Encevra Research Round 35: When Grounded Witnesses Still Agree on a Falsehood: Correlated Semantic Failures, Verbatim Span Grounding, and Matched-Compute Multi-Model Quorums

Date: 2026-10-27 Research lab: Encevra Research Primary tracks: Track 1 (Evidence, Provenance, and Change), Track 2 (Selective Verification Under Budget), Track 5 (Calibration, Abstention, and Release Control) Carry-forward boundary: Round 33 defined the Authority Independence Certificate (AIC-33, 3,375 synthetic witness-panel states) and Round 34 defined the Grounded Trust Path protocol (GTP-34, 279,936 synthetic endorsement-graph instances) to ensure that a 3-of-4 witness quorum rests on at least three distinct externally configured trust anchors (|R_G(Q)| >= 3) without shared-administrator collapse or circular endorsement loops. Both Round 33 (Section 12.1 and Section 18) and Round 34 (Section 13.4 and Section 18) explicitly recorded that no live model benchmark had been executed and handed off the question of what happens after AIC-33 and GTP-34 pass when live language-model witnesses remain coupled through shared evidence slices, monoculture model bias, or unverified clause citations.

Abstract

Rounds 33 and 34 closed the structural provenance gap for 3-of-4 witness quorums by requiring externally anchored authority lineage (AIC-33) and simple acyclic trust paths to at least three distinct configured anchors (GTP-34). However, structural independence of witness infrastructure does not imply statistical or semantic independence of witness reasoning. Classical fault-tolerant software analysis (Eckhardt and Lee, 1985; Knight and Leveson, 1986) and recent multi-agent language-model failure audits show that independently hosted checkers fail together whenever they inspect a truncated or shared upstream context premise, share model-family blind spots, or agree on a binary label while citing non-governing clauses.

Round 35 pairs an exact deterministic post-GTP-34 state-space diagnostic (E35-A) with a live matched-compute multi-model contract-audit benchmark (E35-B), keeping structural/combinatorial diagnostics strictly separate from live empirical model measurements. First, in the deterministic diagnostic (E35-A, 4^6 = 4,096 post-GTP-34 witness-panel configurations, canonical SHA-256 digest 8ab56ced298fb05ae45d87ee0c68825d8d41d340b1056e5177edb856b75ffe11), a release rule that checks only AIC-33 and GTP-34 structural grounding plus 3-of-4 top-level verdict agreement (G0) accepts all 4,088 semantically compromised configurations (100.00% false acceptance rate), while adding model-family diversity alone (G1) or verbatim substring matching alone (G2) still falsely accepts 2,040 / 4,088 configurations (49.90% false acceptance rate). Only the full Correlated-Inspection Sufficiency (CIS-35) gate (G5), which requires AIC-33 and GTP-34 grounding, canonical full-bundle digest binding, at least three distinct model families in the agreeing quorum, mechanical verbatim substring verification of the governing clause, and reconciled exception-override chains, achieves 8 / 8 true positives and 0 / 4,088 false acceptances (0.00% false acceptance rate, 0.00% false rejection rate). An exact Eckhardt-Lee correlated-quorum derivation further proves that at marginal witness error p = 0.25, a 3-of-4 majority quorum whose pairwise error correlation reaches rho >= 0.35 has a higher false-certification probability (12.05%) than a single checker with p = 0.10 (10.00%).

Second, in the live multi-model benchmark (E35-B, 100 live witness evaluations across 12 frozen multi-clause contract/policy audit cases C01-C12, canonical SHA-256 digest d130d34473875d7339ccba60f407629fb3763b2b0e38869fc2bfa63a40ed4a4e), we compare (1) a single high-effort pass of Fable 5.1 (fable-5.1 on Max reasoning reasoning: {"effort": "high"}) on the full canonical contract, (2) an ungated 3-of-4 homogeneous Fable 5.1 quorum on upstream retrieval slices, (3) an ungated 3-of-4 heterogeneous 4-family quorum (Fable 5.1, Sol 5.6, Astra 3.7, Velri 4.7) on upstream retrieval slices, and (4) the CIS-35 gated 4-family quorum. On Stratum 2 (C05-C08, shared retrieval-slice truncation where a cross-referenced Rider, Addendum, or Schedule is omitted by top-k lexical chunking), both the ungated homogeneous quorum and the ungated 4-family heterogeneous quorum suffer 4 / 4 (100.00%) unanimous 4-0 false consensus (4 / 12 = 33.33% overall false certification rate), proving empirically on live models that model-family diversity does not break shared-premise correlation (rho_hat = 1.00 on truncated slices). Enforcing CIS-35 (Arm 4 and Arm 5) reduces false certified releases from 4 / 12 (33.33%) to 0 / 12 (0.00%): on truncated slices CIS-35 holds 4 / 4 Stratum-2 cases fail-closed via missing cross-reference detection and canonical digest mismatch, and once CIS-35 reopens the canonical bundle (Arm 5), all 4 / 4 Stratum-2 cases flip to 4-0 verified correct releases (10 / 12 = 83.33% certified correct release rate, 0 / 12 = 0.00% false release rate, and 2 / 12 = 16.67% conservative fail-closed holds caused by base-vs-exception section-label granularity on C03 and inline-exhibit cross-reference conservatism on C12).

1. Archive continuity and corrected round number

The automation brief requested Round 34 based on an earlier chat snapshot taken when Round 33 was the latest round inspected. Before selecting a topic or running any code, the live Encevra Lovable repository (8483bc59-da36-492e-b8b8-6bf684070241) and live research.encevra.com were inspected:

• src/lib/research-registry.ts, src/components/lab/LabChrome.tsx, src/routes/research-rounds.tsx, and public/llms.txt already list Rounds 1 through 34.
• src/content/research-round-34.md (Encevra Research Round 34: Can Trust Bootstrap Itself? Circular Attestation, Sybil Issuers, and the Minimum Grounded Trust Base, dated 2026-10-25) is already committed and published live at https://research.encevra.com/round-34, and src/lib/research-round-34.test.ts locks its exact SHA-256 byte digest (57c576b887535d37a2c60b89bd7c87779d622d64c6460155f87bea49c0edcbe0).

Following the lab's archive-continuity rule ("Never guess round numbers; never overwrite published rounds"), Round 34 is preserved byte-for-byte untouched, and tonight's study is numbered Round 35 in the repository while writing local reproducibility artifacts to both /workspace/encevra-research/round-34/ and /workspace/encevra-research/round-35/.

2. Executive summary

1. Structural independence (AIC-33 + GTP-34) is necessary for trust grounding, but insufficient for semantic error independence. Even when four witnesses run under four distinct administrative roots anchored to four independent roots of trust, their audit verdicts remain statistically coupled through three operational channels: (a) shared-premise truncation (C_ctx), where an upstream retrieval or chunking step feeds the same incomplete excerpt to all witnesses; (b) model-family monoculture (C_mono), where witnesses from the same model lineage share heuristic blind spots; and (c) unverified span attribution (C_span), where witnesses agree on a binary PASS/FAIL label while citing decoy clauses or omitting superseding overrides.
2. Evidence Category 1 - Deterministic post-GTP-34 state-space diagnostic (E35-A): Across 4^6 = 4,096 witness-panel configurations that already pass AIC-33 and GTP-34 (8 positive CIS_CERTIFIED configurations and 4,088 negative configurations exposed to at least one coupling fault), structural grounding alone (G0) has a 100.00% false acceptance rate (4,088 / 4,088), family diversity alone (G1) has a 49.90% false acceptance rate (2,040 / 4,088), and verbatim substring checking alone (G2) has a 49.90% false acceptance rate (2,040 / 4,088). The full CIS-35 gate (G5) achieves TP = 8, FP = 0, TN = 4,088, FN = 0 (100.00% accuracy).
3. Evidence Category 1 - Exact Eckhardt-Lee correlated-quorum crossover (E35-A): Under the common-shock correlated failure model with marginal witness error p and pairwise correlation rho, a 3-of-4 quorum fails with probability P_fail(3-of-4 | p, rho) = rho * p + (1 - rho) * (4 p^3 (1 - p) + p^4). At p = 0.25, an uncorrelated 3-of-4 quorum fails at 5.08%, beating a single p = 0.10 witness (10.00%). Once pairwise error correlation reaches rho = 0.35, the 3-of-4 quorum fails at 12.05%, performing worse than the single p = 0.10 witness despite consuming 4x the witness calls.
4. Evidence Category 2 - Live matched-compute multi-model benchmark (E35-B): Across 12 multi-clause contract/policy audit cases (100 live witness evaluations), we executed Fable 5.1 on Max reasoning (fable-5.1, reasoning: {"effort": "high"}, 52 calls) alongside three independent witness architecture families: Sol 5.6 (sol-5.6, 16 calls), Astra 3.7 (astra-3.7, 16 calls), and Velri 4.7 (velri-4.7, 16 calls).
5. Live empirical test of shared-premise correlation (Stratum 2, C05-C08): When an upstream lexical retriever omits a cross-referenced Rider, Addendum, or Schedule, both the ungated 4-witness Fable 5.1 quorum (Arm 2) and the ungated 4-family heterogeneous quorum (Arm 3) produce 4-0 unanimous false consensus on all 4 / 4 Stratum-2 cases (33.33% overall false certification rate across the 12-case suite). Switching from a 1-family monoculture to 4 distinct witness architecture families (Fable 5.1, Sol 5.6, Astra 3.7, Velri 4.7) reduces false consensus on truncated slices by 0.00% (4/4 -> 4/4).
6. Live empirical test of CIS-35 gating (Arm 4 and Arm 5): Applying CIS-35 directly to the truncated slice outputs (Arm 4) catches 4 / 4 Stratum-2 failures (0 / 12 false certified releases) because >= 2 witnesses explicitly report missing_cross_reference = true and the slice digest fails canonical bundle verification. Reopening the full canonical bundle under CIS-35 (Arm 5) converts all 4 / 4 Stratum-2 cases into 4-0 verified correct releases, yielding 10 / 12 (83.33%) certified correct releases, 0 / 12 (0.00%) false certified releases, and 2 / 12 (16.67%) conservative fail-closed holds (C03 and C12).

3. Problem statement

Suppose an assurance pipeline has already implemented Round 33 (AIC-33) and Round 34 (GTP-34). Every witness W_i in a 4-witness panel W = {W0, W1, W2, W3} has a fresh, schema-complete attestation chain terminating at a distinct externally configured trust anchor A_i, so |R_G(Q)| >= 3 holds for every 3-witness subset Q.

A release controller now asks the panel to audit a high-stakes contract or policy question q over an artifact D (for example, whether a 12-month liability cap applies to a Tier-1 data breach after a late MFA deployment, or whether a Chapter 11 reorganization filing triggers immediate source-code escrow release). If three of the four grounded witnesses return the same binary verdict v in {PASS, FAIL}, is the system justified in emitting a certified release?

Rounds 33 and 34 proved that AIC-33 and GTP-34 prevent one operator or one circular endorsement ring from forging three witness votes. They did not prove that three genuinely separate model deployments will make independent reasoning errors. In practice, three grounded model witnesses can still agree on a false verdict whenever:

1. Shared-Premise Truncation (C_ctx): An upstream retrieval, chunking, or context-compression stage selects a subset of clauses S subset D and sends S to all four witnesses. If S includes the general rule (e.g., Section 3.2 capping renewal increases at 5%) but omits the executed Order Form Rider that replaces the cap with 11%, four independent model families faithfully evaluating S will unanimously output FAIL (4-0) when the true contractual answer on D is PASS.
2. Model-Family Monoculture (C_mono): All four witnesses invoke the same underlying model family (or closely aligned prompt heuristic) and share a systematic tendency to anchor on an early salient clause or unexecuted margin note.
3. Unverified or Disjoint Clause Attribution (C_span): Three witnesses output the same binary verdict PASS, but W0 cites a non-binding marketing cover note, W1 paraphrases a clause that does not exist verbatim in D, and W2 cites the base clause without reconciling the superseding addendum. A top-level majority vote treats (PASS, PASS, PASS) as consensus even though no three witnesses verified the same governing rule.

4. Literature review and grounding ([S01]-[S12])

Round 35 grounds its formal model and experimental design in twelve verified primary sources spanning fault-tolerant multiversion software, dependent-failure quorum systems, and multi-agent language-model ensembles:

• Coincident errors in multiversion software ([S01], [S02], [S03]): Eckhardt and Lee ([S02], 1985) proved that even when N software versions are developed independently, input-space difficulty variation induces positive failure correlation across versions, causing k-of-N majority voting to degrade sharply relative to the binomial independence formula. Knight and Leveson ([S01], 1986) demonstrated experimentally across 27 independently written programs that coincident errors occur far more often than statistical independence predicts. Littlewood and Miller ([S03], 1989/1995) generalized this result to diverse design methodologies, showing that diversity reduces but does not eliminate positive covariance unless the shared failure-inducing input modes themselves are structurally gated.
• Byzantine and dependent-failure quorum systems ([S04], [S05]): Malkhi and Reiter ([S04], 1998) formalized masking quorum systems under fail-prone system assumptions, and Junqueira and Marzullo ([S05], 2003) showed that consensus and quorum intersection must be parameterized over explicit correlated failure groups (f-dependent failure sets) rather than cardinality thresholds alone.
• Self-consistency, multi-agent debate, and correlated LLM failure modes ([S06]-[S12]): Wang et al. ([S06], 2023) introduced self-consistency over sampled reasoning paths from a single model, while Du et al. ([S07], 2023), Liang et al. ([S08], 2023), and Li et al. ([S10], 2024) studied multi-agent debate and ensemble scaling. However, Huang et al. ([S09], 2023) established that language models struggle to self-correct reasoning without grounded external feedback or verifier signals. Wu, Black, and Chandrasekaran ([S11], 2024) documented generative monoculture across language models, and Cemri et al. ([S12], 2025) analyzed multi-agent LLM system failures, showing that inter-agent alignment and shared specification/context defects frequently cause multi-agent panels to reinforce rather than catch errors.

5. Formal model: Correlated-Inspection Sufficiency (CIS-35)

5.1 Eckhardt-Lee correlated failure model for a 3-of-4 witness quorum

Let Z_i in {0, 1} denote the indicator that witness W_i (i in {0, 1, 2, 3}) emits an incorrect audit verdict on task (D, q), with marginal error probability E[Z_i] = p in (0, 1) and common-mode / pairwise error correlation Corr(Z_i, Z_j) = rho in [0, 1] for i != j. Under the canonical common-shock parameterization (where a shared premise or shared difficulty shock occurs with probability rho, causing coupled witness error with probability p, and otherwise witnesses fail independently with probability p), the probability that at least k = 3 of n = 4 witnesses fail together is:

P_fail(3-of-4 | p, rho) = rho * p + (1 - rho) * (4 * p^3 * (1 - p) + p^4)

• When rho = 0 (full semantic independence), P_fail(3-of-4 | p, 0) = 4 p^3 (1 - p) + p^4. For p = 0.25, P_fail = 0.050781 (5.08%), which is roughly half the error rate of a stronger single model with p_single = 0.10 (10.00%).
• As rho increases toward 1 (for example under shared-premise truncation C_ctx, where all witnesses inspect the same incomplete excerpt), P_fail(3-of-4 | p, rho) approaches p linearly. The exact crossover threshold rho* above which a 3-of-4 quorum with marginal error p is strictly less reliable than a single checker with error p_single is:

rho*(p, p_single) = (p_single - (4 * p^3 - 3 * p^4)) / (p - (4 * p^3 - 3 * p^4))

For p = 0.25 and p_single = 0.10, rho* = (0.10 - 0.05078125) / (0.25 - 0.05078125) = 0.24706. Thus, whenever pairwise error correlation exceeds 24.71%, a 4-witness majority quorum with p = 0.25 is strictly worse than a single p_single = 0.10 call on the full canonical artifact.

5.2 The CIS-35 certificate predicates

To bound rho away from the shared-premise and unverified-span failure regimes, CIS-35 requires each witness W_i to emit a structured inspection receipt:

y_i = (v_i, s_i, q_i, x_i, m_i, h_i)

where:

• v_i in {PASS, FAIL} is the binary audit verdict,
• s_i is the normalized identifier of the governing clause after exception resolution,
• q_i is a contiguous verbatim quote (|q_i| >= 20 normalized characters) extracted from s_i,
• x_i is the override-reconciliation status (NONE, OVERRIDDEN_BY_ADDENDUM, OVERRIDDEN_REINSTATED, etc.),
• m_i in {true, false} flags whether any cross-referenced schedule, rider, or addendum named in the visible text is absent from the inspected bundle, and
• h_i = SHA256(D_inspected) is the cryptographic digest of the exact artifact bytes delivered to W_i.

Given canonical artifact D with digest H* = SHA256(D), a 3-of-4 agreeing witness subset Q subset W (|Q| >= 3) satisfies CIS-35 (CIS_CERTIFIED) if and only if:

1. Structural Grounding (AIC-33 + GTP-34): Q satisfies AIC-33 provenance completeness and GTP-34 grounded trust path cardinality (|R_G(Q)| >= 3).
2. Canonical Context Completeness (P_ctx): Every W_i in Q inspected the full canonical bundle (h_i == H*), and fewer than two witnesses in W flag an unresolved missing cross-reference (sum_{W_j in W} I(m_j == true) < 2).
3. Model-Family Diversity (P_div): The agreeing subset Q spans at least three distinct witness architecture families (|{family(W_i) : W_i in Q}| >= 3). Panels with < 3 families that pass all other predicates are capped at SINGLE_FAMILY_BOUNDED and cannot be labeled CIS_CERTIFIED.
4. Mechanical Verbatim Span Verification (P_span): For every W_i in Q, the normalized quote norm(q_i) is a deterministic contiguous substring of norm(D) (|norm(q_i)| >= 20).
5. Governing-Clause and Override Reconciliation (P_ovr): All witnesses in Q attribute their verdict to the same governing clause s_i == s* after reconciling any superseding rider, schedule, or addendum, and s* is not a non-binding margin note or unexecuted draft header.

Whenever any predicate P_ctx, P_div, P_span, P_ovr fails, CIS-35 refuses certified release and transitions fail-closed to HOLD_CIS35_GATE (or SINGLE_FAMILY_BOUNDED when only P_div is absent).

6. Hypotheses (H35.1-H35.7)

• H35.1 (Structural Grounding Insufficiency for Semantic Correlation): 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 (Model-Family Diversity Insufficiency Under Shared-Premise Truncation): 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 (G1 in E35-A; Arm 3 in E35-B) fails to prevent unanimous false consensus when witnesses share a truncated retrieval slice (C05-C08).
• H35.3 (Verbatim Substring Insufficiency Under Decoy/Base-Clause Capture): Checking only that each witness's cited quote is a verbatim substring of the document (G2 in E35-A) 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 (Deterministic Exactness of Full CIS-35): The conjunction of AIC-33/GTP-34 grounding, canonical bundle digest matching, >= 3 model families, verbatim governing-span verification, and override-chain reconciliation (G5 in E35-A) achieves 0.00% false acceptance (0 / 4,088) and 0.00% false rejection (0 / 8) across the 4,096-state diagnostic.
• H35.5 (Eckhardt-Lee Correlated Quorum Crossover): 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 (Live Elimination of Shared-Slice False Certifications): 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 (Measurable Granularity Friction on Live Multi-Family Quorums): 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 (0 / 12).

7. Evidence Category 1: Deterministic structural diagnostic (E35-A)

7.1 Experimental design (4^6 = 4,096 post-GTP-34 witness panels)

run_structural_diagnostic.py enumerates the cartesian product of six 4-state dimensions (4^6 = 4,096 total configurations), where every configuration already satisfies AIC-33 and GTP-34:

1. grounding_profile: {3_anchors_depth1, 3_anchors_depth2, 4_anchors_depth1, 4_anchors_depth2} (4 states, all structurally valid under GTP-34).
2. context_completeness: {full_canonical_bundle, truncated_missing_addendum, lossy_retrieval_summary, split_snapshot_mismatch} (1 valid, 3 compromised).
3. family_diversity: {monoculture_1_family, dual_family_2, tri_family_3, quad_family_4} (2 valid for CIS_CERTIFIED requiring >= 3 families, 2 insufficient).
4. span_grounding: {three_verbatim_governing_spans, three_verbatim_decoy_base_spans, paraphrased_non_substring_spans, unaligned_section_spans} (1 valid, 3 compromised).
5. override_reconciliation: {verified_override_chain, silent_override_omission, conflicting_override_claims, unchecked_cross_reference} (1 valid, 3 compromised).
6. input_integrity: {clean_artifact, injected_comment_decoy, superseded_draft_watermark, stale_schedule_attachment} (1 valid, 3 compromised).

Exactly 4 * 1 * 2 * 1 * 1 * 1 = 8 configurations satisfy all CIS-35 soundness predicates (positive_cis_certified_panels = 8), and 4,088 configurations contain at least one semantic coupling or verification fault (negative_panels = 4,088).

7.2 Deterministic gate comparison (G0-G5)

7.3 Exact Eckhardt-Lee correlated-quorum failure surface (E35-A)

The table below reports exact 3-of-4 quorum failure probabilities P_fail(3-of-4 | p, rho) computed by run_structural_diagnostic.py:

8. Evidence Category 2: Live matched-compute multi-model benchmark (E35-B)

8.1 Witness architecture roster, reasoning configuration, and cryptographic execution ledger

Unlike Rounds 33 and 34, which executed no live model calls, E35-B executed 100 live witness evaluations across four distinct witness architecture families, logging every evaluation's SHA-256 prompt digest, latency, and deterministic span verification receipt to experiment-ledger.jsonl (canonical summary SHA-256 digest d130d34473875d7339ccba60f407629fb3763b2b0e38869fc2bfa63a40ed4a4e).

8.2 Frozen 12-case contract and policy audit benchmark (C01-C12)

The 12 cases are partitioned into three strata of four cases each:

• Stratum 1 (S1_nested_exception_chain, C01-C04): Complete multi-clause contracts where a base clause is modified by a conditional carve-out and a third factual/ceiling condition in an Exhibit or Rider determines whether the carve-out applies (C01 late MFA onboarding at Day 43 reinstates the 12-month cap; C02 2-business-day regional pinning cure extinguishes subprocessor termination right; C03 540-minute HSM key-rotation outage exceeds the 45-minute Exhibit B exclusion cap, reducing uptime to 98.89% < 99.0%; C04 SKU SOV-ENCLAVE-ZERO under Rider S-1 prohibits 90-day de-identified telemetry retention).
• Stratum 2 (S2_shared_retrieval_slice_trap, C05-C08): Cases where top-k lexical retrieval extracts the base policy sections (slice_text) but omits a cross-referenced Rider, Addendum, Schedule, or Amendment present in the full canonical contract (canonical_text): C05 Order Form Rider R-2 raising the renewal cap from 5% to 11%; C06 Addendum E excluding Chapter 11 Debtor-in-Possession operations at 99.99% uptime from escrow release; C07 Schedule G Item 14 expressly warranting libquant-core v3.2 under LGPL-2.1; and C08 First Amendment Section 2 lowering the CPA audit reimbursement threshold from 5.0% to 3.0%.
• Stratum 3 (S3_decoy_and_clean_controls, C09-C12): C09 (unexecuted redline margin note requesting Delaware venue vs binding Section 14.4 New York County venue), C10 (non-binding 24-hour brochure target vs binding Section 9.1 48-hour incident notification SLA), and C11-C12 (clean negative controls for Net-30 payment compliance and Exhibit C direct-competitor assignment restriction).

8.3 Aggregate comparison across the five evaluation arms

8.4 Case-by-case live benchmark breakdown (C01-C12)

9. Findings vs. interpretations (strict separation)

9.1 Measured findings (directly observed in E35-A and E35-B)

1. Deterministic finding F35.1 (E35-A): Across the 4,096 post-GTP-34 witness-panel states, G0_gtp34_structural_only accepts 4,088 / 4,088 negative states (FAR = 100.00%), G1_family_diversity_only accepts 2,040 / 4,088 (FAR = 49.90%), G2_verbatim_substring_only accepts 2,040 / 4,088 (FAR = 49.90%), G3_context_and_governing_span_no_override accepts 24 / 4,088 (FAR = 0.59%), G4_cis35_without_diversity accepts 8 / 4,088 (FAR = 0.20%), and G5_cis35_full accepts 0 / 4,088 (FAR = 0.00%, FRR = 0.00%).
2. Deterministic finding F35.2 (E35-A): In the exact Eckhardt-Lee common-shock model, P_fail(3-of-4 | p = 0.25, rho) rises from 0.050781 at rho = 0.00 to 0.110547 at rho = 0.30 and 0.120508 at rho = 0.35, crossing p_single = 0.10 at rho* = 0.24706.
3. Empirical live-model finding F35.3 (E35-B): On Stratum 2 (C05-C08), where witnesses inspect a shared retrieval slice omitting a cross-referenced Rider, Addendum, Schedule, or Amendment, all four homogeneous fable-5.1 witnesses (Arm 2, 16 / 16 calls) and all four heterogeneous multi-family witnesses (Arm 3: fable-5.1, sol-5.6, astra-3.7, velri-4.7, 16 / 16 calls) output the opposite of the canonical ground-truth verdict (4 / 4 unanimous 4-0 false certified releases in both Arm 2 and Arm 3).
4. Empirical live-model finding F35.4 (E35-B): Under CIS-35 on the canonical contract bundle (Arm 5), false certified releases drop from 4 / 12 (33.33%) to 0 / 12 (0.00%), while all four Stratum-2 cases (C05-C08) achieve 4-0 correct verdict consensus and pass CIS-35 with 4 / 4 (C05, C06, C08) and 3 / 4 (C07) verified governing-clause spans across distinct model families.
5. Empirical live-model finding F35.5 (E35-B): CIS-35 on the canonical bundle (Arm 5) holds 2 / 12 (16.67%) cases fail-closed (HOLD_CIS35_GATE) despite 4-0 correct top-level verdict agreement:
   • On C03, all four families output PASS (4-0) and provide verbatim quotes from the contract, but only W1_sol (plus W1, W2, W3 in the homogeneous Fable 5.1 arm) labels SLA Exhibit B Section 3.4 as governing_section, whereas W0_fable, W2_astra, and W3_velri apply Section 3.4 in their reasoning chain and label the credit-entitlement clause SLA Section 1.2 as governing_section.
   • On C12, all four families output FAIL (4-0) and cite Section 15.2 with verbatim quotes (4 / 4), but W2_astra and W3_velri set missing_cross_reference = true because Exhibit C (Named Direct Competitors) is summarized inline in the Transaction Record rather than attached as a separate exhibit block.

9.2 Interpretations and architectural implications

1. Why model diversity cannot substitute for canonical input-digest and cross-reference verification: Because instruction-following frontier models are trained to reason strictly from the supplied prompt context, feeding four distinct model families the same truncated RAG excerpt (C05-C08) turns instruction fidelity into a 100% common-mode failure multiplier. AIC-33 and GTP-34 certify who ran the witness; CIS-35 certifies what exact byte digest and clause chain the witness inspected.
2. Why C03 and C12 demonstrate a real calibration trade-off in structured span verification: Requiring witnesses to emit a single scalar governing_section string creates false-hold friction when a legal conclusion depends on a two-clause composition (an entitlement clause SLA Section 1.2 plus an exclusion ceiling SLA Exhibit B Section 3.4). Splitting the witness schema into entitlement_section and decisive_override_section is the natural next refinement to eliminate C03-style label splits without relaxing override verification.

10. Scope limitations and threats to validity

1. Sample size and evaluation scope: E35-B evaluated 12 contract and policy audit cases across 100 live witness evaluations (5 evaluation arms). While the 4 / 4 (16 / 16 witness calls) failure correlation on Stratum 2 is sharp, larger multi-hundred-document legal corpora will exhibit broader paraphrase and cross-reference topologies.
2. Single-pass non-adversarial witness prompting: Witnesses in E35-B ran as zero-shot structured JSON auditors without multi-turn cross-examination. Multi-turn debate can either repair C03-style section-label splits or amplify sycophantic conformity ([S09], [S12]).
3. Orchestration and verification scope: Encevra builds and evaluates the verification, evidence-grounding, and quorum release-control architecture (AIC-33, GTP-34, CIS-35) rather than training foundation model weights; all live measurements in E35-B evaluate witness outputs under deterministic orchestration and verification gates.

11. Next check (Round 36 handoff)

Round 36 should test Compositional Clause-Chain Schemas (CCS-36) on a live matched-compute multi-model benchmark: replacing the single governing_section field in CIS-35 with an ordered (base_entitlement_section, decisive_override_section, inline_fact_sufficiency) tuple to test whether the 2 / 12 (16.67%) conservative fail-closed holds observed on C03 and C12 can be reduced to 0 / 12 while preserving 0.00% false certified releases on truncated and decoy-injected contracts.

12. References ([S01]-[S12])

• [S01] John C. Knight and Nancy G. Leveson (1986). "An Experimental Evaluation of the Assumption of Independence in Multi-Version Programming." IEEE Transactions on Software Engineering, SE-12(1):96-109. https://doi.org/10.1109/TSE.1986.6312924
• [S02] Dave E. Eckhardt and Larry D. Lee (1985). "A Theoretical Basis for the Analysis of Multiversion Software Subject to Coincident Errors." IEEE Transactions on Software Engineering, SE-11(12):1511-1517. https://doi.org/10.1109/TSE.1985.231895
• [S03] Bev Littlewood and Douglas R. Miller (1995). "A Conceptual Model of Multi-Version Software." Twenty-Fifth International Symposium on Fault-Tolerant Computing (FTCS-25 Highlights), pp. 150-159. https://doi.org/10.1109/FTCSH.1995.532633
• [S04] Dahlia Malkhi and Michael Reiter (1998). "Byzantine Quorum Systems." Distributed Computing, 11(4):203-213. https://doi.org/10.1007/s004460050050
• [S05] Flavio P. Junqueira and Keith Marzullo (2003). "Synchronous Consensus for Dependent Process Failures." Proceedings of the 23rd International Conference on Distributed Computing Systems (ICDCS 2003), pp. 274-283. https://doi.org/10.1109/ICDCS.2003.1203476
• [S06] Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou (2023). "Self-Consistency Improves Chain of Thought Reasoning in Language Models." International Conference on Learning Representations (ICLR 2023). https://arxiv.org/abs/2203.11171
• [S07] Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch (2023). "Improving Factuality and Reasoning in Language Models through Multiagent Debate." arXiv preprint arXiv:2305.14325. https://arxiv.org/abs/2305.14325
• [S08] Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Shuming Shi, and Zhaopeng Tu (2023). "Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate." arXiv preprint arXiv:2305.19118. https://arxiv.org/abs/2305.19118
• [S09] Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou (2023). "Large Language Models Cannot Self-Correct Reasoning Yet." arXiv preprint arXiv:2310.01798. https://arxiv.org/abs/2310.01798
• [S10] Junyou Li, Qin Zhang, Yangbin Yu, Qiang Fu, and Deheng Ye (2024). "More Agents Is All You Need." Transactions on Machine Learning Research (TMLR 2024). https://arxiv.org/abs/2402.05120
• [S11] Fan Wu, Emily Black, and Varun Chandrasekaran (2024). "Generative Monoculture in Large Language Models." arXiv preprint arXiv:2407.02209. https://arxiv.org/abs/2407.02209
• [S12] Mert Cemri, Melissa Z. Pan, Shuyi Yang, Lakshya A. Agrawal, Bhavya Chopra, Rishabh Tiwari, Kurt Keutzer, Aditya Parameswaran, Dan Klein, Kannan Ramchandran, Matei Zaharia, Joseph E. Gonzalez, and Ion Stoica (2025). "Why Do Multi-Agent LLM Systems Fail?" arXiv preprint arXiv:2503.13657. https://arxiv.org/abs/2503.13657