<!-- DOC_STATUS: ACTIVE_NON_AUTHORITY -->

FIX7 Real-N6 — Duplicate Fail-Open Closure Report (2026-06-11)

• Host: T1 / Claude Code. Codex consulted: NO. Owner approval: NO. Production mutation: NO.
• Closes: Codex CODEX-N6-DUP-DECLARED, CODEX-N6-DUP-MANIFEST. Object IDs: TKT-OBJ-274 (this report), 281/282/283 (artifacts).
• Verifier: 277daf86…b67cf9cb (rev1) → 922e5246…80224385 (rev2). N6 value, certificate UNCHANGED.

1. Root cause (Codex-found)

Defect	Code locus (rev1)	Why it failed open
duplicate declared corpus member	verify_corpus_membership() → declared_set = set(declared)	the set silently absorbs a duplicated id; missing/extra stay empty; the per-doc loop re-hashes the same file → a candidate certificate is emitted
duplicate manifest record	parse_hash_manifest() → entries[rel] = h	a dict silently overwrites a duplicate path; shasum -c alone never catches duplicate records; verification proceeds and certifies an ambiguous corpus

2. The fix (rev2, fail-closed, no certificate on any)

verify_corpus_membership() — before the set() conversion:

_dup_seen = set()
for _m in declared:
    if _m in _dup_seen:
        raise N6Reject("N6_CORPUS_DUPLICATE", f"duplicate declared member {_m!r}")
    _dup_seen.add(_m)

parse_hash_manifest() — before dict insertion, per line:

if sep != "  " or not rel.strip():
    raise N6Reject("N6_MANIFEST_MALFORMED_LINE", ...)
rel = rel.strip()
if not re.fullmatch(r"[0-9a-f]{64}", h):
    raise N6Reject("N6_MANIFEST_BAD_HASH", ...)
if rel in entries:
    raise N6Reject("N6_MANIFEST_DUPLICATE", ...)

New reject codes: N6_CORPUS_DUPLICATE (declared), N6_MANIFEST_DUPLICATE, N6_MANIFEST_MALFORMED_LINE, N6_MANIFEST_BAD_HASH. Each raises before any value/compute → verify_real_n6 exits 5 with no certificate.

3. Positive proof (the legitimate path is unchanged)

• cmd python3 real_n6_provenance_verifier.py --selftest (cwd packet) → 19/19 PASS, exit 0 (was 14/14; +5 cases: clean-manifest-parses, dup-manifest-rejected, malformed-rejected, non-hex-rejected, dup-declared-rejected).
• recon manifest format safety: the real 32-entry recheck-9 HASH_MANIFEST.txt has 0 lines that the stricter parser would reject (all ^[0-9a-f]{64} .+).
• full ledger RECON=_recon bash commands.sh → REAL-N6 VERIFY: OK n6=d777e87c…, regenerated cert byte-identical to pinned acf8d259…, OVERALL: PASS (5/5).
• fresh-from-KB bash RERUN.sh → RERUN_RESULT: PASS.

4. Negative / adversarial proof (the exact Codex inputs now fail closed)

Replay harness codex_replay.py (Codex's exact inputs), cwd packet, evidence _recon:

Input	Unpatched 277daf86	Patched 922e5246
declared_members = CORPUS_MEMBERS + [CORPUS_MEMBERS[0]]	ACCEPTED cert_emitted=True (FAIL-OPEN)	REJECTED[N6_CORPUS_DUPLICATE] cert_emitted=False
HASH_MANIFEST with duplicate docs/05-…md record	ACCEPTED cert_emitted=True (FAIL-OPEN)	REJECTED[N6_MANIFEST_DUPLICATE] cert_emitted=False
HASH_MANIFEST malformed (single-space) line	ACCEPTED cert_emitted=True (FAIL-OPEN)	REJECTED[N6_MANIFEST_MALFORMED_LINE] cert_emitted=False

Permanent regression probes P25–P28 added to n6_adversarial_probes.py (0f2c94d3…): full suite 31/31 fail-closed, any PASS emitted: False | any seal/official digest emitted: False.

5. Verdict

Both Codex duplicate fail-open classes are CLOSED. No invalid duplicate input emits a certificate/digest/seal. N6 remains the same engineering candidate d777e87c…; it is NOT Codex-ratified. Standing authority blocker SEAL_REAL_N6_NOT_AVAILABLE is unchanged and authority-only.