reviews: refutation battery — twelve external passes from six lineages, with fact-checked verdicts

[dhh1128]

The external adversarial review the battery's README asks for: independent reviewer instances, blind to one another, attacking the two dockets. Six seats, one flagship per lab, none Claude-lineage (the dockets and their source of record are Claude-co-authored, so same-lineage reviewers would share the author's blind spots): GPT 5.6, DeepSeek v4 Pro, GLM 5.2, Qwen 3.8 Max, Kimi K3, Mistral Large. Twelve blind passes, verbatim and sealed under MANIFEST.txt, plus the exact stdin packets for replay and a harness report whose fact-checks are each reproducible without trusting the harness.

Docket 1 — §V.6 SURVIVES. Four lineages NOT REFUTED after prosecuting the rule-1 upgrade route across the V.1 matrix and sweeping well beyond it (Microsoft CCF and Hyperledger Fabric are the strongest new near-misses; witness-cosigned transparency, Cardano Conway, seL4 proof-CI, and the vLEI ecosystem itself were all graded and rejected). The two REFUTED verdicts do not survive fact-checking: Mistral's DNSSEC exhibit quotes three DPS sections that do not exist in the live document (checked 2026-08-14 against the 8th-edition DPS — actual titles reported in the harness report), and DeepSeek's CT upgrade route fails Part 0's own "certifying its cure" clause, with Kimi's inversion closing the door: strict Part 0 grading moves CT's P2 down (the matrix graded issuance acts, not verification runs).

One seam is ruling-grade and owed to 4.3: Part 0's P4 definition does not state the organ requirement that §V.6's "P4 textual but organ-less" dismissal of CT leans on — four seats surfaced it independently. Also convergent, on residue item 7's two named attack surfaces: "died unfunded" over-asserts (drafts' expiration verifiable, funding cause not), and the QVI exhibit conflates issuer with watcher — GLEIF's witness pool proposed as the better existence proof.

Docket 2 — split verdict, sub-claim by sub-claim.

• Margin (i) falls, five genuine seats of five: KERI's own superseding-recovery doctrine is the prior art, verified against the spec of record (reconciliation rules A0/A1; disputed-branch retention).
• Margin (ii) survives on novelty — five seats, convergent near-miss lists (Sheng et al. is threshold-counting, not spectral; SybilGuard/SybilLimit is the nearest genuine conductance use; Chuat et al. quantifies gossip detection without expansion pricing). The one refutation attempt had to invent a "Theorem 4.1" to land, and was struck.
• Margin (iii) falls as stated: the corollary's own example class contains its counterexamples (outpoint double-spends, same-parent blocks, Schneier–Kelsey hash-chained logs), unless "coordinate" absorbs any consumed resource reference — in which case the lemma trends tautological. Four seats, both horns, independently.
• The detection theorem's "iff" falls twice: GPT's temporal-ordering counterexample (connected comparison graph, no time-respecting path carrying both occupants) and Kimi's supersession-laundering (equivocate briefly, then publish a lawful superseding rotation to everyone; nothing in the six clauses preserves or transmits the superseded evidence — full connectivity, permanent evasion). Both check by hand. The convergently prescribed repair imports KERI's own "first seen, always seen, never unseen" as an explicit evidence-retention law, plus evidence-complete comparison messages and a time-indexed detection-by-t.
• Completeness falls to the unstated predecessor-chaining premise (one-clause repair); clause 4 carries a one-line internal inconsistency ("distributed across observers" belongs to undetectedness, not the definition); soundness holds — the one attack against it misstated KERI's rule A1 and was struck.

Both struck verdicts are Mistral's; its fabrications were caught because five other lineages graded the same sources honestly — the battery's design working as intended.

Relates to #77 (the §V.6 graduation gate this battery discharges is queued there via the workspace's audit discipline).

🤖 Generated with Claude Code