When Grounded Witnesses Still Agree on a Falsehood: Correlated Semantic Failures, Verbatim Span Grounding, and Matched-Compute Multi-Model Quorums

When a 3-of-4 witness quorum already satisfies AIC-33 provenance completeness and GTP-34 grounded trust paths, when do live model witnesses still converge on a false audit verdict, and does Correlated-Inspection Sufficiency (CIS-35) prevent false consensus at matched compute?

Pairs an exact enumeration of 4,096 post-GTP-34 witness-panel configurations (E35-A) with a 100-evaluation live multi-model contract-audit benchmark across four distinct witness architecture families: Fable 5.1 (Max reasoning), Sol 5.6, Astra 3.7, and Velri 4.7 (E35-B). On shared retrieval-slice truncation traps (C05-C08), both ungated monoculture and ungated 4-family quorums suffer 4/4 unanimous false consensus (33.33% overall false certification rate), whereas Correlated-Inspection Sufficiency (CIS-35) reduces false certified releases to 0/12 (0.00%) and recovers 4/4 truncated-slice cases to 4-0 verified consensus once the canonical bundle is reopened. Twelve cases and one hundred calls bound the empirical claim; no production system changed.

• In E35-A, structural grounding alone (G0) falsely accepts all 4,088 compromised post-GTP-34 panels (100.00% FAR); family diversity alone (G1) or verbatim substring checking alone (G2) still falsely accept 2,040/4,088 (49.90%). The full CIS-35 gate (G5) achieves 8 TP, 0 FP, 4,088 TN, 0 FN. At marginal witness error p = 0.25, a 3-of-4 quorum with pairwise error correlation rho >= 0.35 fails at 12.05%, worse than a single p = 0.10 checker.
• In E35-B, on Stratum 2 shared retrieval-slice truncation traps (C05-C08), both the ungated homogeneous Fable 5.1 quorum and the ungated 4-family heterogeneous quorum produce 4/4 unanimous 4-0 false consensus (33.33% overall false certification rate). CIS-35 on the canonical bundle (Arm 5) reduces false certified releases to 0/12 (0.00%), certifies 10/12 (83.33%) correct releases, and holds 2/12 (16.67%) fail-closed on C03 and C12 section-label granularity.

• E35-A Post-GTP-34 correlated-inspection state-space diagnostic Executed (diagnostic)
• E35-B Live matched-compute multi-model contract-audit benchmark Executed (diagnostic)

• H35.1 On post-GTP-34 witness panels (|R_G(Q)| >= 3), releasing solely on 3-of-4 top-level verdict agreement (G0) produces a 100.00% false acceptance rate across the 4,088 semantically compromised configurations in E35-A.
• H35.2 Adding 4-family witness architecture diversity (Fable 5.1, Sol 5.6, Astra 3.7, Velri 4.7) without canonical context-completeness and override gating fails to prevent unanimous false consensus when witnesses share a truncated retrieval slice (C05-C08).
• H35.3 Checking only that each witness's cited quote is a verbatim substring of the document (G2) still falsely accepts 49.90% (2,040/4,088) of compromised panels because witnesses cite verbatim text from a superseded base clause rather than the governing override clause.
• H35.4 The conjunction of AIC-33/GTP-34 grounding, canonical bundle digest matching, at least 3 model families, verbatim governing-span verification, and override-chain reconciliation (G5) achieves 0.00% false acceptance and 0.00% false rejection across the 4,096-state diagnostic.
• H35.5 For marginal witness error p = 0.25, a 3-of-4 majority quorum exceeds the 10.00% error rate of a single p_single = 0.10 checker whenever pairwise failure correlation rho >= 0.2471 (and reaches 12.05% at rho = 0.35).
• H35.6 On the 12-case live multi-model benchmark (E35-B), CIS-35 reduces false certified releases from 4/12 (33.33%) under ungated homogeneous and ungated multi-family quorums to 0/12 (0.00%), catching 4/4 truncated-slice traps (C05-C08) fail-closed and recovering 4/4 correct certified releases once the canonical bundle is supplied.
• H35.7 Requiring strict governing-section label agreement and cross-reference completeness across four distinct model families introduces a nonzero conservative hold rate on clean/complete cases (2/12 = 16.67% in E35-B, specifically C03 and C12) without emitting any false certified release.

Research diagnostic and live multi-model benchmark only. Encevra does not train or fine-tune foundation models. No production system, release gate, or Noetis behavior changed.

• Can Round 36 replace the single governing_section field with a compositional clause-chain schema (CCS-36) that reduces the 2/12 conservative fail-closed holds on C03 and C12 to 0/12 while preserving 0.00% false certified releases on truncated and decoy-injected contracts?

Builds on Round 34, Round 33, Round 32, Round 31, Round 30, Round 29, Round 26, Round 10, Round 9.