No AGENTS.md edit; no citations added under content/ (verification section found none). The change moves inbox link capture off the old 'reachable URL ⇒ ai_verified' heuristic to a model-free durable source-request that explicitly does not assert verification — that strengthens §4's honest-verification-label rule and §3.8 (never invent a locator), as the inbox/README and MAINTENANCE updates state plainly. Promotion now routes through the already-existing bounded reader with a separate checking model, tighter caps (three sources → two, plus call/time/$ ceilings) and removal of the house-model refusal fallback; adoption is model-free, checks the exact basis fingerprint, validates a complete prospective case through the production loader, rolls back interrupted writes, records an append-only intake outcome referencing the originating proposal, and writes a history.yaml entry stamped with model, promptVersion and runId (§3.14, §3.15 provenance/inspectability). Nothing here publishes: prepared bundles still go to a needs-approval PR under the normal independent gate, and the docs repeatedly say 'prepared' is not ratification (§3.15 no unratified draft presented as ratified). The one genuine relaxation is that a title-similar source, which promote-core previously blocked as a probable duplicate, now yields a review warning retained in the adoption outcome rather than a hard stop; exact identifier matching still fails closed, the change is disclosed in AUTOMATION.md/DECISIONS.md, and independence remains an explicit review question, so I read this as a documented judgment call rather than a §3.10/§3.15 check-weakening. Governance promotion totals derive from adoption receipts and throw on an unknown proposal, so the ledger cannot silently invent promotions.

The change adds no published content or citation identifiers, and its synthetic fixtures explicitly disclaim being real publications. Inbox URLs are now queued without claiming verification, while proposals require retrieved text, a separate reading check, explicit evidence direction, narrow claims, inference separation, and independence notes before reaching the existing publication gate. Adoption checks the exact current case basis, validates the prospective complete case, appends a stamped history entry, preserves the originating proposal receipts, and rolls back interrupted ledger writes. The removal of the old promoter does not visibly weaken constitutional checks: exact identifier matching remains mechanical, title similarity becomes an explicit review warning rather than false identity, stale proposals are parked, and no AGENTS.md or panel-threshold change is made.

The change unifies inbox URL intake and legacy watch imports under the shared bounded source reader and research proposal envelope, eliminating the previous behavior where reachability alone marked imported URLs as verified. Proposed observations and evidence records continue to require mechanical verification, claim-anchoring, and prospective case validation before installation into working trees for the standard gated publication PR. Exact provenance, case history tracking, and rollback on failed materialization are preserved without bypassing the independent publication arbiter. AGENTS.md is not modified and no constitutional checks are weakened.

The diff unifies inbox capture and legacy watch imports onto the existing bounded source-reader path without editing AGENTS.md or weakening the independent publication gate. Inbox no longer equates URL reachability with ai_verified; it only records immutable source-request history plus submitted context, which is stricter honest verification labeling (§3.8 / phase-1). Reading remains budget-capped, separately checked, and non-publishing; prepareResearchAdoptions materializes only after exact-basis and prospective-case validation, with rollback on interrupted writes and ordinary needs-approval PR staging. No content/ citations or Deep Memory assessments are added (citation tooling found none); tests use synthetic fixtures and claim no research findings. Governance totals derive from adoption receipts rather than rewriting historical import files, preserving provenance and append-only intake history.

The change removes a genuine §4/§3.8 honesty defect — inbox capture previously stamped reachable URLs as `ai_verified` — and now records source requests that explicitly assert no verification, preserving the original contributor file verbatim and keeping submitted context as provenance. The old promotion path's mechanical verbatim-quote check is deleted, but the new format confines verbatim text to a short anchor checked by a separate source-reading pass, complete bundles are validated against the production loader with basis-hash staleness checks, interrupted writes are rolled back, and every addition still faces the ordinary independent publication gate (classifier, citation arbiter, needs-approval PR) — so this reads as a restructure under the gates, not a reclassification to dodge them, and the reading-pass budget actually tightens (3 sources → 2, $1 cap, no house-model fallback). Nothing publishes unilaterally: the materializer only prepares working-tree changes and its outcomes are labeled as non-ratifying, adoption history entries carry aiAssisted/model/promptVersion stamps, and §3.14 changelog entries are appended rather than history rewritten. No AGENTS.md edit, no new citation identifiers (verification section is empty), and all test fixtures are explicitly synthetic with no research finding claimed. The same-vendor reading check is candidly documented as not cross-vendor concurrence, which is consistent with the constitution's disclosure norms rather than a misrepresentation.