Reasoning & Verification

State of knowledge

Claim-level evidence review, bounded repair, revision drift, terminal release gates, and calibration of the final revised answer.

This track draws on 35 published rounds, 113 hypotheses and 73 experiments. 121 reported findings come from published rounds and have not been independently replicated.

Reported findings

• F47.1 (deterministic, E47-A): R47 releases zero of five gold cures held by R46 and retains the universal test against all ten scripted schedule threshold releases. (Round 47)
• F47.2 (deterministic, E47-A): Exploratory R47c releases all five gold cures but permits ten unsafe schedule threshold releases. (Round 47)
• F47.3 (live, E47-B): On Witness C ledgers, R47 improves correct open outcomes from 10 to 13 without an unsafe release; on all 20 Witness B ledgers (6 completed post hoc), R47 improves from 12 to 13, also without an unsafe release. (Round 47)
• F47.4 (live, E47-B): All routing compositions escape zero unsafe outcomes. C1 with R47 routes 17 of 20, compared with 18 of 20 for C1 with R46. (Round 47)
• F47.5 (live, E47-B): Independent Witness D is less accurate with packet and brief than packet alone, 11 against 17 correct, and follows wrong resolver holds as often with both as with brief alone. (Round 47)
• F47.6 (live, E47-C): Live note following is zero on 24 exhibits. Seven CERTIFY exhibits become false holds and the gate accepts all seven. (Round 47)
• F46.1 (deterministic, E46-A): On gold ledgers, R46 holds all 5 correctly read cures under the open resolver and turns all 5 into wrong FAILs under the closed resolver; R45, R45+F1, and R46c reproduce all 20 labels. (Round 46)
• F46.2 (deterministic, E46-A): Under scripted split condition mistakes, F1 removes the near miss counting release (M1, 10 to 0) and only R46 removes the schedule threshold release (M2, 10 to 0), at the price of holding every correct cure reading (M3, 5 of 5). (Round 46)
• F46.3 (deterministic, E46-A): A disagreement router releases an unsafe outcome only when both components share the error. (Round 46)
• F46.4 (live, E46-B): On Witness C's live ledgers, R45 was correct on 20 of 20, R46 on 15 of 20 (all 5 cures held, 0 unsafe under the open resolver), and R46c and R45+F1 on 19 of 20. Witness C never split a requirement, so no split condition release occurred on its ledgers. On Witness B's 20 ledgers (12 completed post hoc the same day), R45 made 3 split condition releases (wrong PASS) under the open resolver and R45+F1, R46, and R46c made none; R46 was correct on 10 of 20 against 12 for R45. (Round 46)
• F46.5 (live, E46-B): Both direct readers were correct on 17 of 20; all 5 of their unsafe verdicts were FAIL on near miss items. No direct reading or reviewer verdict was a wrong PASS. (Round 46)
• F46.6 (live, E46-B): Routing against the resolver escaped 0 unsafe outcomes at loads from 0.30 (R45) to 0.55 (R46); routing two readers against each other escaped 2 at a load of 0.20. (Round 46)
• F46.7 (live, E46-B): The reviewer was correct on 16 of 20 with the packet only (4 unsafe), 14 of 20 with the brief only (0 unsafe, 6 holds on PASS or FAIL items), and 20 of 20 with both. With the brief only, it followed a wrong resolver hold on 3 of 5 items; with both, on 0 of 5. (Round 46)
• F46.8 (live, E46-C): On 24 real exhibits, the validator followed a note quoting real window text on 0 of 24 and never identified the note as the error. (Round 46)
• F45.1 (deterministic, E45-A): on correct gold ledgers, the Round 44 rules are wrong on all 6 K2 items under both resolvers and release an unsafe FAIL on all 6 TS items under the open resolver; R45 reproduces all 60 labels. (Round 45)
• F45.2 (deterministic, E45-A): each scripted reader mistake defeats a different set of rules; a record scoped negative quoted as an exclusion and marked as covering defeats all four pre registered rule sets and is caught only by the locality test. (Round 45)
• F45.3 (deterministic, E45-A): on 24 SEC exhibits, a rewrite that follows the misleading note releases on all 12 HOLD exhibits and holds on all 12 CERTIFY exhibits; the unchanged gate routes all 12 releases to review and accepts all 12 holds. (Round 45)
• F45.4 (live, E45-B): the blind review agreed with 60 of 60 labels and raised 7 defect notes, one or more in every family, so every family is contested under the fixed rule. (Round 45)
• F45.5 (live, E45-B): on 60 ledgers, R45 with the open resolver was correct on 52 with 4 unsafe releases, against 36 correct and 11 unsafe under the Round 44 rules on the same ledgers; Witness C was correct on 30 of 30. (Round 45)
• F45.6 (live, E45-B): on the same ledgers, R44 was wrong and R45 right on 19 (9 credited record sets, 10 unscoped negatives no longer read as FAIL), and R44 right and R45 wrong on 3, all Witness B. (Round 45)
• F45.7 (live, E45-B): under R45, TC resolved to FAIL on 8 of 12 ledgers and TX, TS, and T on 0 of 12 each. (Round 45)
• F45.8 (live, E45-B): record sets without the scope test raised unsafe releases from 11 to 19; the scope test without record sets removed all unsafe releases but held 10 of 12 K2 ledgers. (Round 45)
• F45.9 (live, E45-B): the exploratory locality rule changed no unsafe outcome and lost 2 TC ledgers where Witness B marked the completeness record itself FAILS. (Round 45)
• F45.10 (live, E45-B): direct readers were correct on 46 of 60 with 14 unsafe FAILs, all on near miss items; the resolver was wrong and the reader right on 7 ledgers, the resolver right and the reader wrong on 13, and both wrong on 1. (Round 45)
• F45.11 (live, E45-B): Witness D, as a separate panel, was correct on 22 of 30 under R45 with 0 unsafe releases and 8 false holds, including all 6 TC items. (Round 45)
• F45.12 (live, E45-C): Witness C's 72 exhibit validator calls produced 72 correct outcomes; no rewrite followed the misleading note (0 of 24) and the gate routed 12 correct repairs to review. (Round 45)
• F45.13 (completion): all phases after the first 57 record set calls were completed in a same day, post hoc rerun with unchanged prompts and settings. (Round 45)
• F44.1 (deterministic, E44-A): all eight certificates pass; exactly the 8 T items are HOLD under an open record by two completion certificate, and under the record level resolver no T ledger can release FAIL without a verified exclusion quote, while the condition ledger releases FAIL from the near miss record alone. (Round 44)
• F44.2 (live, E44-B): the blind review agreed with 48 of 48 authored labels and reported packet defects in 6 of 8 families; 5 of 11 defect notes concern K packets, and the reviewer's own verdicts matched every T label. (Round 44)
• F44.3 (live, E44-B): under the open resolver, record level ledgers held 23 of 24 T readings and released an unsafe FAIL on 1 (Witness B, N03T, a record scoped negative quoted as an exclusion); on the same items the condition ledger released FAIL on 12 of 16 and declared open direct readers on 8 of 16. (Round 44)
• F44.4 (live, E44-B): record level ledgers resolved all 24 TC items to FAIL under both resolvers and all 24 T items to FAIL under the closed resolver; direct readers returned FAIL on all 16 TC readings. (Round 44)
• F44.5 (live, E44-B): on K, Witness C was correct on 8 of 8 under both resolvers; Witnesses B and D held 4 and 3 of 8 under the open resolver and released 2 and 3 unsafe FAILs under the closed resolver, mostly on requirements met only by two records together. (Round 44)
• F44.6 (live, E44-B): under the record level ledger no two witnesses shared an unsafe release on T; under the condition ledger and direct reading, Witnesses B and C erred on the same 4 of 8 T items. (Round 44)
• F44.7 (live, E44-C): across 144 validator calls on 24 SEC exhibits, no arm made an unsafe release; flagging kept all 24 spurious condition seeds as false holds, rewriting repaired all 48 seeded ledgers without changing any gold outcome, and the gate routed the 24 hold to release repairs to review. (Round 44)
• F43.1 (deterministic, E43-A): all seven certificates pass; a flag only validator cannot create an unsafe release, while a span valid rewrite turns each of the 12 gold HOLD exhibit ledgers into an unsafe release. (Round 43)
• F43.2 (live, E43-B): undeclared, direct readers released FAIL on 33 of 36 U and T decisions; declared open, U errors fell from 15 of 18 to 0 of 18 and T errors from 18 of 18 to 6 of 18, all FAIL on near miss items (S03T, S06T, S02T). (Round 43)
• F43.3 (live, E43-B): declared closed, direct readers were correct on 36 of 36 U and T decisions and 18 of 18 K decisions; K was also 18 of 18 under the open declaration. (Round 43)
• F43.4 (live, E43-B): extraction readers under the open resolver held 18 of 18 U items and released FAIL on 14 of 18 T items, marking the requirement UNMET from the near miss record; under the closed resolver they were correct on 35 of 36 with no unsafe release. (Round 43)
• F43.5 (live, E43-B): under the open convention, 5 of 6 direct unsafe releases coincided with the same witness's extraction unsafe release (independence 2.33); error correlation was positive for every role pair and every cross witness pair. (Round 43)
• F43.6 (live, E43-C): on 48 exhibit ledgers, the flag only arm made 0 unsafe releases and 12 false holds and repaired 12; the rewriting arm made 0 unsafe releases and 0 false holds and repaired 24; neither corrupted a correct ledger. (Round 43)
• F42.1 (deterministic, E42-A): all contrast, gold resolver, and exhibit certificates pass; a pair's joint unsafe release is bounded by its smaller marginal, and zero events in 18 opportunities allow a true rate of 0.153. (Round 42)
• F42.2 (live, E42-B): on undetermined items direct readers released 11 unsafe verdicts in 18 decisions, all FAIL; extraction readers released none. (Round 42)
• F42.3 (live, E42-B): two witnesses in the direct role released the same wrong verdict on 8 of 18 pair opportunities, equal to the Frechet upper bound (independence 6.33); one witness in two roles did so on 0 of 18. (Round 42)
• F42.4 (live, E42-B): on determined items the direct role was correct 18 of 18; the extraction role gave 4 false holds in 18, turned by the pair rule into 4 of 18 role pair and 7 of 18 cross witness extraction pair false holds. (Round 42)
• F42.5 (live, E42-C): on 20 balanced real exhibits, one unsafe direct release (Z16) and two extraction false holds (Z19); no pair group released a joint unsafe verdict. (Round 42)
• F42.6 (live, pooled over 32 items): joint error was 0 of 96 for role pairs, 8 of 96 for cross witness direct pairs, and 2 of 96 for cross witness extraction pairs; error correlation was negative within every role pair and positive for every cross witness direct pair. (Round 42)
• F41.1 (deterministic, E41-A): disagreement selection keeps 42.2 percent of items and doubles the share of hard and trap items; the forced majority's wrong release rate is 0.285 in sample against 0.139 on the full pool. (Round 41)
• F41.2 (deterministic, E41-A): the measured forced majority advantage over one witness is 0.055 in sample against 0.023 on the full pool and 0.028 on a perfect twin holdout. (Round 41)
• F41.3 (deterministic, E41-A): 13.4 percent of forced majority errors lie on items disagreement never selects, and a 3-of-4 quorum releases all of them. (Round 41)
• F41.4 (live, E41-B): 0 of 10 hard by construction families produced any screening disagreement; each witness scored 9/10 on both twins and 40/40 witness twin pairs were stable. (Round 41)
• F41.5 (live, E41-B): all four witnesses released PASS on both F07 twins when the decisive delivery fact was absent; only the disagreement routed hybrid held it, at the price of one false hold on F03E caused by a joined quote. (Round 41)
• F41.6 (live, E41-C): on 20 real exhibits two disagreements occurred (Y11 and Y13, both execution only, each split Witness A and Witness C against Witness B and Witness D); the forced majority released both on a tie and the quorum and resolver held both. The other 18 exhibits were unanimous and correct. (Round 41)
• F40.1 (deterministic, E40-A): over 243 verdict states a forced majority releases 108 unsafely, the same as one witness, while raising the fewest faults for an unsafe release from 1 to 2; a 3-of-4 quorum releases 36 with 126 false holds. (Round 40)
• F40.2 (deterministic, E40-A): with witnesses that never return HOLD, single and majority topologies hold 0/16 HOLD-truth states and a 3-of-4 quorum holds 6/16. (Round 40)
• F40.3 (deterministic, E40-A): span checked extract then resolve releases 4/12 fault states unsafely (8/15 without the span check); disagreement routing limits unsafe releases to the 4/48 states with identical misreadings and needs 2 faults. (Round 40)
• F40.4 (live, E40-B and E40-C): across 33 items and 330 calls every topology made 0 unsafe releases; accuracy was 33/33 for five topologies and 32/33 for the role pipeline. (Round 40)
• F40.5 (live): Witness B alone released PASS on N01, two same-date lender letters that grant and decline one approval; majority, extraction, routed, and critique topologies all held it. (Round 40)
• F40.6 (live, E40-C): on X24 the validator added five covenants as conditions to a correct ledger and the decider held an instrument effective on its face, the only composed error. (Round 40)
• F39.1 (deterministic, E39-A): across 112 states (108 determinate, 4 hold-required, 56 with conflicting records), first-record, single-lens, and unnormalized-computation gates (G0-G3) release 24/112 states unsafely (21.43%) and the designated-record-only gate (G4) 20/112 (17.86%). Strict CEV-39 (G5) makes 0 unsafe releases with 4 false holds (3.70%); CEV-39 with verdict invariance (G6) makes 0 unsafe releases and 0 false holds by construction, a specification check rather than performance evidence. (Round 39)
• F39.2 (deterministic, E39-A): under a 3-of-4 quorum only a 2-2 lens split changes the outcome. For rule-following witnesses it holds exactly the 4 outcome relevant unresolved states; for lens-only witnesses it still releases 8 unsafe states where the records agree with each other but not with the computation the contract requires. (Round 39)
• F39.3 (live, E39-B): all four witnesses returned the correct verdict on 48/48 evaluations of designated-record and required-computation conflicts under both lenses, and on 8/8 evaluations of the invariant unresolved packet K09. Quotes were verified verbatim in the packet in 96/96 evaluations and inside the cited instrument in 85/96. (Round 39)
• F39.4 (live, E39-B): witnesses returned HOLD on 0/96 evaluations, including 0/16 on the outcome relevant unresolved packets K07 and K08. On K08, every presumptive run returned PASS and every precautionary run returned FAIL, while all 8 runs flagged the conflict. Lens dependence occurred only on K07 (3/4 witnesses) and K08 (4/4), and on 0/40 other witness and packet pairs. (Round 39)
• F39.5 (live, E39-B): across all 16 possible lens assignments, K08 was held in 6/6 balanced two and two assignments and released in 10/10 unbalanced ones. The Round 38 lens panel (A3) released K08 by 3-1; the CEV-39 gate (A8) released 10/12 correctly with 0/12 unsafe releases and 2/12 holds. (Round 39)
• F39.6 (live, E39-C): on 24 real executed exhibits from SEC EDGAR, Witness C matched the structural effectiveness label on 22/24 and Witness B on 9/9 completed; the two witnesses agreed on 8/9 paired exhibits. One mismatch traces to an audit window cut after the effectiveness clause. (Round 39)
• F39.7 (authoring audit): witness rationales identified a defect in packet K07, whose base agreement makes the invoice premature regardless of the conflict. The pre-registered label is kept and K07 excluded results are reported alongside. (Round 39)
• F38.1 (deterministic, E38-A): across 464 feasible post-register-closure states (16 positive, 448 negative), TDR-37 date-only section resolution (G2) admits 41/448 negative states (FAR 9.15%, 32 wrong verdicts) and rejects 13/16 valid states (FRR 81.25%). External operative verification without sub-clause partition (G5) leaves FAR 0.45% and FRR 6.25%; full CIV-38 (G6) reaches 0.00% FAR and 0.00% FRR by construction, a specification check rather than performance evidence. (Round 38)
• F38.2 (deterministic, E38-A): on V01-V12, TDR-37 date-only section resolution matches authored truth on 2/12 lifecycles (V11, V12), whereas the OpAsOf sub-clause resolver matches on 12/12; coarse operative resolution reaches 10/12, condition-only 6/12, and base-only 7/12. (Round 38)
• F38.3 (live, E38-B View R): unlike Round 37 (0/36 split vote sets), role-decorrelated evidentiary instructions produced 3-1 split votes on 9/12 View R packets (V01-V08 and V11). Witness B, assigned a precautionary burden-of-proof lens, was correct on 11/12; Witness A, Witness C, and Witness D, assigned presumptive register effectiveness, were correct on 4/12 each. The splits were instructed, not spontaneous. (Round 38)
• F38.4 (live, E38-B View R): because a 3-1 split still satisfies a 3-of-4 majority, ungated quorum (A3) and TDR-37 (A4) released 8/12 (66.67%) false certifications. All 36/36 receipts on the 9 verification-dependent packets set operative_validity_verifiable = false, and pre-reopen CIV-38 (A6) held those 9 packets with 0/12 false releases while releasing V09, V10, and V12. (Round 38)
• F38.5 (live, E38-B View V): with DOC-VERIF attached, 48/48 witness evaluations returned the exact verdict, controlling instrument, and sub-clause with verified quotes (A2 and A7 both 12/12 correct, 0 false, 0 holds). TDR-37 on View V (A5) held 10/12 valid packets (V01-V10) because its date-only resolver insisted on the inoperative or sibling-sub-clause instrument. (Round 38)
• In E37-A (1,536 feasible contract-lifecycle evidence states: 16 positive, 1,520 negative;), CCS-36 reference closure (G1) admits 208 negative states (FAR 13.68%, 80 with a wrong verdict); adding latest-executed-wins resolution (G2) rejects 8/16 valid states (FRR 50.00%); a self-attested register (G4) admits 24 and a stale register (G5) admits 8; the full TDR-37 gate (G6) admits 0, which holds by construction and is a specification check, not performance evidence. On L01-L12 the as-of resolver matches authored truth 12/12, executed-before-event 9/12, latest-executed-wins 7/12, base-only 5/12. (Round 37)
• In E37-B (96 live witness evaluations;), base-only packets produced unanimous 4-0 votes on 12/12 lifecycles and unanimous wrong verdicts on 7/12; ungated voting and oracle-free CCS-36 both released all 7 falsely and 0/48 receipts reported an unattached reference. TDR-37 held 11/12 on the register gap with 0 false releases. With the full lifecycle and register extract, 48/48 witness answers were correct and cited the as-of controlling instrument, so every arm, including a single Witness A witness, released 12/12 correctly. (Round 37)
• In E37-C (48 live witness evaluations;), removing the most recently executed instrument from both the packet and its register extract made 20/20 answers wrong on the 5 lifecycles where that instrument governs; TDR-37 trusting the packet's extract released all 5 falsely, while a gate-side register of record released 0 and held 11/12. Witness completeness self-reports were inconsistent (true on 5, 10, 12, and 7 of 12 packets by family) and on L08 all 4/4 witnesses asserted completeness and were wrong. (Round 37)
• In E36-A (4^7 = 16,384 multi-document configurations: 8 positive CCS_CERTIFIED states, 16,376 negative compromised states;), GTP-34 quorum agreement alone (G0) falsely accepts 4,088 states (FAR 24.96%); local single-document CIS-35 (G1) still falsely accepts 126 cross-document override states (FAR 0.77%) while rejecting 6/8 valid multi-document compositions (FRR 75.00%); multi-document attachment without precedence-DAG verification (G2) falsely accepts 56 states; and rigid scalar schemas (G3) reject 6/8 valid compositions (FRR 75.00%). Full CCS-36 (G5) achieves TP = 8, FP = 0, TN = 16,376, FN = 0 (0.00% FAR, 0.00% FRR, 100.00% accuracy). At independent witness error p_indep = 0.15, per-hop retrieval recall q = 0.85 across m = 3 documents raises 3-of-4 quorum failure from 1.20% to 28.62% (q* = 96.50% required to beat a single p_single = 0.08 full-packet auditor). (Round 36)
• In E36-B (108 live witness evaluations across M01-M12 and Witness A, Witness B, Witness C, Witness D;), local single-document CIS-35 (Arm 2) and ungated 4-family voting (Arm 3) both suffer 6/6 unanimous 4-0 false certified releases on Stratum 1 dangling overrides (M01-M06, 50.00% overall false certification rate), and Arm 2 also holds M09 and M10 fail-closed (4/12 = 33.33% correct releases). CCS-36 before reopen (Arm 4) catches all 6/6 dangling overrides fail-closed (0/12 false releases), and CCS-36 on the closed transitive bundle (Arm 5) achieves 12/12 (100.00%) certified correct releases, 0/12 false releases, and 0/12 holds, resolving both Round 35 C03/C12 hold archetypes (M09 and M10). (Round 36)
• 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. (Round 35)
• In E35-B, on Stratum 2 shared retrieval-slice truncation traps (C05-C08), both the ungated homogeneous Witness A 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. (Round 35)
• Distinct-key, immediate-issuer, terminal-component and reachability-only policies produce 216,648, 169,992, 58,320 and 165,888 false certifications in the declared uniform finite universe. Depth-2 fail-open produces 109,944 false certifications. These are combinatorial facts, not empirical rates. (Round 34)
• Depth-2 and depth-3 fail-closed validation produce zero false certifications but miss 7,776 and 648 valid graphs. Four visible edges resolve every anchor or cycle only in this bounded graph. GTP-34 is proposed; no production trust store, key, policy or deployment changed. (Round 34)
• Across 3,375 enumerated topologies (211 positive, 3,164 negative), distinct signatures and self-attested unique roots accept every world and falsely certify all 3,164 negative worlds. Optimistic one- and two-layer external audits still produce 1,364 and 434 false certifications. These are combinatorial properties of the stated uniform universe, not empirical estimates. (Round 33)
• The complete three-layer arm classifies the finite universe exactly by construction, and the fail-closed fixed-trio arm certifies 64 of 211 positive worlds with zero false certifications. The readiness audit remains 29 present, 4 partial, 1 absent with critical failures A3, I1, R4, R5 and C6; the model-facing pilot remains BLOCKED. (Round 33)
• On 256 authored synthetic cases from 16 author-known fault families, B3 classified 240/256 exactly and detected 208/224 critical faults. All 16 shared-administrator aliases were accepted; their observable-pattern sets matched clean cases. These are diagnostic results, not production reliability or out-of-template generalization. (Round 32)
• The 34-control audit remains 29 present, 4 partial, 1 absent, with critical failures A3, I1, R4, R5 and C6. Model-facing readiness remains BLOCKED. (Round 32)
• On seven authored conformance fixtures only, B0 and B1 each made 1/7 correct decisions with 6 false accepts; B2 made 5/7 with 2 false accepts; B3 made 7/7 with 0 false accepts. These do not estimate field failure rates, model accuracy, or causal effect size. (Round 31)
• The sandbox re-audit found 29 present, 4 partial, 1 absent controls, with five critical failures: A3, I1, R4, R5, C6. Live E30-B remains BLOCKED; the improved count does not transfer to production. (Round 31)
• 34 controls; 8 present, 3 partial, 23 absent; 26 critical controls partial or absent; fail-closed verdict BLOCKED. The checker validates the declared manifest, not the correctness of the human code review. (Round 30)
• Frozen snapshots, source versions, hashes, packet spans, model traces and usage telemetry are useful foundations but do not establish root-action fidelity, separation or unresolved structural equivalence. (Round 30)
• Three content probes identify all 36 observable response functions but root provenance in 0% of the 72 structural worlds. Provenance twins are interventionally equivalent under every admissible content probe. Exact diagnostic result, not a population estimate. (Round 29)
• Greedy entropy and minimax splitting identify all behaviors with 2.50 mean probes. Equating strict local response with root independence is wrong in 50% of the deliberately paired worlds. Verified root rotation establishes a prospective boundary, not historical independence. (Round 29)
• Declared-variant model intervals covered exact conditional probabilities in 100.00% of weighted histories under all three authored priors. Held-out hidden-common-origin coverage was 72.08% (uniform), 60.57% (benign), and 65.43% (issuer shift). These are exact constructed results, not production estimates. (Round 28)
• All hidden-common-origin histories remained compatible with the declared set. Passive adequacy AUROC ranged from 0.488 to 0.577. TSAR did not systematically dominate scalar routing at matched coverage; most cells tied. (Round 28)
• A model-set unanimity gate separates verdict safety from scheduling efficiency under fixed verification budgets: four policies using the 116-scenario gate produced no wrong categorical verdict in any of the 3,712 in-model policy-world runs, while a nominal target-entropy policy resolved 100 percent of cases but falsely emitted NO_CHANGE in 11.23 percent of changed cases in one regime. (Round 27)
• An out-of-model common-origin stress shows that extra evidence can increase false confidence: with the hidden_common_origin variant excluded from the planning set, the exact MSAV policy's false-clean rate reached 33.68 percent among changed worlds at budget 9. Myopic adaptivity can underperform a fixed target-first schedule. (Round 27)
• In the declared one-fault universe, SACP leaves eight opposite-target ambiguity pairs; adding one independently rooted external commitment reduces that count to zero. Under two faults, SACP plus external commitment leaves 26 pairs; adding an independently rooted behavior probe reduces that count to zero. These are exact finite-model counts, not prevalence-weighted rates. (Round 26)
• Against three failed roots, four genuinely independent target-sensitive roots identify the target in the declared model. If the witness shares the issuer root, no available set identifies change against three failed roots. Root independence is an assumption. (Round 26)
• On 80 authored traces, SACP alerted on 55 of 60 gap traces and none of the 20 negative controls (trace recall 0.9167, precision 1.0000, claim-weighted recall 0.8944). Revalidate-all achieved 1.0000 recall but alerted on all 20 negative controls and created 400 claim-review tasks, compared with 161 for SACP. A sequence-only policy falsely alerted on all five benign-reordering controls. These numbers show that the implementation conforms to the authored distinctions. They do not estimate production precision, recall, incident prevalence, reliability, cost, or user impact. (Round 25)
• On 23 authored fixtures, the proposed authority-lineage selector preserved every expected claim edge with one conservative extra edge, while repository-only selectors missed every deliberately non-repository fixture. Conformance to an authored model, not evidence of production effectiveness. (Round 24)
• The hybrid selector found all three artifacts labeled relevant in the real interval, including one uniquely added by an independently derived import view. It also selected four irrelevant artifacts and invalidated two claims that adjudication later found unaffected. Design evidence from one interval, not a general validation result. (Round 23)
• On the actual 127-path diff, the measurement-only watchlist selected 0 of the 17 audited critical paths, the filename heuristic 12 of 17 with 25 false positives, retest-all all 127, and the authored CTAG manifest exactly the 17. The 17-path set and tags were manually authored from the same change set, so this is not discovery accuracy or compute savings. (Round 22)
• On 32 authored mutation fixtures CTAG routed 32 of 32 exactly (direct watchlist 9, filename keyword 7, retest all 3). Specification tests on authored fixtures, not evidence of out-of-sample accuracy. (Round 22)
• Two historical claims preserved; five current claims made stale, then restored after direct inspection; two hypotheses left untested. (Round 22)
• Under the current three-write best-effort persistence model, seven of eight possible write states are incomplete yet do not block release; ABMG audit-qualifies none of those seven. Logical state-space result, not a measured production rate. (Round 21)
• Ten envelope-only mutations collide under the current content-only fingerprint, 24 equal-byte different-version source pairs and 24 distinct service-response pairs collapse under the current records, and five bad stream fixtures are falsely accepted by a development model of the current parser; the ABMG projection separates or rejects all of them. Constructed fixtures, not production results. (Round 21)
• Endpoint and generic operational records miss evidence-specific corruption: A0 detected 2 of 12 mutation families and A1 detected 4 of 12, while content-bound receipts detected all 12 across 96 episodes with zero clean false positives. Constructed-fixture results, not model performance. (Round 20)
• An operator-only digest chain does not survive a privileged rewrite by that operator: after simulated rewrite, operator-only arms detected 0 of 12 families. Receipts bind observable delivery under assumptions; they do not prove source truth, completeness before emission, service-internal processing, semantic support, or causal evidence use. (Round 20)
• Constructed collision: terminal output plus aggregate usage cannot determine whether the corrective source was delivered. A narrow indistinguishability result over constructed records, not model performance. (Round 19)
• Static audit: 4 of 18 collector-contract requirements present, 3 partial, 11 absent at the required granularity. (Round 19)
• Constructed example: a timestamp-only newest-recorded rule returns a late January correction (60) for a March current query whose correct fixture answer is 75. A counterexample over hand-specified records, not model performance. (Round 18)
• Exact check: 24 temporal-query checks hold across all six arrival orders of the three-record example, plus 12 checks of baseline failure, conflict handling, dependency repair, alternative support, and source integrity. (Round 18)
• A supported fact does not establish decision sufficiency: the exact fictional policy counterexample leaves both review decisions possible. (Round 16)
• Exact finite-world enumeration yields 3,778 determined and 3,134 underdetermined pairs among 6,912 pairs. These are checker diagnostics, not model performance. (Round 16)
• A valid decision-changing pair establishes a gap within the modeled scope; failed bounded search does not certify completeness. Scope expansion can reopen a decision. (Round 16)
• Exact constructed example: 97.5% snapshot-label accuracy coexists with a reversed comparison (true −2 points, judged +3 points). This is a possibility result on invented labels, not a measured Noetis failure rate. (Round 14)
• Exact constructed example: a pooled confusion correction restores the overall mean to 79% while the corrected comparison stays in the wrong direction (+3.4054 points). (Round 14)
• Exact constructed example: six valid identical-artifact couplings narrow whole-cohort bounds from [0, 20] to [6, 14] points without resolving either artifact's correctness. (Round 14)
• Replaying the current scorer reproduced all 119 nonempty stored scores. (Round 13)
• Appending the full source as a labeled appendix removed detected-issue credit in all 114 outputs that had it. (Round 13)
• Removing appended approved findings changed detection counts in 44 of 59 nonempty panel outputs. (Round 13)
• Saved results keep the final brief and latest check, but not before/after pairs, so a repair flag cannot distinguish a fix from a new error. (Round 13)
• Usage data existed for 19 panel cases and no baseline cases. (Round 13)
• Exact synthetic diagnostics show how candidate thresholds trade coverage for an upper risk bound at fixed δ = 5% and family size 3. These are worked calculations, not model performance. (Round 12)

Rounds in this track

1. Round 47 · 2026-11-05
   When a Replacement Is Not a Cure: Supersession, Schedule Thresholds, and Independent Review

   Does explicit supersession release correctly read cures without weakening schedule threshold protection, and what must an independent reviewer see?

   Notebook
2. Round 46 · 2026-11-04
   Who Must Look Again? Disagreement Routing, False Hold Load, and What a Reviewer Must See

   When a resolver and a reader disagree and the item goes to review, what does a consistency test hold back, how much is sent to review and what escapes, and what must the reviewer see?

   Notebook
3. Round 45 · 2026-11-03
   The Rule or the Reader? Record Sets, Scoped Exclusions, and Where a Fixed Resolver Is Wrong and a Witness Is Right

   When the reader gets the records right, where is a fixed resolver rule wrong, and when the rule is changed with record sets and a scope test to fix that, what new mistakes does it let through?

   Notebook
4. Round 44 · 2026-11-02
   This Record Fails, or No Record Exists? Record Level Status Ledgers, Completeness Records, and Independently Reviewed HOLD Labels

   When an evidence ledger records what each record shows and asks separately whether anything excludes other records, do readers stop treating a defective record as proof that no valid record exists, and do the authored HOLD labels survive a blind independent review?

   Notebook
5. Round 43 · 2026-11-01
   Open Record or Closed Record? Declared Record Conventions, Near Miss Records Read as Negation, and Flag Only Against Rewriting Validators

   When the record convention is declared, do readers still read a missing or defective record as proof that an event did not happen, and should a validator flag ledger defects or rewrite them?

   Notebook
6. Round 42 · 2026-10-31
   Role or Witness? Measured Joint Misreading of Undetermined Facts by One Witness in Two Roles and by Two Witnesses in One Role

   Does one witness reading in two roles, a direct verdict and an evidence ledger resolved by deterministic code, misread undetermined facts jointly less often than two different witnesses reading in the same role?

   Notebook
7. Round 41 · 2026-10-30
   Hard by Construction or Hard by Selection? Disagreement Mined Item Pools, Isomorphic Twin Holdouts, and the Consensus Errors That Disagreement Cannot Find

   Is selecting items on which witnesses disagree a sound way to build the hard pool that separates composition topologies, or does it overstate how much composition helps and leave consensus errors exactly where it cannot look?

   Notebook
8. Round 40 · 2026-10-29
   Composition Before Capability: Six Noetis Composition Topologies Compared on Release Safety, Abstention, and Calls per Decision

   Holding the evidence questions fixed, which way of composing four witnesses into one release decision gives the fewest unsafe releases and the best abstention per model call: a single witness, a forced majority, extract then deterministic resolve, a role specialized pipeline, critique and revision, or a disagreement routed hybrid?

   Notebook
9. Round 39 · 2026-10-29
   When the Records Disagree: Conflicting Collateral Evidence, Evidentiary Lens Balancing, and Abstention Failure in Multi-Witness Contract Audits

   Once every condition to effectiveness has an attached verification record (CIV-38), what should a multi-witness gate do when two attached collateral records of the same event disagree, and can a balanced two and two assignment of precautionary and presumptive evidentiary lenses make a 3-of-4 quorum abstain when the contract does not resolve the conflict?

   Notebook
10. Round 38 · 2026-10-28
    Signed Is Not Operative: Conditional Effectiveness, Counterpart Authority, and Sub-Clause Scope Partition in Multi-Witness Contract Audits

    Once a gate-side register of record has closed the instrument set and resolved the event date as-of (TDR-37), what happens when an attached, dated, signed instrument is still non-operative (due to an unsatisfied condition precedent, unilateral counterpart execution, a defective notice, or service to the wrong channel) or amends only one sub-clause while leaving a sibling sub-clause in the same section intact?

    Notebook
11. Round 37 · 2026-10-28
    When the Contract Changed After Signing: Event-Date Resolution, Register-Anchored Completeness, and Stale Registers in Multi-Witness Contract Audits

    When a later amendment, waiver, termination notice, or revocation changes a contract obligation without any forward pointer in the base agreement, what lets a multi-witness audit decide which instrument governed on the event date, and what makes the completeness of the instrument set checkable rather than self-asserted?

    Notebook
12. Round 36 · 2026-10-28
    When One Complete Document Is Still Half the Truth: Cross-Document Reference Closure, Precedence DAGs, and Compositional Clause Chains Across Multi-Instrument Quorums

    When each witness inspects a 100% byte-complete primary agreement inside a multi-document transaction graph G_doc = (D, E_ref, pi_prec), how do dangling incorporation-by-reference edges, precedence inversions, and compositional clause chains cause false certifications or false holds, and does Cross-Document Completeness and Sufficiency (CCS-36) achieve 0/12 false releases and 0/12 false holds at matched compute?

    Notebook
13. Round 35 · 2026-10-27
    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?

    Notebook
14. Round 34 · 2026-10-25
    Can Trust Bootstrap Itself? Circular Attestation, Sybil Issuers, and the Minimum Grounded Trust Base

    When witness-independence claims are supported by chains of cross-certificates or attestations, what observable condition distinguishes a genuinely grounded quorum from a circular or Sybil-generated endorsement structure, and how should bounded verification behave when a path cannot be completed?

    Notebook
15. Round 33 · 2026-10-24
    Who Certifies Independence? Provenance Completeness and the Limits of Self-Attested Witness Quorums

    Can a collaboration system certify that a 3-of-4 witness quorum is administratively independent using only witness-generated identities and provenance claims, or does a useful certificate require externally anchored and sufficiently complete dependency evidence?

    Notebook
16. Round 32 · 2026-10-23
    Who Watches the Witness? Boundary Hardening and a Blinded Mutation Holdout for the RISE Gate

    Under a fixed, replayable evidence bundle, can separately keyed witness, registry, and monitor services with append-only receipts detect execution and evidence-history faults on randomized hidden instances without rejecting clean cases, and what remains undetectable when all services share one administrator?

    Notebook
17. Round 31 · 2026-10-22
    Did the Root Actually Move? A Signed Sandbox Dry Run of the RISE Gate

    Can a signed, separately witnessed sandbox prove enough about one file-backed root action to distinguish clean execution from specified fidelity, freshness, isolation, and identifiability failures before any model inference is allowed?

    Notebook
18. Round 30 · 2026-10-21
    Can We Intervene for Real? A Fail-Closed Readiness Gate for Active Evidence-Lineage Trials

    Can the current Encevra/Noetis workflow execute an active evidence-lineage trial whose causal interpretation is warranted, and what minimum fail-closed protocol is required before E29-B may run?

    Notebook
19. Round 29 · 20 October 2026
    Round 29 - Make the Evidence Move: Active Source Interventions and the Limits of Dependency Discovery

    Under a fixed verification budget, which controlled interventions can distinguish independent evidence roots from shared-root facades in a collaborative reasoning system, and what evidence justifies an independence claim rather than an explicit unresolved-equivalence finding?

    Notebook
20. Round 28 · 19 October 2026
    Round 28 - Confidence About What? Separating Target Probability, Structural Ambiguity, and Model Adequacy

    Under a fixed evidence and compute budget, does separating target probability, structural ambiguity, and model adequacy support safer abstention than a single confidence scalar?

    Notebook
21. Round 27 · 18 October 2026
    Round 27 - Which Check Should Run Next? Model-Set-Safe Verification Under a Fixed Budget

    Under a fixed observation budget, can a sequential verification policy improve correct resolution while avoiding harmful false confidence when the cost, efficacy, and independence of checks are uncertain?

    Notebook
22. Round 26 · 17 October 2026
    Round 26 - What Can the Watchers Ever Know? Minimum Observables for Detecting Omitted Change

    For each omitted-change failure class, which changed and unchanged worlds remain observationally identical, and what minimum independently rooted observable or trust assumption separates them?

    Notebook
23. Round 25 · 16 October 2026
    When Silence Looks Like Stability: Detecting Missing Change Evidence in Collaborative AI Systems

    Can an AI assurance system distinguish a genuinely quiet authority from a broken, delayed, stale, selectively silent, or common-mode observation path?

    Notebook
24. Round 24 · 15 October 2026
    Do Two Watchers Count as Two? Authority Lineage and Common-Origin Failure in Claim Revalidation

    How should Encevra represent the lineage and authority of change evidence so that multiple analyses of one source are not mistaken for independent coverage, and so that changes outside the repository can trigger bounded claim revalidation?

    Notebook
25. Round 23 · 14 October 2026
    Can the Gate Catch Tomorrow's Change? A Prospective Shadow Trial of Claim Revalidation

    Can a prospective gate notice when a genuinely later software change should make a current claim stale, without unnecessarily withdrawing unrelated claims or rewriting history?

    Notebook
26. Round 22 · 13 October 2026
    When Does an Audit Expire? Dependency-Triggered Revalidation for Rapidly Changing Agent Systems

    Can a typed claim-dependency graph automatically mark only the assurance claims affected by code, configuration, data, or external-contract changes as stale, preserve historically valid results, and require the right revalidation?

    Notebook
27. Round 21 · 12 October 2026
    Where Did the Trace Break? An Attempt-Bound, Fail-Closed Measurement Gate for Noetis

    Can a boundary-tapped, attempt-bound, atomic measurement gate reconstruct every Noetis external-service attempt and refuse audit-qualified release when the trace is incomplete, at fixed model compute?

    Notebook
28. Round 20 · 11 October 2026
    What Must the Log Reveal? Minimal-Disclosure Receipts for Verifiable Evidence Delivery

    Can salted, ordered, independently witnessed invocation receipts preserve exact application-observable evidence delivery for authorized audit while keeping raw prompts and source text out of routine telemetry, with acceptable leakage, false-alarm, storage, latency, and reconstruction costs?

    Notebook
29. Round 19 · 10 October 2026
    Did the Evidence Reach the Agent? Reconstructable Evidence Lifecycles for Reliable Experiments

    Can an Evidence Lifecycle Observation Contract reconstruct which immutable source version was admitted, what ordered material crossed each model invocation boundary, and how claims and the terminal output were bound to those events, better than final records and generic traces?

    Notebook
30. Round 18 · 9 October 2026
    When Evidence Changes: Versioned Support and Bounded Belief Repair

    At fixed total work and source access, does an append-only record that distinguishes valid time from recorded time and tracks claim dependencies improve current, historical, as-known, and downstream-consequence answers over strong overwrite, chronological-history, and newest-recorded baselines, without increasing false invalidation, unhelpful dispute states, or cost?

    Notebook
31. Round 16 · 7 October 2026
    Before the Answer Closes: Detecting Missing Decision Conditions

    At a fixed total inference and source-access budget, does requiring a validated pair of evidence-compatible scenarios with different decisions reduce unsupported commitment more effectively than a strong missing-information checklist, without reducing useful completion or increasing harmful revision?

    Notebook
32. Round 14 · 5 October 2026
    When the judges are uncertain: measuring revision gains without an infallible evaluator

    Under what evaluation design does an apparent revision gain survive plausible uncertainty and differential bias in the labels used to measure it?

    Notebook
33. Round 13 · 4 October 2026
    From final scores to measurable corrections

    Can saved Noetis records show whether a correction actually helped?

    Notebook
34. Round 12 · 4 October 2026
    After the last revision: calibrating release risk

    Can the final revised answer be released under a defined selective-risk target?

    Notebook
35. Round 11 · 3 October 2026
    When should collaborators spend their next unit of compute?

    Should verification effort go where a specific check is predicted to repair, rather than where uncertainty is highest?

    Notebook

Hypotheses

• H47.1R47 releases cures that R46 holds on gold and live ledgers, with no added unsafe release.Rejected
• H47.2R47 still removes the scripted schedule threshold release.Supported
• H46.1On live ledgers, R46 makes no more unsafe releases than R45 and removes the split condition releases.Supported
• H46.2R46 adds false holds on R items relative to R45; R46c recovers them without an added unsafe release (exploratory).Supported
• H45.1On the same ledgers, R45 under the open resolver gives more correct outcomes than R44 without more unsafe releases.Supported
• H45.3The scope test removes unsafe FAILs on TS under the open resolver.Supported
• H44.1With a record level ledger, near miss T items hold under the open resolver, and the unsafe FAIL rate on T falls well below the condition ledger's.Supported
• H43.1Declaring an open record removes the direct FAIL releases on U and T items.Rejected
• H42.2The extraction role holds undetermined items more often than the direct role.Supported
• H42.3The extraction role pays a measurable false hold cost on determined items.Supported
• H40.2Separating extraction from decision with a span check removes fabricated quotes as a route to release.Supported
• H39.3Witnesses apply a designated record and a contract-required computation regardless of assigned lens.Supported
• H39.4Witnesses return HOLD when records of equal standing conflict, the contract is silent, and the outcome depends on the conflict.Rejected
• H39.5A balanced two and two lens assignment holds the clean unresolved packet while every unbalanced assignment releases it.Supported
• H38.3Role-decorrelated witness instructions break Round 37's unanimity in View R, producing 3-1 split votes on packets whose operative validity depends on an unattached DOC-VERIF log.Supported
• H38.4When DOC-VERIF is omitted (View R), a 3-of-4 quorum (A3) and TDR-37 (A4) still release 8/12 false certifications because three presumptive witnesses outvote one precautionary witness.Supported
• H38.5When DOC-VERIF is attached (View V), TDR-37 (A5) holds 10/12 valid packets because its date-only check rejects witnesses that rightly disqualify inoperative or sibling-sub-clause instruments.Supported
• H37.1With base-only packets, pointer-free post-execution instruments produce unanimous false consensus that CCS-36 reference closure does not intercept.Supported
• H37.3TDR-37 with a gate-side register holds base-only packets with zero false releases.Supported
• H36.1Applying Round 35's CIS-35 gate to a 100% locally complete primary document without cross-document reference-graph closure (G1 in E36-A; Arm 2 in E36-B) permits unanimous false certifications whenever an unattached higher-precedence companion instrument reverses the base obligation.Supported
• H36.3Attaching all referenced documents and checking verbatim spans without verifying the contractual Order-of-Precedence DAG and active amendment execution status (G2 in E36-A) still falsely accepts 56 compromised states where witnesses apply a subordinate, superseded, or unexecuted clause.Supported
• H36.5Across all 4^7 = 16,384 enumerated multi-document states in E36-A, the complete CCS-36 gate (G5_ccs36_full) achieves TP = 8, FP = 0, TN = 16,376, FN = 0 (0.00% FAR, 0.00% FRR, 100.00% accuracy).Supported
• H36.6On the 12-packet live 4-architecture benchmark (E36-B), applying CCS-36 prior to bundle reopen (Arm 4) intercepts 6/6 Stratum-1 dangling-override packets (M01-M06) fail-closed (HOLD_CCS36_DANGLING_REF_CLOSURE) with 0/12 false certified releases, while ungated 4-family voting (Arm 3) and local CIS-35 (Arm 2) both suffer 6/6 unanimous 4-0 false certified releases.Supported
• H35.1On 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.Supported
• H35.2Adding 4-family witness architecture diversity (Witness A, Witness B, Witness C, Witness D) without canonical context-completeness and override gating fails to prevent unanimous false consensus when witnesses share a truncated retrieval slice (C05-C08).Supported
• H35.3Checking 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.Supported
• H35.4The 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.Supported
• H34.1For any fixture whose complete signer paths remain inside a strongly connected component with no configured anchor, GTP-34 will never return `certified`, regardless of the number of valid signatures inside the component.Proposed
• H34.2When the action requires `q` independent roots, GTP-34 will not count multiple witness paths terminating at the same administrative root more than once.Proposed
• H34.3Timeouts, missing intermediates, depth limits, and lookup failures will increase `unproven`, never create a new accepted root.Proposed
• H34.4An in-band anchor update will be accepted only when authorized under the current anchor policy and the proposed new policy, except during a separately authenticated recovery event.Proposed
• H34.6On independently authored structural fixtures, GTP-34 will reduce critical false certification relative to distinct-key, immediate-issuer, reachability-only, and fail-open baselines without changing the anchor store after seeing fixture labels.Proposed
• H33.1On adversarial common-controller aliases, distinct keys plus self-attested root claims will not reduce false certification relative to key count alone.Proposed
• H33.2Conservative externally anchored coverage will never increase the set of certified witness coalitions when it reveals a shared root.Proposed
• H33.3Fail-closed handling of missing critical layers will reduce false certifications relative to optimistic handling, at the cost of more unproven outcomes.Proposed
• H32-AA separately administered witness plus an independently monitored append-only registry will eliminate the shared-admin alias under the stated threat model, provided at least one authority remains honest and reachable.Proposed
• H32-BWhen model calls and prompts are held fixed, adding externally anchored execution evidence will reduce false correction and harmful revision more than it reduces beneficial correction.Proposed
• H32-DA threshold witness scheme will outperform a single external witness under one-witness compromise but will introduce availability and latency regressions that require explicit abstention rules.Proposed
• H31.1When the same executor authors both the action and success claim, adding a valid signature will not materially improve detection of off-target, partial, false-success, or stale-readback faults relative to unsigned self-report.Proposed
• H31.2Comparing frozen intent with a separately signed direct observation will detect at least 95% of off-target, partial, false-success, and stale-readback holdout faults at no more than 2% false rejection on clean cases.Proposed
• H31.3Cache and memory receipts plus hidden canaries reduce undetected spillover by at least 75% relative to a direct-action witness without exposure telemetry, with false-positive rate below 5%.Proposed
• H31.4A candidate-set representation with `unresolved_equivalence` reduces false root-independence by at least 30% relative to forced binary classification while retaining at least 70% structural-decision coverage.Proposed
• H31.5Moving the witness signer and receipt registry into separately administered, policy-enforced services and running each arm in an ephemeral isolated environment will close A3, I1, R4, and R5 without introducing more than 2% accounting overhead or unacceptable operational failure.Proposed
• H31.6Once the infrastructure gate passes, full RISE will reduce false-root independence and harmful revision relative to the strongest active baseline at matched total compute, without more than a 1.5 percentage-point answer-accuracy loss.Proposed
• H30.1At the audited snapshot, the current E29-B path fails at least one critical Root, Intervention, Separation, Receipt, and Equivalence control.Proposed
• H30.2The dual-path comparator detects at least 80% of injected off-target, partial and false-success actions and improves detection by at least 20 percentage points over signed self-report at the same action count.Proposed
• H30.3Per-arm cache and memory receipts plus canaries reduce undetected contamination by at least 75% relative to logical arm labels alone.Proposed
• H30.4An unresolved-equivalence state reduces false root-independence by at least 30% while retaining at least 70% structural-decision coverage.Proposed
• H30.5Full RISE lowers false-root-independence and harmful revision relative to the best active baseline at matched total compute, without more than a 1.5 percentage-point loss in accuracy or a 2% compute-budget overrun.Proposed
• H29.1Within a correctly specified finite response-function family, an adaptive separating policy will decide the target behavioral-dependency predicate more often than passive repetition at equal total cost.Proposed
• H29.2When separate-root and shared-root implementations are interventionally equivalent under the admissible content interface, no content-only policy will outperform the prior on root provenance.Proposed
• H29.3Requiring verified action receipts will reduce false-clean dependency judgments under unknown and off-target interventions compared with trusting intended targets.Proposed
• H29.4Verified root rotation plus path isolation will reduce prospective shared-failure correlation and false corroboration more than content perturbation alone.Proposed
• H29.5Stopping when all surviving candidates share the same decision label will preserve target decision accuracy while using fewer interventions than full graph identification.Proposed
• H29.6Certificates formed without reserved out-of-version-space validation will exhibit a higher false-clean rate on omitted mechanisms than certificates with a reserved validation action.Proposed
• H28.1For every evaluated history generated by a declared model with matched prior and likelihood, the exact conditional target probability lies inside the reported model interval.Supported
• H28.2At matched coverage, TSAR's conservative set risk yields lower delayed-correction risk than pooled scalar risk across both declared variants and all preregistered priors.Inconclusive
• H28.3Surprise, interval width, or compatibility detects the held-out hidden-common-origin variant with AUROC at least 0.75 and useful precision above prevalence.Rejected
• H28.4Replacing one passive repeat check with a provenance-changing intervention will raise held-out model detection above AUROC 0.75 and reduce false-clean rate at equal total cost.Proposed
• H28.5On a preregistered multi-agent benchmark with delayed adjudication, an interoperable TSAR schema will improve correction-risk Brier score and 90%-coverage risk over a one-number confidence message, without reducing task accuracy or increasing total compute/tool cost.Proposed
• H27.1For every scenario inside a correctly specified finite model set, a target-unanimity gate will never emit the wrong categorical target.Proposed
• H27.2At matched compute in strata with correlated evidence roots, a model-set gate will have a lower harmful false-clean rate than a nominal independence gate, while retaining nonzero correct resolution.Proposed
• H27.3Given the same declared model, costs, and planning distribution, finite-horizon optimization will weakly dominate any included fixed schedule on expected categorical resolution.Proposed
• H27.4A greedy next-check policy can underperform a target-first fixed schedule because an attractive immediate reduction can block an affordable completing path.Proposed
• H27.5When apparently independent checks share an omitted root, increasing budget can increase the conditional false-clean rate.Proposed
• H27.6A scheduler that models both target value and check efficacy will improve cost-to-correct-resolution or selective risk under shifts in check reliability.Proposed
• H27.7Among abstained cases, the recorded number and structure of cross-target scenario pairs will predict which additional checks later change the answer.Proposed
• H26.1Under one injected silent-root fault, SACP plus one genuinely independent target channel has fewer harmful false-stability certificates than SACP at matched total cost.Proposed
• H26.2Three root-disjoint target channels are sufficient against two modeled silent-root failures in the implementation.Proposed
• H26.3Removing or falsifying root-lineage metadata increases false stability under common-origin witness failures.Proposed
• H26.4Health-only checks improve abstention and reduce false certification for visible path failures but do not detect healthy-looking selective non-emission.Proposed
• H26.5Additional LLM agents sharing the same evidence root do not match one independent external channel at equal inference and observation cost.Proposed
• H26.6Ambiguity-directed probes reduce residual pair weight per observation unit relative to fixed and random probes.Proposed
• H26.7OSC's safety gain is not fully explained by abstention.Proposed
• H22-1Under equal total verification budget, CTAG will produce a lower stale-claim escape rate than no invalidation, fixed-file watchlists, filename heuristics, and periodic expiry.Proposed
• H22-3Execution-path closure and unknown-path fail-closed rules will reduce misses for negative claims compared with direct-file watchlists.Proposed
• H21-2Under injected single-write, process-crash, and retry-race failures, false audit_qualified_released is at most 0.1 percent with a two-sided 95 percent upper confidence bound at or below 0.5 percent.Proposed
• H21-3Under truncated streams, malformed events, and HTTP-200-then-error cases, false-complete classification is at most 0.1 percent and recall for valid complete streams is at least 99.5 percent.Proposed
• H20-1Under authorized reveal, MDIR reaches at least 0.95 exact exposure reconstruction with no more than a 0.03 absolute deficit from inline logging, while routine raw sensitive bytes remain zero.Proposed
• H19-2B1 claim lineage improves exact first-divergence localization over B0 without increasing false causal attribution.Proposed
• H18-2Dependency-directed repair improves downstream-consequence correctness beyond bitemporal selection alone.Proposed
• H17-2An explicit scope check recovers confident omissions on changed-question cases.Proposed
• H16-1Validated alternatives improve supported task completion and reduce unsupported commitment relative to a strong checklist at the same budget.Proposed
• H14-3Valid joint constraints reduce evaluation uncertainty at useful cost.Proposed
• H13-1Evidence receipts at the revision boundary improve task success over ordinary reconsideration.Proposed
• H12-1Terminal evidence-and-revision features improve held-out release coverage by at least five points at a 2% selective-risk target.Proposed
• H11-2Increased checking does not raise harmful revisions beyond 0.5 points.Proposed
• H25-1Gap recall: At fixed total monitoring and revalidation budget, SACP reduces independently adjudicated affected-claim gap misses versus receipts-only, heartbeat-only, checkpoint-sequence, and periodic-expiry baselines.Proposed
• H25-2Availability: SACP preserves more correctly current claim-hours than revalidate-all while staying within a preregistered non-inferiority margin for affected-claim gap recall.Proposed
• H25-3Lineage: Removing root-lineage checks increases misses for mirrored/common-origin failures without reducing detection of independently rooted conflicts.Proposed
• H25-4Maturity window: Removing the reorder maturity window increases false observation gaps on delayed but complete streams.Proposed
• H25-5Challenges: Removing challenge canaries increases detection delay for observer crashes and stale checkpoint freezes.Proposed
• H25-6Independent reconciliation: Removing separately rooted reconciliation increases misses for selective non-emission and equivocation.Proposed
• H25-7Expectation weighting: At the same total challenge count, authority-specific expectation-weighted scheduling improves gap recall or detection delay without increasing false suspicion versus uniform scheduling.Proposed
• H25-8Scoped state: Claim-scoped gap projection reduces unnecessary claim withdrawal versus authority-wide or registry-wide fail-closed policies without increasing stale-claim escape.Proposed
• H25-9User-facing honesty: Explicit `unknown_observation_gap` status reduces unsupported current claims without causing more abandonment or misunderstanding than a generic stale label in a blinded review.Proposed
• H25-10Complexity: A simpler checkpoint-sequence-plus-end-to-end-probe policy matches SACP on safety, availability, and cost. Support for this hypothesis would reject the need for full SACP.Proposed
• H24-1authority-gap recall: Across a preregistered prospective window, ALR will leave fewer independently adjudicated affected claims current after non-repository changes than the Round 23 repository union at the same total event-inspection and claim-revalidation budget.Proposed
• H24-2unique authority value: Removing each of deployment, live-database, service-contract, and runtime observers will create at least one unique stale-claim escape in the corresponding held-out authority stratum.Proposed
• H24-3typed selectivity: At matched stale-claim escape, typed authority-to-claim mapping will produce fewer false-invalidated claims and fewer reviewer minutes than authority broadcast or retest-all.Proposed
• H24-4unknown handling: Authority-bounded fail-closed handling will reduce stale-claim escape relative to treating unknown events as irrelevant, without increasing false-invalidated claim time by more than a preregistered noninferiority margin.Proposed
• H24-5scope honesty: Adding `current_within_observed_authorities` and `unknown_authority_gap` will reduce public universal-current claims unsupported by current authority evidence to zero, without reducing accurately scoped publication coverage by more than five percentage points.Proposed
• H24-6lineage integrity: Every event used to restore a claim will retain a verifiable chain to its root authority, observer version, issuer, freshness policy, and revalidation procedure; any missing field must block restoration.Proposed
• H24-7common-cause awareness: When two views share a root authority, ancestry-aware integration will produce fewer false assurances than naive view counting at equal claim coverage, measured on correlated-miss probes and prospective events.Proposed
• H24-8total cost: Monitoring, verification, graph maintenance, privacy review, and revalidation labor for ALR will remain below retest-all at matched stale-claim escape over the full prospective window.Proposed
• H23-1Stale-claim escape: across a preregistered multi-release window, the hybrid will leave fewer independently adjudicated affected claims current than the frozen exact-path manifest at the same claim-revalidation budget.Proposed
• H23-2Unique-view value: removing the static-import view will increase affected-artifact false negatives on at least one held-out release without an equal reduction in false-invalidated claims.Proposed
• H23-3Unknown-path value: removing the fail-closed unknown rule will increase stale-claim escape on held-out adapter, configuration, or execution-path changes.Proposed
• H23-4Selectivity: at equal stale-claim escape, the hybrid will select fewer artifact-review tasks and fewer unaffected claims than retest-all.Proposed
• H23-5Availability: the hybrid plus cached claim-specific procedures will maintain greater useful-current availability over time than a policy that expires all current claims on every production change.Proposed
• H23-6Lifecycle integrity: neither selector execution nor revalidation will rewrite historical findings or upgrade untested hypotheses in any evaluated release.Proposed
• H23-7Cost: end-to-end review labor, including graph maintenance and selector triage, will be lower for the hybrid than retest-all at matched stale-claim escape.Proposed
• H23-8Mismatch trigger: revalidation triggered by a typed dependency mismatch will achieve lower unnecessary revalidation than periodic expiry at matched stale-claim escape.Proposed

Experiments

• E47-AStructural certificates for supersession, schedule thresholds, routing, and public exhibitsExecuted (diagnostic)
• E47-BLive supersession, disagreement routing, and three independent reviewer viewsExecuted (bounded diagnostic)
• E47-CValidators on 24 new public exhibit ledgers with notes quoting real window textExecuted (bounded diagnostic)
• E46-AStructural certificates, gold ledgers under four rule sets, scripted split condition mistakes, routing certificate, and exhibit and gate certificatesExecuted (diagnostic)
• E46-BBlind label review, record set ledgers under four rule sets, disagreement routing, and three reviewer views on authored familiesExecuted (bounded diagnostic)
• E46-CValidators on real SEC exhibit ledgers with a note quoting real window textExecuted (bounded diagnostic)
• E45-AStructural certificates, gold ledgers under four rule sets, scripted reader mistakes, and exhibit and gate certificatesExecuted (diagnostic)
• E45-BBlind label review and live record set ledgers resolved by four fixed rule sets on authored familiesExecuted (bounded diagnostic)
• E45-CValidators on real SEC exhibit ledgers with a misleading reviewer noteExecuted (bounded diagnostic)
• E44-AStructural certificates for variants, completion labels, resolver contrast, exhibit ledgers, and the rewrite gateExecuted (diagnostic)
• E44-BBlind label review and live record level ledgers on authored near miss familiesExecuted (bounded diagnostic)
• E44-CGated validators on real SEC exhibit ledgers with harder seeded defectsExecuted (bounded diagnostic)
• E43-AStructural certificates for variants, conventions, exhibit novelty, and validator monotonicityExecuted (diagnostic)
• E43-BLive declared convention study on authored near miss familiesExecuted (diagnostic)
• E43-CFlag only against rewriting validators on real SEC exhibit ledgersExecuted (diagnostic)
• E42-AStructural certificates, gold resolver ledger, and Frechet bounds for reader pairsExecuted (diagnostic)
• E42-BLive contrast families read in two roles by three witnessesExecuted (diagnostic)
• E42-CLive two role reading of real SEC exhibitsExecuted (diagnostic)
• E41-AExact disagreement selection enumeration with twin holdoutsExecuted (diagnostic)
• E41-BLive isomorphic twin screen and holdout on authored familiesExecuted (diagnostic)
• E41-CLive screen on real SEC exhibits chosen for surface trapsExecuted
• E40-AExact composition topology enumerationExecuted (diagnostic)
• E40-BLive six-topology benchmark on authored packetsExecuted (diagnostic)
• E40-CLive six-topology benchmark on real SEC exhibitsExecuted (bounded diagnostic)
• E39-AConflicting collateral evidence state-space diagnostic and lens composition analysisExecuted (diagnostic)
• E39-BLive 4-witness conflicting-evidence benchmark under two evidentiary lensesExecuted (diagnostic)
• E39-CReal SEC EDGAR exhibit track: conditional versus unconditional effectivenessExecuted (bounded diagnostic)
• E38-APost-register-closure operative-validity state-space diagnostic and resolver comparisonExecuted (diagnostic)
• E38-BLive 4-architecture operative-validity benchmark (View R vs View V)Executed (diagnostic)
• E37-AContract-lifecycle evidence state-space diagnostic and resolver consistency checkExecuted (diagnostic)
• E37-BLive 4-architecture contract-lifecycle benchmark (base-only vs full lifecycle)Executed (diagnostic)
• E37-CLive stale-register probeExecuted (diagnostic)
• E36-AMulti-document reference-graph and precedence-DAG state-space diagnosticExecuted (diagnostic)
• E36-BLive matched-compute 4-architecture multi-document contract-audit benchmarkExecuted (diagnostic)
• E35-APost-GTP-34 correlated-inspection state-space diagnosticExecuted (diagnostic)
• E35-BLive matched-compute multi-model contract-audit benchmarkExecuted (diagnostic)
• E34-ARecursive endorsement graph enumerationExecuted (diagnostic)
• E34-BExternal bootstrap and semantic handoffProposed
• E33-AThree-root external pilot and matched-compute AIC-33 experimentProposed
• E31-CBoundary hardening and blinded mutation holdoutProposed
• E30-BProposed matched-compute shadow trial of the RISE GateProposed
• E29-AExact finite deterministic interventional-lineage diagnosticExecuted (diagnostic)
• E29-BProposed matched-cost live Interventional Lineage TestProposed
• E28-AExact fixed-compute constructed trace replayExecuted (diagnostic)
• E27-AExact finite constructed diagnostic of budgeted verification policiesExecuted (diagnostic)
• E27-BFixed-compute shadow comparison of model-set adaptive verificationProposed
• E26-ADeterministic finite-model identifiability analysisExecuted (diagnostic)
• E26-BProspective fixed-budget observation-sufficiency shadow experimentProposed
• D25-AAdversarial-omission authored conformance diagnostic (D25-A)Executed (bounded diagnostic)
• E25-BProspective fixed-budget multi-authority silence-coverage trial with hidden independent adjudicationProposed
• D24-AAuthority coverage audit and authored conformance diagnostic (D24-A)Executed (bounded diagnostic)
• E24-BProspective multi-authority authority-lineage revalidation experiment with independent adjudicationProposed
• D23-AProspective shadow trial of claim revalidation (D23-A)Executed (bounded shadow diagnostic)
• E23-BProspective multi-release CTAG shadow evaluation with independent adjudicationProposed
• D22-ADeterministic code-change routing diagnostic (D22-A)Executed (diagnostic)
• E22-BProspective CTAG evaluation on unseen Encevra changes at controlled verification computeProposed
• E21-ADeterministic attempt-bound measurement gate diagnostic (D21-A)Executed (diagnostic)
• E21-BShadow ABMG evaluation on real external-service traffic at fixed model computeProposed
• E20-ADeterministic minimal-disclosure receipt development harnessExecuted (diagnostic)
• E20-BHeld-out receipt comparison against strong generic tracesProposed
• E19-ADeterministic evidence-lifecycle validator and static collector auditExecuted (diagnostic)
• E19-BReconstruction under frozen trajectoriesProposed
• E18-AExact deterministic constructed bitemporal and repair diagnosticExecuted (diagnostic)
• E18-BFive-arm versioned evidence study (O0/A0/T0/B0/B1)Proposed
• E16-AExact finite-world sufficiency diagnosticsExecuted (diagnostic)
• E16-BBounded missing-condition studyProposed
• E14-AExact offline label-uncertainty diagnosticsExecuted (diagnostic)
• E14-BMasked-replay comparison of evaluation methodsProposed
• E13-AAudit of saved Noetis evaluation recordsExecuted (audit)
• E13-BLocal revision pilot with evidence receiptsProposed
• E12-AExact synthetic calibration diagnosticsExecuted (diagnostic)
• E12-BTerminal release calibration studyProposed
• E11-AMarginal-value verification controllerProposed

Architecture changes in Noetis

• A-03 Revision-drift guard: Flag synthesis that omits supported risks or adds untraced claims, rather than presenting it as consensus. (In Noetis (preview))
• A-02 Uncalibrated release-risk contract: Return a null risk score until version-bound, externally labelled calibration exists. (In Noetis (preview))
• A-01 Conservative role triage with full-panel fallback: Skip optional roles only on confident, valid routing; otherwise run every role. (In Noetis (preview))

Open questions

• Can condition level discharge semantics release cured items without weakening the universal schedule threshold test? (Round 47)
• Does independent extraction before brief exposure reduce reviewer deference to wrong resolver and direct reading outcomes? (Round 47)
• Does a supersession field (which record replaces or amends which) let a universal condition test hold the schedule threshold mistake without holding correctly read cures? (Round 46)
• Does a packet and brief reviewer stay correct when it did not write the ledger in the brief, and on a pool where readers make split condition mistakes in practice? (Round 46)
• Does a record set consistency test (no record set may use a record the same ledger marks FAILS) remove the unsafe releases on new families without new holds, and does it survive a requirement split into independent conditions? (Round 45)
• Does a disagreement between the resolver and a direct reading make a cheap review trigger, and does a weaker validator follow a misleading reviewer note on real exhibit ledgers so that the gate's catch rate can be measured? (Round 45)
• Does a ledger with record sets and a scoped exclusion field keep near miss items held and satisfied requirements released across four witnesses, and where are the resolver's fixed rules wrong while a witness reading the records together is right? (Round 44)
• Does a completeness record drive the FAIL on TC items, or would any added record do? (Round 44)
• Does an extraction ledger that records, for each cited record, whether it satisfies, fails, or is irrelevant to the condition hold near miss items under an open record, and does a second validator witness keep the rewriting arm safe on harder seeded defects? (Round 43)
• When the packet is declared an open or a closed record, do direct readers still read absence as negation, and once extraction readers can err unsafely, do the same witness's unsafe releases coincide across roles? (Round 42)
• On a pool built by construction, with undetermined fact items labelled HOLD and PASS and scored on held out twins, is the joint error of one witness in direct and extraction roles lower than that of two different witnesses in the same role? (Round 41)
• On a hard pool selected for witness disagreement, with externally reviewed HOLD labels, do disagreement routed topologies keep their safety without excessive review load, and do flag only validators avoid the errors that rewriting validators introduce? (Round 40)
• With a neutral instruction that offers HOLD and no evidentiary lens, and with expected verdicts audited outside the authoring process, do witnesses abstain on outcome relevant unresolved conflicts, and can a four-witness quorum on real SEC exhibits tie each conditioned amendment to the filer's later statement that it became effective? (Round 39)
• Can a balanced 2-2 assignment of precautionary and presumptive evidentiary lenses block unverified modifications through voting alone when the contract does not explicitly name DOC-VERIF, and how should a gate behave when two attached verification records conflict (CEV-39)? (Round 38)
• With register coverage and as-of resolution held fixed, can witnesses distinguish an executed instrument from an effective one when a waiver depends on an unmet condition precedent, a termination notice is defective, an amendment lacks authorized signature, or two registered amendments overlap within one clause, and does deliberate witness decorrelation produce informative split votes? (Round 37)
• Can Round 37 formulate and test Temporal Drift and Post-Execution Amendment Supersession (TDR-37) when a post-execution Amendment, Waiver, or Termination Notice modifies an earlier obligation silently without any forward pointer in the base agreement? (Round 36)
• 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? (Round 35)
• Can Round 35 execute an externally administered bootstrap-and-rotation intervention with hidden fixtures before a matched-compute model experiment, only after all critical readiness controls are present? (Round 34)
• Can Round 34 provision at least three witnesses with independently administered identities, key custody, and deployment controls, obtain one external binding per witness per critical layer, and run AIC-33 before action with hidden shared-root and stale-binding cases adjudicated by a separate authority? (Round 33)
• Can Round 33 establish separately administered witnesses and registry services, external history observation and real isolation before an independently authored holdout and any paired model-answer runner? (Round 32)
• Can Round 32 harden witness authority, arm isolation and external receipt registration, freeze a blinded mutation holdout and implement a paired matched-compute runner before any model pilot? (Round 31)
• Can Round 31 build a sandboxed vertical slice with root identity, a non-destructive root action, independent witness, separation and budget records, and equivalence-aware outcomes before piloting E30-B? (Round 30)
• Can Round 30 execute E29-B in shadow mode with independently verifiable root interventions, receipts, leakage prevention and matched total compute? If prerequisites fail, report the infrastructure gap. (Round 29)
• Can a fixed-cost active intervention discover a missing evidence dependency that passive confidence, disagreement, and surprise cannot detect? (Round 28)
• Does model-set adaptive verification reduce harmful false-clean rates at fixed compute with hidden independent adjudication in E27-B? (Round 27)
• Does OSC reduce harmful false-stability certifications at fixed total cost with hidden independent adjudication in E26-B? (Round 26)
• Does SACP reduce independently adjudicated affected-claim gap misses at fixed total budget in a preregistered prospective multi-authority trial (E25-B)? (Round 25)
• Does authority-lineage revalidation reduce stale-claim escape in a prospectively frozen, independently adjudicated, multi-authority, fixed-budget experiment (E24-B)? (Round 24)
• Does the hybrid reduce stale-claim escape across a preregistered multi-release window with independent adjudication (E23-B)? (Round 23)
• Does a prospective CTAG shadow gate reduce stale-claim escape under fixed verification budget on unseen real commits? (Round 22)
• Which independent dependency view adds unique catches? (Round 22)
• Can unknown execution paths be caught and historical claims kept immutable in a shadow slice? (Round 22)
• Does ABMG capture at least 99.5 percent of live attempts exactly? (Round 21)
• Does synchronous gate overhead meet budget at fixed model compute? (Round 21)
• Does failure labeling improve reviewer reconstruction, and does the privacy design withstand review? (Round 21)
• Does MDIR improve real audit accuracy at acceptable cost against strong generic traces with independent immutable export? (Round 20)
• Does an independent checkpoint recover detection after a privileged operator rewrite? (Round 20)
• Can reviewer burden and operational overhead stay within the frozen development-gate thresholds? (Round 20)
• Do exposure ledgers improve exact exposure reconstruction over final records and generic traces? (Round 19)
• Does claim lineage improve first-divergence localization without false attribution? (Round 19)
• Can hash, version and locator binding reduce captured sensitive content while preserving identity reconstruction? (Round 19)
• Do bitemporal fields beat deterministic newest-recorded assembly on historical and as-known questions at matched cost? (Round 18)
• Does dependency-directed repair improve downstream consequences without excessive false invalidation? (Round 18)
• Does an explicit disputed state reduce unsupported commitment without collapsing useful coverage? (Round 18)
• Does validated alternative generation improve supported task completion over a strong checklist at matched budget? (Round 16)
• Does scope discovery help when a relevant rule is absent from the representation? (Round 16)
• Can decision relevance reduce unsupported closure without needless questions or reduced useful completion? (Round 16)
• Does evaluation error differ meaningfully between ordinary and receipt-based revision outputs? (Round 14)
• Does a protected-audit protocol make more correct, informative promotion decisions than naive or disagreement-only evaluation at the same budget? (Round 14)
• Do valid joint constraints reduce evaluation uncertainty at useful cost on naturally occurring outputs? (Round 14)
• Does evidence packaging change outcomes at a fixed state? (Round 13)
• Does Noetis outperform a strong single agent at matched evidence access and measured cost? (Round 13)
• Do terminal evidence and revision features improve useful coverage at identical generation compute? (Round 12)
• Do revised and conflicted episodes need separate risk calibration? (Round 12)
• Do localized checks plus short bundles beat a whole-answer router? (Round 11)
• Can harmful revisions be bounded while spending more on checks? (Round 11)