The diff is pipeline/tooling only: a shared source-identity module, a read-only interface over existing watch/promotion/archive decision files, a model-free `intake-report.mjs` inspector, tests, and doc updates. No content/ records, citations, or assessments are added or altered (the verification section found no identifiers), and AGENTS.md is untouched. The checks are tightened rather than weakened: `alreadyCarried` still defers on title-similarity near-duplicates (§3.10), promotion still records duplicates/failures to the ledger, the case-input hash is re-checked before writing so decisions on stale inputs are discarded, and the existing per-run promotion budget and admission gates are preserved. Provenance is treated conservatively — missing historical model/inputHash/promptVersion stay `null`, ambiguous legacy promotions stay unscoped rather than being attributed to an invented case, and malformed memory throws instead of silently reading as empty (§3.8, §8 'no silent data repair'). The triage prompt explicitly labels prior decisions as untrusted context and warns that knowing a source does not establish that an observation was assessed, which supports §3.2/§3.6 rather than undermining them. Making previously-suppressed proposals reconsiderable (retryable operational failures, revised rationale, changed case inputs) widens review rather than dodging it, and the docs/DECISIONS.md entry honestly scopes what is not yet implemented.

The change adds no research content or citation identifiers under content/, and the mechanical citation report found nothing requiring existence verification. Source identity is kept separate from assessment: title similarity is advisory, decisions are scoped by case and input hashes, and known sources are explicitly not treated as proof that every observation was considered. Historical decisions remain in their original files with exact file references, missing metadata stays null, and inferred legacy case ownership is visibly labeled rather than presented as recorded provenance. The visible diff appends decision context, adds fail-closed parsing and concurrent-input checks, and does not remove the existing admission or multi-model publication gates.

The change establishes a shared source-identity and read-only intake memory interface without altering or weakening existing admission or multi-model review gates (§3.15). Exact provenance and historical receipts are preserved honestly, keeping missing legacy review hashes null rather than inventing metadata (§3.8, §3.14). Deduplication logic distinguishes exact identifiers from advisory title similarity and scopes promotion receipts per case, protecting independence (§3.10) while preventing cross-case suppression. AGENTS.md is untouched, and no content under content/ is added or modified.

The diff touches only automation docs, intake scripts, and tests; it does not edit AGENTS.md, research content under content/, or public UI. Shared source-identity and the read-only intake-memory layer keep original decision files authoritative, leave missing historical model/input receipts null rather than inventing them, and treat prior decisions as context rather than terminal verdicts or an “already considered” gate. Promotion resting is narrowed to same case + candidate hash + case input hash with retryable operational failures preserved, which strengthens rather than weakens the §3.15 checks. Citation tooling reported no new content/ identifiers; synthetic test DOIs/arXiv strings are labeled as syntax fixtures, not research citations.

The change is tooling and documentation only: it adds a shared source-identity module, a read-only intake-memory interface over existing decision files, and per-case promotion receipts, explicitly leaving the existing admission and publication gates in place and rewriting no historical records (missing model/input receipts are left null rather than invented, honoring §3.15's no-fabrication rule). The triage prompt correctly labels prior intake decisions as untrusted records rather than verdicts, and new archive entries gain model/promptVersion/inputHash stamps, strengthening rather than weakening the audit trail (§3.14). Test identifiers (10.1234/..., arXiv 2605.01190) are synthetic and labeled as such, no citations were added under content/, AGENTS.md is untouched, and no dependencies or build checks were weakened. Scoping promotion rest/retry to case+candidate+inputs (rather than a global handled-URL set) does not undermine §3.10, since near-duplicate sources still route through the existing identity-review gate and the code explicitly refuses to treat a known source as evidence that an observation was assessed.