without treating as violations: the evidence `origin` stamp records runId/date/extractedBy but not the concrete model name or promptVersion that §3.15 asks for, and bench-core plus its 'founder-confirmed' comment are scope not disclosed in the PR body — both should be tightened, neither is a fabrication, provenance removal, or check weakening.

The change introduces a promotion pipeline that drafts Source and Evidence records but cannot publish them; they are only added on a separate needs-approval PR, preserving multi-model/human gating and append-only history (§3.14, §3.15). Evidence records are mechanically checked for valid directions and strengths enums, claim anchors existing in the case, and minimum-length sourceStatement/editorInference fields, supporting atomic, anchored claims and explicit evidence direction and separating source statement from inference (§3.2, §3.5, §3.6, §3.9). The `unverifiedQuotes` logic enforces that every quoted span in sourceStatement must be found (whitespace/hyphen-insensitive) in the fetched text, reinforcing exact though not paginated provenance without fabricating locators (§3.7, §3.8, §3.15). Deduplication via identifiers and title similarity is explicitly aimed at not double-counting the same work, aligning with independence handling (§3.10). The script refuses to promote a proposal if no evidence draft survives these checks and records all dispositions (promoted/duplicate/failed) in an auditable ledger, avoiding silent failures and retaining calibrated uncertainty and process transparency (§3.11, §3.13, §3.14). No citations, locators, or provenance details are fabricated in the code; identifiers are taken from proposals or model output and then mechanically constrained, and there is no change to AGENTS.md or weakening of constitutional checks. The automation description added to docs/DECISIONS.md accurately reflects the implemented safeguards and does not misrepresent the verification level or review status (§3.15).

The PR introduces an automated promotion pipeline that strictly adheres to the constitution's constraints. It operates entirely behind the established multi-model concurrence gate by opening a 'needs-approval' PR instead of publishing directly (§3.15). The pipeline guarantees exact provenance by mechanically verifying every quoted span against the fetched text (§3.8), fail-closing on any misquotes or bad identifiers. Furthermore, AI-generated records are honestly labeled as 'ai_verified' and 'ai_extracted' (§3.15), and the script distinguishes source statements from editor inferences (§3.9). No core checks are bypassed or weakened.

The diff does not edit AGENTS.md. It adds a gated promotion path that drafts source/evidence records from already-verified inbox proposals, mechanically fail-closes on quote-verbatim checks, claim-anchor/enum shape, dedupe (§3.10), and budgets, then opens a separate needs-approval PR rather than publishing or riding the auto-tier maintenance PR (§3.15). Drafted records are labeled ai_extracted / ai_verified with origin, runId, date, prompt version, and a history.yaml entry, and they distinguish sourceStatement from editorInference (§3.9, §4). No content/ citations are added in this change (verification tooling found none); DECISIONS.md only documents the pipe and does not weaken the publication gates. Ordinary gaps (identifier not registry-checked in-script; unused bench-core 4/5 helper still defers veto to the freeze/arbiter path) are residual tooling limits, not a constitutional break.

The pipe drafts records but cannot publish: the new `promote` job opens a needs-approval PR routed through the classifier, citation-checking arbiter, and founder reviewer, so the §3.15 gate is preserved rather than weakened, and nothing in the diff touches the existing fail-closed checks or AGENTS.md. Drafted records are honestly labeled (`reviewState: "ai_extracted"`, `verification: "ai_verified"`, origin stamped with runId/date/prompt version), and the verificationNote explicitly tells reviewing seats to judge the readings, not the existence — no unratified draft is presented as ratified. The §3.8 surface is guarded mechanically: every quoted span in a drafted sourceStatement must verify verbatim against the fetched text or the record is discarded and logged as failed in the auditable promotions ledger, and evidence direction enums match §3.5. No citation identifiers are added to content (verification section confirms none), so the real-citations rule is not engaged; the DOI/arXiv strings in test files are obvious fixtures exercising regex/dedupe logic, not content citations. The LLM-drafted `identifier` field lacks a mechanical in-text check, but any hallucinated locator would be caught by the downstream citation-checking arbiter before publication, so this is ordinary imperfection rather than a check-weakening. Live fetching and model calls live in CI maintenance scripts, not the site itself, consistent with §4. DECISIONS.md is appended, not AGENTS.md, and documenting a shipped material decision there fits the §9 workflow.