The founder-supplied home page (SRC-JHONG-2026) is admitted with a permission recorded to the §3.15 standard — grantor (Eugene Jhong, own work), date (2026-09-09), channel (inbox statement), and where the correspondence is held (inbox/processed/…) — so no confidence-material problem arises, and the only added URL (github.com/ejhong/birdmen) RESOLVES per the tooling. All 25 works the page names but that were not retrieved are dispositioned `blocked` with a route and nothing is quoted from them, which is exactly the honest handling §3.7/§3.8 asks for. Verification labels are conservative (`unverified` on the source, every evidence record `weak`), disputed second-reader judgments are preserved in each record's `limitations` rather than suppressed (E015, E017, E022, E023, E024), and the verify run's refusals and splits are reasoned and anchored in verification.md. The edition's article and assessment repeatedly say the ledger rests on one self-published, unaudited source, grade most featured claims `unresolved`, and mark prior-driven grades as such — the article does not assert more than the ledger holds. The panel ran with 5 of 5 seats and their verdicts (four `unresolved`, one `weakly_supported`) are all within one step of the draft, each displayed in its own words. Two minor imperfections worth a note but not violations: the source's verificationNote ("not retrievable when drafted (no URL)") is stale and sits oddly beside an identifier naming a commit and md5 for a page the run did retrieve as a supplied document; and DEEP-E022/E024 keep a `supports` direction toward claims their own second reader says they do not bear on, which is at least visible on the record.

Several admitted Evidence records retain inaccurate directions or claim links even though their own second-reader notes identify the mismatch. DEEP-E015 and DEEP-E017 are marked `supports` although their statements only qualify the scope of the numerical claims; DEEP-E022 is linked as support for DEEP-C037 even though it addresses only DEEP-C038; and DEEP-E024 supports DEEP-C015 despite stating only the Kurdish-name claim in DEEP-C016. Most clearly, DEEP-E023 treats a pointer to an unretrieved wear study as support for that study's alleged finding, contrary to the rule that a source or citation is not itself evidence. These records should be reclassified as `qualifies` or `context only`, or have the unsupported claim links removed; recording the dispute in `limitations` does not cure the structured evidence-direction error.

The intake of the Deep Memory homepage strictly adheres to epistemic standards and phase-1 rules. The source (SRC-JHONG-2026) is marked with an honest verification label (`unverified`), explicit provenance, and a verbatim record of supplier permission satisfying the §3.15 confidentiality rule. The verification pass thoroughly enforced atomic propositions and explicit directionality, rejecting compound statements and noting second-reader disputes directly on evidence records. The new edition for Deep Memory is calibrated, transparent about the ledger's evidentiary limits, and evaluated by an independent 5-seat panel whose concurrence and disagreements are faithfully preserved without check-weakening.

The deep-memory intake admits one founder-supplied source (SRC-JHONG-2026) with a recorded permission stamp (who/when/channel/held-at), honest `unverified` label, and weak self-report framing; the GitHub URL resolves. Verify rejected 31 non-atomic or overreaching candidates with reasons, split survivors, and attached second-reader direction disputes on several evidence records—consistent with §3.2, §3.5, §3.8–3.9. The edition and five check assessments are append-only, stamped, labeled `humanReviewed: false` / role draft|check, verdict unresolved (4 of 5 seats) with thin-ledger honesty, competing accounts, and explicit cruxes. Vasocomputation adds only panel checks. No AGENTS.md edit, no confidence leak, no fabricated locator visible in the material or citation-verification output.

The diff does not touch AGENTS.md, adds no unresolvable identifiers (the one added URL github.com/ejhong/birdmen and the three DOIs named in the blocked-source dispositions all RESOLVE in the verification section), and the intake pipeline's behavior is exactly what the constitution demands: 31 of 42 proposed records rejected with specific reasons (non-atomicity, anchor overreach, locator mismatch), compound claims and evidence split into atomic parts, second-reader disputes over direction recorded verbatim in the limitations of E015/E017/E022/E023/E024 rather than hidden. The founder-supplied page enters only with a recorded permission line (who, date, channel, holding location), is labeled `verification: unverified` and self-published, is graded weak self-report throughout, and nothing is quoted from the 25 unretrieved works, which are dispositioned `blocked` with retrieval routes. The article says only what the ledger holds — it repeatedly grades claims 'unresolved' or 'provisionally supported on priors' and flags that the decisive bundle null has no supporting record — and the edition and check files are stamped runId/model/promptVersion with `humanReviewed: false` and role labels, with five-seat panel verdicts and disagreement (Gemini's weakly_supported vs four unresolved seats) displayed rather than buried. The budget-capped report run refused with nothing sent, which is fail-closed behavior, not a dodge. Minor imperfections (a stale verificationNote saying 'not retrievable when drafted' that the same record supersedes, an edition rationale saying 48 claims where the ledger holds 47, a garbled mechanical disposition key for E021) are routine record-keeping noise, not violations of any named rule.