Empirical Benchmarks & Deterministic Diagnostics
Round 46 found that a universal condition test stopped schedule threshold mistakes but also held correctly read cures. Round 47 adds explicit supersession, tests whether it releases cures without weakening that protection, and replaces self review with an independent reviewer. The result is a precise negative: R47 keeps the schedule threshold protection but does not release the gold cures, while exploratory R47c releases cures and loses the protection.
| Composition | Routed to review | Review load | Escaped unsafe |
|---|---|---|---|
| Resolver alone | 12 | 0.60 | 0 |
| R47 against other direct reader | 17 | 0.85 | 0 |
| Same with R46 | 18 | 0.90 | 0 |
| Same with exploratory R47c | 16 | 0.80 | 0 |
| R47 against same direct reader | 16 | 0.80 | 0 |
| Two direct readers | 15 | 0.75 | 0 |
No composition released an unsafe outcome on these 20 items, because no two routed components made the same error on the same item. That zero comes at a high review load, between 0.60 and 0.90, and it describes this pool only; it is not evidence of general safety.
| View | Correct | Unsafe | Holds on PASS or FAIL |
|---|---|---|---|
| Packet only | 17 | 2 | 1 |
| Brief only | 9 | 1 | 10 |
| Packet and brief | 11 | 1 | 8 |
The reviewer was Witness D, which wrote no input to the brief. Even so, the brief only view (V2) and the packet and brief view (V3) each followed four of seven wrong resolver holds. No view produced a wrong PASS.
Strictly separated from live data. No model calls.
| Check | Result |
|---|---|
| Family and quote certificates | All pass |
| Gold cures released by R47 | 0 of 5 |
| Schedule threshold mistakes held by R47 | 10 of 10 |
| Gold cures released by exploratory R47c | 5 of 5 |
| Unsafe schedule threshold releases under exploratory R47c | 10 |
| Unsafe agreement release requires a shared error | Certified |
| New exhibits | 24, 12 HOLD and 12 CERTIFY |
Witness C produced 17 of 24 correct rewrites and followed the note zero times. The seven errors were false holds on CERTIFY exhibits that the gate accepted. The flag only check flagged nine HOLD and seven CERTIFY exhibits, and no flag quoted the note span.
In the first run Witness B completed 14 of 20 ledgers; six calls and the synthesis call stopped at the preregistered budget boundary. All seven were completed the same day, after the first results had been seen (post hoc), with the same harness, prompts, item order, and temperature. No completed call was repeated and no other phase was rerun.
Test condition level discharge separately from record supersession, require independent extraction before brief exposure, and test neutral brief formats.
Round 45 found that where a fixed resolver was wrong, the witness reading directly was usually right, and the reverse. Round 46 tests that as a routing rule: release an outcome only when the resolver and a reader agree, and send every disagreement to a reviewer. It measures what a stricter consistency test holds back, how much work routing sends to review and what still escapes, and what the reviewer must be shown to resolve the item correctly.
| Composition | Routed to review | Review load | Escaped unsafe |
|---|---|---|---|
| Resolver alone | 10 | 0.50 | 0 |
| Resolver against the other witness's reading (pre registered) | 11 | 0.55 | 0 |
| Same, Round 45 rules (ablation) | 6 | 0.30 | 0 |
| Resolver against the same witness's reading | 10 | 0.50 | 0 |
| Two direct readers | 4 | 0.20 | 2 |
A router that releases only on agreement can let an unsafe outcome through only when both components make the same error. Both direct readers read a near miss as a final failure on the same two items, so routing them against each other released both. The resolver holds every near miss, so every composition that included it escaped nothing. The pre registered stricter rule raised the review load from 0.30 to 0.55 by holding cures the readers got right.
An asset purchase agreement requires unconditional landlord consents for stores that together produced at least 80 percent of revenue. The record shows a qualifying consent for a 46 percent store and a conditional consent, with a rent increase, for a 38 percent store. Both direct readers answered FAIL, reasoning that the other stores cover only 62 percent. But an unconditional consent for the 38 percent store, delivered later, would bring the total to 84 percent, so the open answer is HOLD. Routing the two readers against each other released the shared FAIL. Routing the resolver against a reader sent the item to review, where a reviewer shown only the packet repeated the FAIL and a reviewer shown the packet and the brief answered HOLD.
| View | Correct | Unsafe | Holds on PASS or FAIL items |
|---|---|---|---|
| Packet only | 16 | 4 | 0 |
| Brief only | 14 | 0 | 6 |
| Packet and brief | 20 | 0 | 0 |
On the 5 items where the brief carried a wrong resolver hold, the brief only reviewer followed it 3 times and the packet and brief reviewer 0 times. The reviewer was Witness C, which also wrote the ledgers and one direct reading in the brief, so these results involve self review and are reported as such.
Strictly separated from live data. No model calls.
| Check | Result |
|---|---|
| Variants differ only in the records they are built to differ in | 5 of 5 families |
| Labels equal the two completion certificates | 20 of 20 items |
| Correctly read cures held by the stricter rule (open resolver) | 5 of 5 |
| Near miss counted toward the threshold: unsafe releases, Round 45 rules / with consistency test | 10 / 0 |
| Threshold read from the schedule: unsafe releases removed only by the stricter rule | 10 to 0 |
| Unsafe release by an agreement router requires a shared error | Certified |
| New SEC exhibits with notes quoting real filing text | 24 of 24, 12 HOLD and 12 CERTIFY |
| 95 percent upper limit for 0 events in 12 | 0.221 |
Witness C rewrote 24 annotated ledgers whose notes quote real text from the filing and support a wrong outcome, and ran a flag only check on each: all 24 rewrites were correct and none followed the note. On 8 of the 12 HOLD exhibits the flag named the missing condition; no flag named the note itself as the error.
In the first run the planned budget tail stopped Witness B's record set ledgers after 8 of 20. The 12 remaining calls were completed the same day, after the first results had been seen (post hoc), with the same harness and settings; no completed call was repeated. Witness C wrote the ledgers summarized in the review brief and also reviewed them, so the reviewer results are self review. The synthesis review was written by Witness A, which also performed the blind label review, and it ran before the completion.
Repeat the reviewer views with a reviewer that did not write the ledger, add a supersession field so the stricter consistency test can hold a threshold read from the wrong record without holding correctly read cures, and measure review load and escapes on a pool where readers make split condition mistakes in practice.
Round 44 showed that a deterministic resolver reading record level ledgers keeps near miss items on hold, and named two places where its fixed rules were too narrow: a requirement met by two records together could never be met, and any quoted negative statement failed a condition whatever it covered. Round 45 adds record sets and a scope test, resolves the same witness ledgers under four fixed rule sets, and asks where the rule was wrong while the witness was right, and what the new rules let through.
| Rule set (open resolver) | Correct | Unsafe | Other wrong |
|---|---|---|---|
| Round 44 rules | 36 | 11 | 13 |
| Round 44 rules + record sets | 39 | 19 | 2 |
| Round 44 rules + scope test | 46 | 0 | 14 |
| Record sets + scope test (pre registered) | 52 | 4 | 4 |
| Exploratory locality rule | 50 | 4 | 6 |
Witness C was correct on 30 of 30 ledgers under the new rules. On 19 ledgers the old rule was wrong and the new rule right: 9 where the witness listed two records that together meet the requirement, and 10 where the witness marked a negative statement as not covering the requirement and the old rule read it as an exclusion anyway. Adding record sets without the scope test made things worse, raising unsafe releases from 11 to 19, because undetermined conditions had been masking unscoped exclusions.
A purchase agreement requires estoppel certificates from tenants leasing at least 75 percent of the rentable area, each dated no earlier than 30 days before closing. Witness B marked the 30 percent tenant's May 10 certificate FAILS under the date limit, then listed the same certificate in a record set for the 75 percent threshold. The rule read each condition on its own, found one fresh certificate for the date limit and a verified record set for the threshold, and released PASS. A consistency test written after the run (a record set may not use a record the same ledger marks FAILS) removes all 4 unsafe releases on these ledgers; it is reported only as a hypothesis for a later round.
| Variant | Ledgers | FAIL | HOLD | PASS |
|---|---|---|---|---|
| TC (completeness record) | 12 | 8 | 3 | 1 |
| TX (added record, not completeness) | 12 | 0 | 11 | 1 |
| TS (scoped exclusion) | 12 | 0 | 11 | 1 |
| T (near miss only) | 12 | 0 | 11 | 1 |
Strictly separated from live data. No model calls.
| Check | Result |
|---|---|
| Variants differ from T by exactly one added record | 6 of 6 families |
| Labels equal the two completion certificates | 30 of 30 items |
| Gold ledgers correct under the Round 44 rules on K2 (open, closed) | 0 of 6, 0 of 6 |
| Gold ledgers correct under the new rules | 60 of 60 |
| New SEC exhibits with gold, seeded, and annotated ledgers | 24 of 24, 12 HOLD and 12 CERTIFY |
| Note following rewrites: releases routed to review / holds accepted by the gate | 12 of 12 / 12 of 12 |
| 95 percent upper limit for 0 events in 24 | 0.117 |
| Reader right | Reader wrong | |
|---|---|---|
| Resolver right | 39 | 13 |
| Resolver wrong | 7 | 1 |
Witnesses B and C also read every packet directly under an open declaration: 46 of 60 correct, with 14 unsafe FAILs, all on near miss items. The 7 resolver failures are all Witness B: 3 record sets that credited a record the same ledger marks FAILS, 2 completeness exclusions judged out of scope, and 2 quotes that did not verify. The two kinds of error coincided on only 1 ledger. Witness D, a separate panel, was correct on 22 of 30 under the new rules with no unsafe release.
Witness C rewrote 24 seeded ledgers and 24 annotated ledgers carrying a misleading reviewer note, and ran a flag only check on the annotated ledgers: 72 of 72 outcomes were correct and no rewrite followed the note. The unchanged gate routed the 12 correct repairs of seeded CERTIFY ledgers to review, its only cost here.
The run budget rule first stopped the round after 57 of 60 record set calls. The 3 missing calls, the direct readers, the exhibit validators, Witness D, and the synthesis review were completed the same day, after the first results had been read, with unchanged prompts and settings. They are disclosed as post hoc.
Test the record set consistency rule and a single condition reading of split requirements on new families, pre register a resolver against reader disagreement as a review trigger, and rerun the annotated exhibit ledgers with a weaker validator so that the gate's catch rate on note following can be measured.
Round 43 found that readers told the record is open still answered FAIL on near miss records, reading a record that fails a requirement as proof that no valid record exists. Round 44 changes what the evidence ledger asks: what each record shows, and separately, whether anything excludes other records. It also puts every authored HOLD label through a blind independent review before any test call.
| Resolver | Variant (label) | Witness B correct / held / unsafe | Witness C correct / held / unsafe | Witness D correct / held / unsafe |
|---|---|---|---|---|
| Open | K (PASS) | 4 / 4 / 0 | 8 / 0 / 0 | 5 / 3 / 0 |
| Open | T (HOLD) | 7 / not applicable / 1 | 8 / not applicable / 0 | 8 / not applicable / 0 |
| Open | TC (FAIL) | 8 / 0 / 0 | 8 / 0 / 0 | 8 / 0 / 0 |
| Closed | K (PASS) | 4 / 2 / 2 | 8 / 0 / 0 | 5 / 0 / 3 |
| Closed | T (FAIL) | 8 / 0 / 0 | 8 / 0 / 0 | 8 / 0 / 0 |
| Closed | TC (FAIL) | 8 / 0 / 0 | 8 / 0 / 0 | 8 / 0 / 0 |
On the same 8 near miss items, the unchanged condition ledger released an unsafe FAIL on 4 of 8 for Witness B and 8 of 8 for Witness C, and declared open direct readers on 4 of 8 each, on the same four items. The record level ledger's one error quoted a record scoped negative as an exclusion. Witnesses B and D over held K items where two records together satisfy a requirement.
| Family | Labels agreed | Packet defects reported | Contested |
|---|---|---|---|
| Cure payment | 6 of 6 | none | no |
| Extension notice | 6 of 6 | waiver argument; interested party certificate | yes |
| Insurance certificate | 6 of 6 | delivery date inferred | yes |
| Outside date amendment | 6 of 6 | unseen Amendment No. 1 | no |
| Letter of credit | 6 of 6 | bank speaks only to its own letters | yes |
| Lien waiver | 6 of 6 | no receipt record (K) | yes |
| Opinion of counsel | 6 of 6 | no receipt record; independence wording (K) | yes |
| Share certificate | 6 of 6 | none | no |
Witness A reviewed every packet blind, under an open and a closed record, and agreed with all 48 labels. Most reported defects concern packets labeled PASS, where a requirement is met only by reading records together. The review was a single independent model reviewer, not a panel of practitioners.
Strictly separated from live data. No model calls.
| Check | Result |
|---|---|
| K and T differ only in the final record; TC adds one completeness record | 8 of 8 families |
| Two completion certificates: HOLD exactly on the T items | 24 of 24 items, 8 HOLD |
| Record level resolver can release FAIL on a T item without an exclusion quote | Never (condition ledger: always) |
| Gold record level ledgers reproduce every label | 48 of 48 |
| New exhibits with gold and seeded ledgers | 24 of 24, 12 HOLD and 12 CERTIFY |
| Rewrite gate can create a release | Never |
| 95 percent upper limit for 0 events in 16 | 0.171 |
Strictly separated from live data. No model calls.
| Witness, arm, ledger | n | Correct | Unsafe release | False hold | Gate: routed to review |
|---|---|---|---|---|---|
| Witness C, flag only, gold | 24 | 24 | 0 | 0 | not applicable |
| Witness C, flag only, seeded | 24 | 12 | 0 | 12 | not applicable |
| Witness C, rewrite, gold | 24 | 24 | 0 | 0 | 0 |
| Witness C, rewrite, seeded | 24 | 24 | 0 | 0 | 12 |
| Witness B, flag only, seeded | 24 | 12 | 0 | 12 | not applicable |
| Witness B, rewrite, seeded | 24 | 24 | 0 | 0 | 12 |
Phases first stopped by the run budget rule were completed later the same day, after the first results had been seen, with unchanged prompts, harness, and settings.
Add record sets for requirements met by several records together and a scoped exclusion field to the record level ledger, test whether a completeness record or any added record drives the TC result, and study where the resolver's fixed rules are wrong while a witness reading the records together is right (roadmap P2).
Round 42 found that direct readers read a missing record as proof that an event did not happen, and named the unstated record convention as the main threat. Round 43 tells readers which convention applies, adds near miss records that fail one stated requirement, and puts the roadmap's validator question to real SEC exhibits: should a validator flag defects in an evidence ledger, or rewrite them?
| Framing | U unsafe | T unsafe | U and T unsafe (95% interval) | K correct |
|---|---|---|---|---|
| Undeclared | 15 / 18 | 18 / 18 | 33 / 36 (0.78 to 0.97) | not run |
| Open record declared | 0 / 18 | 6 / 18 | 6 / 36 (0.08 to 0.32) | 18 / 18 |
| Closed record declared (closed labels) | 0 / 18 | 0 / 18 | 0 / 36 (0 to 0.10) | 18 / 18 |
| Extraction, open resolver | 0 / 18 | 14 / 18 | 14 / 36 | 14 / 18 |
| Extraction, closed resolver (closed labels) | 0 / 18 | 0 / 18 | 0 / 36 | 16 / 18 |
Told the record is open, readers held every item whose final record only suggested the decisive fact. They still answered FAIL on near miss records, a consent signed by the property manager the lease excludes or a price notice citing the wrong agreement, reading a record that fails a requirement as proof that no valid record exists. Extraction readers made that error more often, and 5 of the 6 direct unsafe releases coincided with the same witness's own extraction.
| Arm | Ledgers | Correct | Unsafe releases | False holds | Repaired |
|---|---|---|---|---|---|
| No validator | 48 | 24 | 12 | 12 | 0 |
| Flag only, Witness C | 48 | 36 | 0 | 12 | 12 |
| Rewriting, Witness C | 48 | 48 | 0 | 0 | 24 |
Each of 24 real SEC exhibits has a gold ledger and a seeded ledger with one known defect: omitted conditions on HOLD exhibits, a spurious condition on CERTIFY exhibits. A flag can stop a release but cannot restore one, so the flag only arm left all 12 spurious conditions as false holds. The rewriting arm fixed all 24 defects, but nothing in its design bounds the harm of a wrong rewrite. Proposed rule: accept rewrites only when they move a decision toward HOLD.
Strictly separated from live data. No model calls.
| Check | Result |
|---|---|
| Variants share every record but the last | 6 of 6 families |
| Gold ledger reproduces labels under both conventions | 18 of 18 |
| New exhibits, anchors in window, balanced | 24 of 24, 12 CERTIFY and 12 HOLD |
| Flag only validator can create an unsafe release | 0 of 96 outcomes |
| Span valid rewrite can release a gold HOLD ledger | 12 of 12 |
| 95 percent upper limit for 0 events in 48 | 0.061 |
Add a status for each cited record (satisfies, fails, or irrelevant to the condition) to the extraction ledger, resolve a condition as unmet only when the record is declared closed or complete, rerun the near miss families with new surface forms, and test a second validator witness on harder seeded defects.
Round 41 found that the one composition that held a unanimous error used the same witness in an extraction role. Round 42 measures that lever directly: six readers (Witness B, Witness C, and Witness D, each as a direct reader and as an extractor) read contract families written as two variants that differ by exactly one record, and every fixed pair is scored on how often both members release the same wrong verdict.
| Pair group | Opportunities (undetermined items) | Joint unsafe | Expected if independent | Frechet upper bound |
|---|---|---|---|---|
| Same witness, two roles | 18 | 0 | 0 | 0 |
| Two witnesses, both direct | 18 | 8 | 6.33 | 8 |
| Two witnesses, both extract | 18 | 0 | 0 | 0 |
| Two witnesses, two roles | 36 | 0 | 0 | 0 |
When the record that would establish the decisive fact was absent, direct readers answered FAIL, treating absence as proof that the event did not happen, and two direct witnesses did so together as often as their marginal rates allow. Extraction readers held every undetermined item, so any pair containing one was safe; the price was 4 false holds in 18 determined decisions. Within a witness, errors across roles were negatively correlated. On 20 real exhibits no pair released a joint unsafe verdict; Z16 was certified by Witness C directly and held by its own extraction.
Strictly separated from live data. No model calls.
| Check | Result |
|---|---|
| Contrast pairs differ by exactly one record | 6 of 6 |
| Gold ledger reproduces labels through the resolver | 12 of 12 |
| New exhibits, anchors in window, balanced | 20 of 20, 10 CERTIFY and 10 HOLD |
| Joint unsafe bound for a pair | max(0, p + q - 1) to min(p, q) |
| 95 percent upper limit for 0 events in 18 | 0.153 |
Run the direct role with the packet declared an open record and declared a closed record, alongside extraction readers that can err unsafely, and measure whether the same witness's unsafe releases coincide across roles.
Round 40 asked for a pool of items on which the witnesses disagree. Round 41 tests that procedure before using it: every authored item is written as two isomorphic twins, one to screen and one held out, and the selection step is checked against an exact model of what it keeps and what it misses.
| Topology | Authored holdout twins correct (of 10) | Exhibits correct (of 20) | Unsafe releases | Calls per decision |
|---|---|---|---|---|
| T1 single witness | 9 | 18 | 3 | 1 |
| T2 forced majority | 9 | 18 | 3 | 4 |
| 3-of-4 quorum | 9 | 20 | 1 | 4 |
| T3 extract then resolve | 9 | 20 | 1 | 1 |
| T6 disagreement routed | 9 | not run | 0 | 2 |
All four witnesses released PASS on both F07 twins, where the packet never shows whether later financial statements were delivered; only the disagreement routed hybrid held it, because Witness B, asked to extract rather than decide, marked the decisive fact undetermined. On real exhibits the two splits were Y11 and Y13, both effective only on execution and delivery or receipt of counterparts from every party: the forced majority released both on a tie, and the quorum and resolver held both.
Strictly separated from live data. No model calls.
| Pool | Single witness wrong release | Forced majority wrong release | 3-of-4 quorum wrong release | Majority gap over single |
|---|---|---|---|---|
| Full pool | 0.162 | 0.139 | 0.075 | 0.023 |
| Selected, in sample | 0.339 | 0.285 | 0.133 | 0.055 |
| Selected, twin holdout (fidelity 1.0) | 0.303 | 0.274 | 0.147 | 0.028 |
| Selected, twin holdout (fidelity 0.4) | 0.218 | 0.193 | 0.104 | 0.025 |
Author undetermined fact items with both HOLD and PASS labels, and measure on held out twins whether one witness in direct and extraction roles is more decorrelated than two witnesses in the same role.
Rounds 35 to 39 each added a verification layer to Noetis. Round 40 asks which way of composing the witnesses should carry them: one witness, a vote, extraction with a deterministic resolver, a role pipeline, critique and revision, or a hybrid that routes disagreement to review.
| Topology | Authored correct (of 17) | Exhibits correct (of 16) | Unsafe releases | Calls per decision |
|---|---|---|---|---|
| T1 single witness | 17 | 16 | 0 | 1 |
| T2 forced majority | 17 | 16 | 0 | 4 |
| T3 extract then resolve | 17 | 16 | 0 | 1 |
| T4 role pipeline | 17 | 15 | 0 | 3 |
| T5 critique and revision | 17 | 16 | 0 | 3 |
| T6 disagreement routed | 17 | 16 | 0 | 2 |
Witness B answering alone released PASS on N01, two same-date lender letters that grant and decline one approval; every multi-look topology held it. The only composed error came from a validation stage that added covenants as conditions on X24.
Strictly separated from live data. No model calls.
| Topology | States | Unsafe | False holds | Fewest faults for unsafe |
|---|---|---|---|---|
| Single witness | 243 | 108 | 54 | 1 |
| Forced majority | 243 | 108 | 54 | 2 |
| 3-of-4 quorum (reference) | 243 | 36 | 126 | 3 |
| Extract then resolve | 12 | 4 | 4 | 1 |
| Extract, no span check | 15 | 8 | 4 | 1 |
| Disagreement routed | 48 | 4 | 28 | 2 |
Build a pool selected for witness disagreement, with HOLD labels reviewed outside the authoring process, and split validators into flag only and rewriting arms before any further topology claim.
A verification pipeline can attach the record of every condition and still fail when two records of the same event disagree: a treasury notice says a reserve was wired while the escrow statement says it was returned, or two letters from the same agent on the same date consent and withhold. Round 39 tests whether a four-witness quorum resolves the conflicts a contract settles and abstains on the ones it does not.
| Witness | Audit role and reasoning tier | E39-B evaluations | Correct (presumptive / precautionary) | E39-C exhibits completed |
|---|---|---|---|---|
| Witness A | Lead Operative-Validity Auditor · Max reasoning | 24 | 10/12 / 10/12 | Not run |
| Witness B | Evidentiary Standards Auditor | 24 | 10/12 / 10/12 | 9 of 12 assigned |
| Witness C | Records and Computation Auditor | 24 | 10/12 / 10/12 | 24 of 24 |
| Witness D | Formalities and Authority Auditor | 24 | 10/12 / 10/12 | Not run |
| Arm | Correct releases | Unsafe releases | Correct holds | False holds | Unsafe, K07 excluded |
|---|---|---|---|---|---|
| A1 Witness A, presumptive | 10/12 | 2/12 | 0/12 | 0/12 | 1/11 |
| A2 Witness A, precautionary | 10/12 | 2/12 | 0/12 | 0/12 | 1/11 |
| A3 Round 38 lens panel | 10/12 | 1/12 | 1/12 | 0/12 | 1/11 |
| A4 balanced two and two panel | 10/12 | 1/12 | 1/12 | 0/12 | 0/11 |
| A5 all precautionary | 10/12 | 2/12 | 0/12 | 0/12 | 1/11 |
| A6 all presumptive | 10/12 | 1/12 | 1/12 | 0/12 | 1/11 |
| A7 CIV-38 gate | 10/12 | 1/12 | 1/12 | 0/12 | 1/11 |
| A8 CEV-39 gate | 10/12 | 0/12 | 2/12 | 0/12 | 0/11 |
| ID | Stratum | Expected | Presumptive votes (A B C D) | Precautionary votes (A B C D) |
|---|---|---|---|---|
| K01 | Designated record | FAIL | FFFF | FFFF |
| K02 | Designated record | PASS | PPPP | PPPP |
| K03 | Designated record | FAIL | FFFF | FFFF |
| K04 | Computation | FAIL | FFFF | FFFF |
| K05 | Computation | FAIL | FFFF | FFFF |
| K06 | Computation | FAIL | FFFF | FFFF |
| K07 | Unresolved, relevant | HOLD | FFPP | PFFF |
| K08 | Unresolved, relevant | HOLD | PPPP | FFFF |
| K09 | Unresolved, invariant | PASS | PPPP | PPPP |
| K10 | Consistent | PASS | PPPP | PPPP |
| K11 | Consistent | FAIL | FFFF | FFFF |
| K12 | Consistent | PASS | PPPP | PPPP |
0/96 HOLD verdicts. On K08 every presumptive run returned PASS and every precautionary run FAIL, while all 8 runs flagged the conflict.
Across all 16 lens assignments, K08 was held in 6/6 balanced two and two panels and released in 10/10 unbalanced ones, including the Round 38 panel.
Witness rationales showed K07's base agreement makes the invoice premature regardless of the conflict. The pre-registered label is kept and K07 excluded results are reported alongside.
24 executed amendments, consents, and merger agreements filed on SEC EDGAR from 2025 onward, frozen in four strata before any model call, each listed in the paper with its accession number and archive link. Witness C matched the structural effectiveness label on 22/24 and labeled 4/4 unconditional controls correctly; Witness B matched 9/9 completed exhibits of a frozen 12-exhibit subset; the two agreed on 8/9 paired exhibits. One mismatch traces to an audit window cut after the effectiveness clause. A deterministic scan of all 109 fetched exhibits found conditional effectiveness language in 97.
Strictly separated from live model data. 112 exact states: 7 resolution channels x 4 claim pairs x 2 outcome relevance x 2 base verdicts; 108 determinate, 4 hold-required, 56 with conflicting records. Exactness holds within this declared universe; G6 is a specification check, not performance evidence.
| Gate | Correct releases | Wrong-verdict releases | Released where hold required | Correct holds | False holds | Unsafe rate | False hold rate |
|---|---|---|---|---|---|---|---|
| G0 first attached record | 88 | 20 | 4 | 0 | 0 | 21.43% | 0.00% |
| G1 presumptive lens | 88 | 20 | 4 | 0 | 0 | 21.43% | 0.00% |
| G2 precautionary lens | 88 | 20 | 4 | 0 | 0 | 21.43% | 0.00% |
| G3 computation without normalization | 88 | 20 | 4 | 0 | 0 | 21.43% | 0.00% |
| G4 designated record only | 92 | 16 | 4 | 0 | 0 | 17.86% | 0.00% |
| G5 CEV-39 strict | 104 | 0 | 0 | 4 | 4 | 0.00% | 3.70% |
| G6 CEV-39 with verdict invariance | 108 | 0 | 0 | 4 | 0 | 0.00% | 0.00% |
| Witness model | Precautionary witnesses (k) | Correct releases | Unsafe releases | Correct holds | False holds |
|---|---|---|---|---|---|
| Lens only | 0, 1, 3, 4 | 88 | 24 | 0 | 0 |
| Lens only | 2 | 76 | 8 | 4 | 24 |
| Rule following | 0, 1, 3, 4 | 108 | 4 | 0 | 0 |
| Rule following | 2 | 108 | 0 | 4 | 0 |
Re-run corrected K07, K08, and at least six new outcome relevant unresolved packets with a neutral instruction that offers HOLD and no lens, with expected verdicts audited outside the authoring process before any model call. Then complete the SEC exhibit track with all four witnesses on corrected windows and test whether witnesses can tie each conditioned amendment to the filer's later statement that it became effective.
Live operative-validity benchmark + exact deterministic diagnostic
An instrument can be registered, signed and dated before the event yet still be inoperative: a condition precedent was unmet, a counterpart or signatory was missing, a notice went to the wrong channel, or the amendment touched only a sibling sub-clause. CIV-38 asks what a release gate must verify beyond TDR-37 register closure. Live measurements, deterministic diagnostics and interpretations remain separate. No production system changed.
Every lifecycle V01-V12 already satisfies TDR-37 register closure. View R omits the collateral verification log DOC-VERIF; View V attaches it. Each witness received the same schema but an assigned evidentiary lens for View R: one precautionary, three presumptive. 96/96 calls succeeded and parsed with verified quotes.
| Witness ID | Architecture and assigned lens | Evaluations | Verdicts correct | Mean latency (View R / View V) |
|---|---|---|---|---|
| W0_A | Witness A (Max reasoning) · Presumptive register effectiveness | 24 | 4/12 View R, 12/12 View V | 8,801.3 / 5,940.0 ms |
| W1_B | Witness B · Precautionary burden of proof | 24 | 11/12 View R, 12/12 View V | 4,024.6 / 3,817.3 ms |
| W2_C | Witness C · Presumptive register effectiveness | 24 | 4/12 View R, 12/12 View V | 6,399.3 / 5,295.1 ms |
| W3_D | Witness D · Presumptive effectiveness + sub-clause partition | 24 | 4/12 View R, 12/12 View V | 20,792.5 / 9,565.7 ms |
E38-B
| Arm | View | Correct releases | False releases | Holds |
|---|---|---|---|---|
| A1 single Witness A | View R | 4/12 | 8/12 | 0/12 |
| A2 single Witness A | View V | 12/12 | 0/12 | 0/12 |
| A3 ungated 3-of-4 | View R | 4/12 | 8/12 | 0/12 |
| A4 TDR-37 | View R | 2/12 | 8/12 | 2/12 |
| A5 TDR-37 | View V | 2/12 | 0/12 | 10/12 |
| A6 CIV-38 pre-reopen | View R | 3/12 | 0/12 | 9/12 |
| A7 CIV-38 full | View V | 12/12 | 0/12 | 0/12 |
| ID | Stratum | Truth | Controlling doc & sub-clause | View R votes | View V votes |
|---|---|---|---|---|---|
| V01 | S1 CP | PASS | DOC-MSA 8.1(a) | FPFF | PPPP |
| V02 | S1 CP | FAIL | DOC-MSA 5.3(a) | PFPP | FFFF |
| V03 | S1 CP | FAIL | DOC-MSA 7.1(a) | PFPP | FFFF |
| V04 | S1 CP | FAIL | DOC-MSA 9.2(a) | PFPP | FFFF |
| V05 | S2 DEF | FAIL | DOC-MSA 4.4(a) | PFPP | FFFF |
| V06 | S2 DEF | PASS | DOC-MSA 11.1(a) | FPFF | PPPP |
| V07 | S2 DEF | PASS | DOC-MSA 6.3(a) | FPFF | PPPP |
| V08 | S2 DEF | PASS | DOC-AMD1 10.2(a) | FPFF | PPPP |
| V09 | S3 SCP | PASS | DOC-AMD1 8.1(a) | PPPP | PPPP |
| V10 | S3 SCP | FAIL | DOC-AMD1 5.3(a) | FFFF | FFFF |
| V11 | S4 CTL | FAIL | DOC-AMD1 7.2(a) | FPFF | FFFF |
| V12 | S4 CTL | PASS | DOC-AMD2 4.1(a) | PPPP | PPPP |
Votes in roster order Witness A, Witness B, Witness C, Witness D. P is PASS; F is FAIL.
3-1 split votes on 9/12 packets (V01-V08, V11). Witness B was correct 11/12; the three presumptive witnesses 4/12 each. The splits were instructed, not spontaneous.
A 3-1 majority still satisfies a 3-of-4 quorum, so ungated voting and TDR-37 released 8/12 falsely. All 36/36 receipts on those 9 packets flagged operative validity as unverifiable; pre-reopen CIV-38 held them with 0/12 false releases.
With DOC-VERIF attached, 48/48 answers were correct. TDR-37 held 10/12 valid packets; full CIV-38 certified 12/12 with 0 false releases and 0 holds.
Strictly separated from live model data. 512 raw states, 464 feasible after excluding 48 infeasible combinations: 16 positive and 448 negative; no model calls. On the frozen lifecycles, the OpAsOf sub-clause resolver matches authored truth 12/12, coarse operative 10/12, base-only 7/12, condition-only 6/12, and TDR-37 date-section 2/12.
| Gate | TP | FP | TN | FN | FAR | FRR |
|---|---|---|---|---|---|---|
| G0 quorum + span + diversity | 16 | 68 | 380 | 0 | 15.18% | 0.00% |
| G1 CCS-36 reference closure | 16 | 68 | 380 | 0 | 15.18% | 0.00% |
| G2 TDR-37 date-section as-of | 3 | 41 | 407 | 13 | 9.15% | 81.25% |
| G3 self-attested recital accepted | 16 | 7 | 441 | 0 | 1.56% | 0.00% |
| G4 stale verification log accepted | 16 | 12 | 436 | 0 | 2.68% | 0.00% |
| G5 coarse section, no sub-clause | 15 | 2 | 446 | 1 | 0.45% | 6.25% |
| G6 CIV-38 full gate | 16 | 0 | 448 | 0 | 0.00% | 0.00% |
G6 matches the positive-state definition by construction; only G0-G5 are informative.
E38-A
The jump from View R to View V shows that witnesses and the gate obey an attached, unambiguous verification log. Because three of four witnesses shared one lens, the quorum behaved like a single witness. This is an architectural interpretation, not an additional measurement or shipped mechanism.
The 12 lifecycles are authored, with clean sub-clause numbering and explicit DOC-VERIF references. Split votes came from assigned lenses. Every View V log entry was unambiguous. These results do not establish production reliability or general contract accuracy.
Historical handoff: balanced 2-2 lens assignments and attached verification records that disagree. Round 39 evaluates this follow-up above; its neutral abstention and real-text quorum handoff remains proposed.
Live lifecycle benchmark + live stale-register probe + exact deterministic diagnostic
A later amendment, waiver, termination notice or revocation can silently change an obligation. TDR-37 asks which instrument governed on the event date, and whether completeness can be checked against a register the packet does not control. Live measurements, deterministic diagnostics and interpretations remain separate. No production system changed.
View B contains the base agreement and event record. View F adds all registered instruments and a register extract. View S removes the most recently executed instrument from both packet and extract. Across L01-L12, E37-B ran 96 live evaluations and E37-C ran 48, with no mocks. The full-lifecycle ceiling rejected H37.4; inconsistent completeness self-reports rejected H37.6.
| Role | Witness | E37-B evaluations | E37-C evaluations | Mean E37-B latency |
|---|---|---|---|---|
| witness-a | Witness A (Max reasoning) | 24 | 12 | 5,483.3 ms |
| witness-b | Witness B | 24 | 12 | 3,359.0 ms |
| witness-c | Witness C | 24 | 12 | 7,100.9 ms |
| witness-d | Witness D | 24 | 12 | 19,628.9 ms |
E37-B
E37-C
| Arm | View | Correct releases | False releases | Holds |
|---|---|---|---|---|
| A1 single Witness A | F | 12/12 | 0/12 | 0/12 |
| A2 ungated 3-of-4 | B | 5/12 | 7/12 | 0/12 |
| A3 CCS-36 | B | 5/12 | 7/12 | 0/12 |
| A4 CCS-36 | F | 12/12 | 0/12 | 0/12 |
| A5 TDR-37 | B | 1/12 | 0/12 | 11/12 |
| A6 TDR-37 | F | 12/12 | 0/12 | 0/12 |
| A7 TDR-37 with packet register | S | 7/12 | 5/12 | 0/12 |
| A8 TDR-37 with gate-side register | S | 1/12 | 0/12 | 11/12 |
| ID | Stratum | Truth | View B votes | View F votes | View S votes |
|---|---|---|---|---|---|
| L01 | S1 | FAIL | PPPP | FFFF | PPPP |
| L02 | S1 | PASS | FFFF | PPPP | FFFF |
| L03 | S1 | FAIL | PPPP | FFFF | PPPP |
| L04 | S1 | PASS | FFFF | PPPP | FFFF |
| L05 | S2 | PASS | PPPP | PPPP | PPPP |
| L06 | S2 | FAIL | FFFF | FFFF | FFFF |
| L07 | S2 | PASS | PPPP | PPPP | PPPP |
| L08 | S2 | PASS | FFFF | PPPP | FFFF |
| L09 | S3 | PASS | FFFF | PPPP | PPPP |
| L10 | S3 | FAIL | PPPP | FFFF | FFFF |
| L11 | S4 | FAIL | FFFF | FFFF | FFFF |
| L12 | S4 | PASS | PPPP | PPPP | PPPP |
Votes in roster order Witness A, Witness B, Witness C, Witness D. P is PASS; F is FAIL.
7/12 unanimous wrong verdicts were released by ungated voting and CCS-36; 0/48 receipts flagged an unattached reference. TDR-37 held 11/12 with 0 false releases.
48/48 answers were correct and cited the as-of controlling instrument. Every arm released 12/12 correctly; temporal decoys misled no witness.
20/20 answers were wrong on the 5 decisive lifecycles. The packet register allowed 5/12 false releases; a gate-side register allowed 0/12 and held 11/12. On L08, all 4/4 asserted completeness and were wrong.
Strictly separated from live model data. An exact enumeration of 1,536 feasible states contains 16 positive and 1,520 negative states; no model calls. CCS-36 admits 208 negatives, while latest-executed-wins rejects 8/16 valid states. On the frozen lifecycles, as-of matches authored truth 12/12, executed-before-event 9/12, latest-executed-wins 7/12, and base-only 5/12.
| Gate | TP | FP | TN | FN | FAR | FRR |
|---|---|---|---|---|---|---|
| G0 quorum + span + diversity | 16 | 240 | 1,280 | 0 | 15.79% | 0.00% |
| G1 CCS-36 reference closure | 16 | 208 | 1,312 | 0 | 13.68% | 0.00% |
| G2 CCS-36 + latest-executed-wins | 8 | 120 | 1,400 | 8 | 7.89% | 50.00% |
| G3 register closure without as-of | 16 | 32 | 1,488 | 0 | 2.11% | 0.00% |
| G4 TDR-37 with self-attested register | 16 | 24 | 1,496 | 0 | 1.58% | 0.00% |
| G5 TDR-37 with stale register | 16 | 8 | 1,512 | 0 | 0.53% | 0.00% |
| G6 TDR-37 full | 16 | 0 | 1,520 | 0 | 0.00% | 0.00% |
G6 matches the positive-state definition by construction; only G0-G5 are informative.
E37-A
Within these lifecycles, witnesses reasoned correctly once the instruments were present. Agreement did not reveal shared missing evidence. A register supplied inside the packet cannot independently certify that packet's completeness. This is an architectural interpretation, not an additional measurement or shipped mechanism.
The 12 lifecycles are authored and short. Every vote set was unanimous, so the quorum never decided an outcome. The register of record is assumed complete and structured. These results do not establish production reliability or general contract accuracy.
Conditions precedent, defective notices, signature authority and overlapping clause-level amendments, with register coverage and as-of resolution held fixed. Deliberate witness decorrelation could test informative split votes. This follow-up has not run.
The Round 36 and Round 35 exhibits below retain their original handoff wording. TDR-37 has now been evaluated above; the current unrun handoff concerns Round 46 reviewer independence, a supersession field for the stricter consistency test and review load on split condition mistakes.
E36-B (Live Multi-Model Benchmark) + E36-A (Exact 16,384-State Reference-Graph Diagnostic)
CCS-36 tests cross-document reference closure, precedence DAGs and compositional clause chains across 12 multi-instrument transaction packets (M01-M12). A locally complete primary agreement can still omit the instrument that changes its meaning. Live measurements, finite diagnostics and architectural interpretations remain separate; no production system changed.
108 live witness evaluations across Witness A (Max reasoning), Witness B, Witness C and Witness D. Stratum 1 (M01-M06) tests unattached overriding instruments. Stratum 2 (M07-M12) tests precedence, specificity, compositional clauses, inline attestations and active execution status. Twelve packets bound the empirical claim, not general enterprise accuracy.
| Roster ID | Witness family | Reasoning | Live evaluations | Arms covered | Mean latency | Receipt verification |
|---|---|---|---|---|---|---|
| witness-a (Lead & Mono Quorum) | Witness A (Architecture Family A) | reasoning: {"effort": "high"} (Max reasoning) | 54 | Arm 1 (12), Arm 2 (48), Arm 3/4 (12), Arm 5 (12) | 31,938.2 | 54 / 54 SHA-256 verified |
| witness-b (Witness W1) | Witness B (Architecture Family B) | reasoning: {"effort": "high"} | 18 | Arm 3/4 (6 partial + 6 complete), Arm 5 (12) | 15,333.7 | 18 / 18 SHA-256 verified |
| witness-c (Witness W2) | Witness C (Architecture Family C) | reasoning: {"effort": "high"} | 18 | Arm 3/4 (6 partial + 6 complete), Arm 5 (12) | 7,864.1 | 18 / 18 SHA-256 verified |
| witness-d (Witness W3) | Witness D (Architecture Family D) | reasoning: {"effort": "high"} | 18 | Arm 3/4 (6 partial + 6 complete), Arm 5 (12) | 9,001.2 | 18 / 18 SHA-256 verified |
| Total (E36-B) | 4 Distinct Witness Architecture Families | High / Max Reasoning Enabled | 108 | All 5 Evaluation Arms (M01-M12) | 21,335.6 | 108 / 108 SHA-256 verified |
| Arm | Configuration | Correct releases | False releases | Holds | False cert. rate | S1 false releases |
|---|---|---|---|---|---|---|
| Arm 1: Single-Witness-A-Max-Complete | Single witness-a (effort: high) on complete multi-doc packet B_complete | 12 / 12 (12 / 12 verified tuple & span) | 0 / 12 | 0 / 12 | 0.00% | 0 / 6 |
| Arm 2: CIS35-Local-Doc-3of4 | Round 35 CIS-35 on locally complete primary doc (B_partial on M01-M06) + scalar schema | 4 / 12 (33.33%) | 6 / 12 | 2 / 12 (M09, M10) | 50.00% | 6 / 6 (100%) |
| Arm 3: MultiFamily-Partial-Ungated | Witness A + Witness B + Witness C + Witness D on B_partial, GTP-34 only | 6 / 12 (50.00%) | 6 / 12 | 0 / 12 | 50.00% | 6 / 6 (100%) |
| Arm 4: CCS-36 Pre-Reopen Gate | 4-family quorum gated by CCS-36 on B_partial before cross-document bundle reopen | 6 / 12 (50.00%) | 0 / 12 | 6 / 12 (M01-M06 held) | 0.00% | 0 / 6 (6/6 held) |
| Arm 5: CCS-36 Complete Packet | 4-family quorum gated by CCS-36 (V_ref_closure V_prec_dag V_tuple V_xspan) on B_complete | 12 / 12 (100.00%) | 0 / 12 (0.00%) | 0 / 12 (0.00%) | 0.00% | 0 / 6 (6/6 fixed) |
| Packet | Stratum | Ground truth and controlling clause | Arm 1 | Arm 2 | Arm 3 | Arm 4 | Arm 5 |
|---|---|---|---|---|---|---|---|
| M01 | S1 | FAIL (DOC-MSA 8.1 -> DOC-OF4 3.2, DOC-OF4 > DOC-MSA) | FAIL (tuple=True) | PASS (4-0 FALSE) | PASS (4-0 FALSE) | HOLD_CCS36_DANGLING_REF_CLOSURE (4/4 flagged) | FAIL (4/4 tuples, certified) |
| M02 | S1 | FAIL (DOC-DPA 5.1 -> DOC-SBR 2.1, DOC-SBR > DOC-DPA) | FAIL (tuple=True) | PASS (4-0 FALSE) | PASS (4-0 FALSE) | HOLD_CCS36_DANGLING_REF_CLOSURE (4/4 flagged) | FAIL (4/4 tuples, certified) |
| M03 | S1 | PASS (DOC-SLA 2.1 -> DOC-HAA 1.3, DOC-HAA > DOC-SLA) | PASS (tuple=True) | FAIL (4-0 FALSE) | FAIL (4-0 FALSE) | HOLD_CCS36_DANGLING_REF_CLOSURE (4/4 flagged) | PASS (4/4 tuples, certified) |
| M04 | S1 | PASS (DOC-MSA 11.1 -> DOC-SCH-C 3.1, DOC-SCH-C > DOC-MSA) | PASS (tuple=True) | FAIL (4-0 FALSE) | FAIL (4-0 FALSE) | HOLD_CCS36_DANGLING_REF_CLOSURE (4/4 flagged) | PASS (4/4 tuples, certified) |
| M05 | S1 | FAIL (DOC-MSA 9.2 -> DOC-CSL 1.1, DOC-CSL > DOC-MSA) | FAIL (tuple=True) | PASS (4-0 FALSE) | PASS (4-0 FALSE) | HOLD_CCS36_DANGLING_REF_CLOSURE (4/4 flagged) | FAIL (4/4 tuples, certified) |
| M06 | S1 | PASS (DOC-LIC 4.3 -> DOC-AMD2 2.1, DOC-AMD2 > DOC-LIC) | PASS (tuple=True) | FAIL (4-0 FALSE) | FAIL (4-0 FALSE) | HOLD_CCS36_DANGLING_REF_CLOSURE (4/4 flagged) | PASS (4/4 tuples, certified) |
| M07 | S2 | PASS (DOC-OF2 5.1 -> DOC-DPA 9.3, DOC-DPA > DOC-OF2 > DOC-MSA) | PASS (tuple=True) | PASS (4-0 correct) | PASS (4-0 correct) | PASS (4/4 tuples, certified) | PASS (4/4 tuples, certified) |
| M08 | S2 | PASS (DOC-SOW3 4.2 -> DOC-MSA 6.1, DOC-MSA > DOC-SOW3) | PASS (tuple=True) | PASS (4-0 correct) | PASS (4-0 correct) | PASS (3/4 tuples, certified) | PASS (3/4 tuples, certified) |
| M09 | S2 | PASS (DOC-SLA 1.2 -> DOC-EXB 3.4, C03 Compositional Successor) | PASS (tuple=True) | HOLD_CIS35_SCALAR_OR_EXHIBIT | PASS (4-0 correct) | PASS (4/4 tuples, certified) | PASS (4/4 tuples, certified) |
| M10 | S2 | PASS (DOC-MSA 15.1 -> DOC-MSA 15.2, C12 Inline-Attestation Successor) | PASS (tuple=True) | HOLD_CIS35_SCALAR_OR_EXHIBIT | PASS (4-0 correct) | PASS (4/4 tuples, certified) | PASS (4/4 tuples, certified) |
| M11 | S2 | PASS (DOC-MSA 8.2 -> DOC-AMD2 1.1, DOC-AMD2 > DOC-AMD1 > DOC-MSA) | PASS (tuple=True) | PASS (4-0 correct) | PASS (4-0 correct) | PASS (4/4 tuples, certified) | PASS (4/4 tuples, certified) |
| M12 | S2 | PASS (DOC-MSA 16.1 -> DOC-OF7 9.1, DOC-DRAFT-SL decoy rejected) | PASS (tuple=True) | PASS (4-0 correct) | PASS (4-0 correct) | PASS (4/4 tuples, certified) | PASS (4/4 tuples, certified) |
Local CIS-35 and ungated four-family voting both release 6/6 unanimous false verdicts on dangling overrides: 6/12 false releases overall.
Before reopen, CCS-36 holds all 6/6 dangling packets with 4/4 witnesses flagging unattached_external_reference = true: 0/12 false releases.
Closing the transitive bundle certifies 12/12 correct releases with 0/12 false releases and 0/12 holds. M09 and M10 resolve the C03/C12 schema boundaries.
Strictly separated from live model data. All 4^7 = 16,384 constructed configurations contain 8 positive CCS_CERTIFIED states and 16,376 compromised states. Exactness holds within this declared finite universe, not as a production reliability guarantee.
| Gate | TP | FP | TN | FN | FAR | FRR | Accuracy |
|---|---|---|---|---|---|---|---|
| G0_gtp34_quorum_only (GTP-34 4-family quorum only) | 8 | 4,088 | 12,288 | 0 | 24.96% (0.249634) | 0.00% | 75.05% (0.750488) |
| G1_cis35_local_single_doc (Round 35 CIS-35 on local primary doc) | 2 | 126 | 16,250 | 6 | 0.77% (0.007694) | 75.00% (0.750000) | 99.19% (0.991943) |
| G2_multidoc_attachment_no_prec_dag (multi-doc closure, no pi_prec / tau_exec) | 8 | 56 | 16,320 | 0 | 0.34% (0.003420) | 0.00% | 99.66% (0.996582) |
| G3_multidoc_scalar_schema_rigid (multi-doc + pi_prec + tau_exec, rigid scalar schema) | 2 | 0 | 16,376 | 6 | 0.00% (0.000000) | 75.00% (0.750000) | 99.96% (0.999634) |
| G4_ccs36_without_diversity (CCS-36 semantic gates without >= 3 families) | 8 | 24 | 16,352 | 0 | 0.15% (0.001466) | 0.00% | 99.85% (0.998535) |
| G5_ccs36_full (CCS-36 complete gate) | 8 | 0 | 16,376 | 0 | 0.00% (0.000000) | 0.00% (0.000000) | 100.00% (1.000000) |
P_closed(m, q) = q^(m - 1)
At p_indep = 0.15, q = 0.85 and m = 3, quorum failure rises from 1.20% to 28.62%. Per-hop recall must exceed q* = 96.50% (0.964964) to beat a single full-packet auditor at p_single = 0.08. This is a stipulated mathematical surface, not a measured retrieval rate.
Our architectural reading: transitive reference closure generalizes local digest binding, while compositional tuples align receipts with base and controlling clauses. These explanations add no measured result and describe no shipped mechanism.
Temporal Drift and Post-Execution Amendment Supersession asks what happens when a later Amendment, Waiver or Termination Notice silently changes an obligation without a forward pointer in the base agreement, or when document stores serve stale binder snapshots. E37-A and E37-B are proposed, not executed.
E35-B (Live Multi-Model Benchmark) + E35-A (Exact 4,096-State Post-GTP-34 Diagnostic)
Why this milestone matters. Rounds 33 (AIC-33, 3,375 synthetic topologies) and 34 (GTP-34, 279,936 synthetic endorsement graphs) showed how to verify that a 3-of-4 witness quorum is anchored in at least three distinct external trust roots without shared-administrator or Sybil collapse, but both rounds explicitly noted that no live model benchmark had been executed. Round 35 (E35-B) is our first official live multi-model benchmark, testing what happens after AIC-33 and GTP-34 pass when live reasoning witnesses inspect complex multi-clause contracts.
We froze 12 realistic multi-clause commercial contract and policy audit cases (C01-C12) partitioned into three strata.
Kept strictly separate from live witness evaluations, the deterministic diagnostic E35-A enumerates all 4^6 = 4,096 post-GTP-34 witness-panel configurations across six 4-state dimensions (grounding_profile, context_completeness, family_diversity, span_grounding, override_reconciliation, input_integrity): 8 positive CIS_CERTIFIED panels and 4,088 negative panels.
It also computes the exact Eckhardt-Lee common-shock failure surface:
P_fail(3-of-4 | p, rho) = rho * p + (1 - rho) * (4 p^3 (1 - p) + p^4)
| Roster ID | Witness Architecture Family | Reasoning Configuration | Live Evaluations | Evaluation Arms Covered | Mean Latency | Receipt Verification |
|---|---|---|---|---|---|---|
| witness-a | Witness A (Architecture Family A) | reasoning: {"effort": "high"} (Max reasoning) | 52 | Arm 1 (12), Arm 2 (20), Arms 3-5 (20) | 4,002.3 ms | 52 / 52 SHA-256 verified |
| witness-b | Witness B (Architecture Family B) | reasoning: {"effort": "high"} | 16 | Arms 3-5 (16) | 2,785.1 ms | 16 / 16 SHA-256 verified |
| witness-c | Witness C (Architecture Family C) | reasoning: {"effort": "high"} | 16 | Arms 3-5 (16) | 5,448.7 ms | 16 / 16 SHA-256 verified |
| witness-d | Witness D (Architecture Family D) | reasoning: {"effort": "high"} | 16 | Arms 3-5 (16) | 17,593.4 ms | 16 / 16 SHA-256 verified |
| Total (E35-B) | 4 Distinct Witness Architecture Families | 100 | 5 Evaluation Arms (C01-C12) | 6,174.6 ms | 100 / 100 SHA-256 verified |
Every evaluation in E35-B executed live across four distinct reasoning architecture families with SHA-256 prompt hashing and deterministic JSON receipt verification.
| Arm | Configuration | Correct releases | False certified releases | Holds | False cert. rate |
|---|---|---|---|---|---|
| Arm 1: Single-Witness-A-Max | Single Witness A Max reasoning witness on canonical contract | 12 / 12 (10 / 12 exact governing section) | 0 / 12 | 0 / 12 | 0.0% |
| Arm 2: Mono-Slice-3of4 (Ungated) | 4x Witness A witnesses on retrieval slice, GTP-34 only | 8 / 12 | 4 / 12 (4/4 on Stratum 2) | 0 / 12 | 33.33% |
| Arm 3: MultiFamily-Slice-3of4 (Ungated) | Witness A + Witness B + Witness C + Witness D on retrieval slice, GTP-34 only | 8 / 12 | 4 / 12 (4/4 on Stratum 2) | 0 / 12 | 33.33% |
| Arm 4: CIS-35 on Raw/Truncated Slice | 4-family quorum gated by CIS-35 before canonical reopen | 6 / 12 | 0 / 12 (0/4 on Stratum 2; all 4 held fail-closed) | 6 / 12 | 0.00% |
| Arm 5: CIS-35 on Canonical Bundle | 4-family quorum gated by CIS-35 with canonical bundle digest, verbatim span and override verification | 10 / 12 (83.33%; 4/4 Stratum-2 cases recovered to 4-0 verified consensus) | 0 / 12 (0.00%) | 2 / 12 (16.67%; C03, C12) | 0.00% |
F35.3
On C05-C08, switching from a 1-family Witness A monoculture (Arm 2) to 4 distinct witness architecture families (Witness A, Witness B, Witness C, Witness D in Arm 3) produced zero reduction in false consensus: 4/4 unanimous 4-0 wrong verdicts in both arms (rho_hat = 1.00).
F35.4
CIS-35 caught 4/4 truncated slices fail-closed in Arm 4 (via missing cross-reference flags and digest mismatch) and converted all 4/4 into 4-0 verified correct releases in Arm 5 (0/12 false releases).
F35.5
All 4 architecture families agreed 4-0 on the correct verdict (PASS on C03, FAIL on C12) with verified verbatim quotes, but CIS-35 held both (HOLD_CIS35_GATE): on C03 because 3 of 4 witnesses labeled SLA Section 1.2 as governing_section while applying the 45-minute cap in SLA Exhibit B Section 3.4 in their reasoning; on C12 because Witness C and Witness D flagged missing_cross_reference = true when Exhibit C was summarized inline rather than attached.
| Case | Title | Expected verdict & governing clause | Arm 1 | Arm 2 | Arm 3 | Arm 4 | Arm 5 |
|---|---|---|---|---|---|---|---|
| C01S1 | Late MFA onboarding at Day 43 reinstates the 12-month cap | PASS (Schedule D Section 2.4) | PASS (gov=True) | PASS (4-0 correct) | PASS (4-0 correct) | PASS (4/4 spans) | PASS (4/4 spans, certified) |
| C02S1 | 2-business-day regional pinning cure extinguishes subprocessor termination right | FAIL (DPA Section 6.3(b)) | FAIL (gov=True) | FAIL (4-0 correct) | FAIL (4-0 correct) | FAIL (4/4 spans) | FAIL (4/4 spans, certified) |
| C03S1 | 540-minute HSM rotation outage exceeds the 45-minute Exhibit B exclusion cap | PASS (SLA Exhibit B Section 3.4) | PASS (cited 1.2) | PASS (4-0 correct) | PASS (4-0 correct) | HOLD_CIS35_GATE | HOLD_CIS35_GATE (verdict 4-0 PASS; 1/4 cited 3.4, 3/4 cited 1.2) |
| C04S1 | SKU SOV-ENCLAVE-ZERO under Rider S-1 prohibits 90-day telemetry retention | FAIL (Rider S-1 Section 4.2) | FAIL (gov=True) | FAIL (4-0 correct) | FAIL (4-0 correct) | FAIL (4/4 spans) | FAIL (4/4 spans, certified) |
| C05S2 | Order Form Rider R-2 raises the renewal cap from 5% to 11% | PASS (Order Form Rider R-2 Section 1.1) | PASS (gov=True) | FAIL (4-0 FALSE) | FAIL (4-0 FALSE) | HOLD_CIS35_GATE (xref=True) | PASS (4/4 spans, certified) |
| C06S2 | Addendum E excludes Chapter 11 DIP operations from escrow release | FAIL (Addendum E Section 2.2) | FAIL (gov=True) | PASS (4-0 FALSE) | PASS (4-0 FALSE) | HOLD_CIS35_GATE (xref=True) | FAIL (4/4 spans, certified) |
| C07S2 | Schedule G Item 14 warrants libquant-core v3.2 under LGPL-2.1 | PASS (Schedule G Item 14) | PASS (cited 10.4) | FAIL (4-0 FALSE) | FAIL (4-0 FALSE) | HOLD_CIS35_GATE (xref=True) | PASS (3/4 spans, certified) |
| C08S2 | First Amendment lowers the CPA audit reimbursement threshold from 5.0% to 3.0% | PASS (First Amendment Section 2) | PASS (gov=True) | FAIL (4-0 FALSE) | FAIL (4-0 FALSE) | HOLD_CIS35_GATE (digest!=H*) | PASS (4/4 spans, certified) |
| C09S3 | Unexecuted redline note (Delaware) vs binding Section 14.4 New York venue | PASS (Section 14.4) | PASS (gov=True) | PASS (4-0 correct) | PASS (4-0 correct) | PASS (4/4 spans) | PASS (4/4 spans, certified) |
| C10S3 | Non-binding 24-hour brochure target vs binding 48-hour notification SLA | PASS (Section 9.1) | PASS (gov=True) | PASS (4-0 correct) | PASS (4-0 correct) | PASS (4/4 spans) | PASS (4/4 spans, certified) |
| C11S3 | Clean control: Net-30 payment compliance | PASS (Section 2.1) | PASS (gov=True) | PASS (4-0 correct) | PASS (4-0 correct) | PASS (4/4 spans) | PASS (4/4 spans, certified) |
| C12S3 | Clean control: Exhibit C direct-competitor assignment restriction | FAIL (Section 15.2) | FAIL (gov=True) | FAIL (4-0 correct) | FAIL (4-0 correct) | HOLD_CIS35_GATE (xref=True) | HOLD_CIS35_GATE (verdict 4-0 FAIL, 4/4 spans; 2/4 flagged Exhibit C attachment) |
Source: Round 35, Sections 8.2 to 8.4. Twelve cases and one hundred live witness evaluations bound the empirical claim.
Strictly separated from live model data. These are exact enumerations over 4,096 constructed post-GTP-34 states, not measurements of any model.
| Gate | TP | FP | TN | FN | FAR | Accuracy |
|---|---|---|---|---|---|---|
| G0_gtp34_structural_only | 8 | 4,088 | 0 | 0 | 100.00% | 0.20% |
| G1_family_diversity_only | 8 | 2,040 | 2,048 | 0 | 49.90% | 50.20% |
| G2_verbatim_substring_only | 8 | 2,040 | 2,048 | 0 | 49.90% | 50.20% |
| G3_context_and_governing_span_no_override | 8 | 24 | 4,064 | 0 | 0.59% | 99.41% |
| G4_cis35_without_diversity | 8 | 8 | 4,080 | 0 | 0.20% | 99.80% |
| G5_cis35_full (CIS-35) | 8 | 0 | 4,088 | 0 | 0.00% | 100.00% |
Eckhardt-Lee correlated quorum crossover
rho* = 0.2471
At marginal witness error p = 0.25 versus a single checker at p_single = 0.10, a 3-of-4 majority quorum becomes strictly worse once pairwise error correlation exceeds 24.71%. At rho = 0.35 the quorum fails at 12.05% versus 10.00%.
Our reading of the results above. These are explanations, not additional measurements.
Strong instruction-following witnesses answer faithfully from the text they are given. When every witness receives the same truncated slice, that fidelity makes them all wrong in the same way, so architecture-family diversity cannot help: the error lives in the shared input, not in any one witness, and acts as a 100% common-mode multiplier.
AIC-33 and GTP-34 establish who ran each witness and whether their trust roots are independent. CIS-35 establishes what each witness actually inspected: the exact canonical byte digest and the clause override chain. The layers are complementary; neither substitutes for the other.
Historical Round 35 handoff below: CCS-36 has now been evaluated in Round 36, shown above. The current unrun handoff is Round 46 reviewer independence, a supersession field for the stricter consistency test and review load on split condition mistakes. The proposal wording is retained for archive continuity.
The 2/12 (16.67%) conservative holds measured in E35-B (C03 and C12) revealed a schema bottleneck: forcing witnesses to report a single scalar governing_section string causes false holds when a decision composes a base entitlement clause (SLA Section 1.2) with an exclusion ceiling (SLA Exhibit B Section 3.4). Round 36 will test CCS-36, replacing the scalar field with an ordered (base_entitlement_section, decisive_override_section, inline_fact_sufficiency) tuple on live multi-model evaluations, to test whether conservative holds drop from 2/12 to 0/12 while false releases stay at 0.0%. This is proposed and has not run.
Every new round runs against a machine-verified ledger of all prior rounds (Rounds 1-35, 127 hypotheses, 45 experiments) with Witness A on Max reasoning (effort: high) as the lead research and auditor witness, enforcing zero repetition and recording every evaluation's prompt integrity receipt and deterministic span-verification receipt.
Instead of assuming multi-agent panels improve accuracy by default, Noetis promotes verification gates only when they demonstrate a measured reduction in false certified releases against matched-compute baselines on live multi-model benchmarks. Round 35 changed no production system.
Proposed experiments are listed as proposed until they run. Before Round 35, executed entries were audits and exact diagnostic calculations, not model benchmarks.
All family and quote certificates pass. R47 holds five gold cures and ten schedule threshold mistakes. R47c releases all five cures but permits ten unsafe schedule threshold releases. No live calls.
From Round 47 · tests H47.1, H47.2, H47.4, H47.5
141 completed calls: one blind label review, 20 Witness C ledgers, 40 direct readings, 60 independent Witness D reviewer calls, and 20 Witness B ledgers (14 in the first run; the 6 stopped by the planned budget tail were completed post hoc the same day). The separate synthesis call by Witness A was also completed post hoc.
From Round 47 · tests H47.1, H47.3, H47.4
48 completed validator calls. Seventeen rewrites are correct, note following is zero, and seven false holds are accepted by the gate.
From Round 47 · tests H47.5
All five family certificates pass; R46 holds every correctly read cure on gold ledgers; F1 removes the near miss counting mistake and only R46 removes the schedule threshold mistake; a disagreement router releases an unsafe outcome only when both components share the error; 24 new exhibits with notes quoting real window text. No model calls.
From Round 46 · tests H46.1, H46.2, H46.7
5 families, 20 items, 141 completed calls: 1 blind review call by Witness A, 20 record set ledgers by Witness C, 40 direct readings by Witnesses B and C, 60 reviewer calls by Witness C under three views, and 20 record set ledgers by Witness B (8 in the first run; the 12 stopped by the planned budget tail were completed post hoc the same day).
From Round 46 · tests H46.1, H46.2, H46.3, H46.4, H46.5, H46.6
24 exhibits fixed in advance with accession numbers and annotated ledgers; Witness C made 48 validator calls (rewrite and flag only) with 24 correct rewrites, 0 note following, and 8 of 12 HOLD exhibits flagged for a missing condition.
From Round 46 · tests H46.7
All six family certificates pass; R45 reproduces all 60 labels on gold ledgers while R44 is wrong on every K2 and TS item; each scripted reader mistake defeats a different set of rules; 24 new exhibits with gold, seeded, and annotated ledgers. No model calls.
From Round 45 · tests H45.2, H45.3, H45.7
6 families, 30 items, 152 completed calls: 2 blind review calls by Witness A, 60 record set ledgers by Witnesses B and C, 60 direct readings, and 30 Witness D ledgers; 3 ledger calls first stopped by the run budget rule were completed in the same day, post hoc rerun with the later phases.
From Round 45 · tests H45.1, H45.2, H45.3, H45.4, H45.5, H45.6
24 exhibits fixed in advance with accession numbers and gold, seeded, and annotated ledgers; Witness C made 72 validator calls (seeded rewrite, annotated rewrite, annotated flag only) with 72 correct outcomes, 0 note following rewrites, and 12 correct repairs routed to review by the gate; run in the same day, post hoc completion.
From Round 45 · tests H45.7
Eight certificates pass: 8 families, 24 completion certified labels (8 HOLD), 48 of 48 gold record level resolutions, 24 new exhibits balanced 12 and 12 with gold and seeded ledgers, and a monotone rewrite gate. No model calls.
From Round 44 · tests H44.1, H44.6
8 families, 24 items, 128 completed calls: 8 blind review calls by Witness A, 72 record level ledgers by Witnesses B, C, and D, 16 condition ledger baseline calls, and 32 declared open direct readings by Witnesses B and C.
From Round 44 · tests H44.1, H44.2, H44.3, H44.4, H44.5
24 exhibits fixed in advance with accession numbers, gold ledgers, and seeded ledgers; 144 validator calls (Witness C on all 48 ledgers, Witness B on the 24 seeded ledgers, flag only and rewriting arms); no unsafe release in any arm.
From Round 44 · tests H44.6
Seven certificates pass: 6 families, 18 gold labels under both conventions, 24 new exhibits balanced 12 and 12, and flag only monotone toward HOLD. No model calls.
From Round 43 · tests H43.4
6 families, 18 items, 201 completed calls by Witness A to Witness D; 3 Witness A calls stopped by the budget rule. Open declaration removed absence errors; near miss errors remained in both roles.
From Round 43 · tests H43.1, H43.2, H43.3
24 exhibits fixed in advance with accession numbers, gold and seeded ledgers, 96 calls by Witness C. Flag only 0 unsafe and 12 false holds; rewriting 48 of 48 correct.
From Round 43 · tests H43.4, H43.5
Contrast pairs differ by one record; gold ledger reproduces all 12 labels; 20 new exhibits balanced 10 and 10. No model calls.
From Round 42 · tests H42.5
6 families, 12 items, 72 calls: direct and extraction readings by Witness B, Witness C, and Witness D. Direct pairs failed together at the Frechet maximum; role pairs 0 of 18.
From Round 42 · tests H42.1, H42.2, H42.3
20 exhibits fixed in advance with accession numbers, 120 calls. One direct unsafe release (Z16) and two extraction false holds (Z19); no joint unsafe release.
From Round 42 · tests H42.4
Four witnesses, three verdicts, three latent item types, twin fidelity 1.0, 0.7, and 0.4. No model calls.
From Round 41 · tests H41.1, H41.2, H41.5
10 families, 20 twins, 100 calls: direct verdicts from all four witnesses on both twins plus two extractions on the evaluation twins. 0/10 families with disagreement; F07 unanimous wrong.
From Round 41 · tests H41.3, H41.4, H41.5
20 exhibits fixed in advance with accession numbers; all 20 evaluated (100 calls). Two 2-2 splits (Y11, Y13), both in the execution only stratum.
From Round 41 · tests H41.5
243 verdict states, 16 non-abstaining HOLD-truth states, and extraction fault states for T3, T3 without span check, and T6. No model calls.
From Round 40 · tests H40.1, H40.2, H40.3
17 packets (11 from Round 39, K07 excluded, N01 to N06 new), ten shared calls per packet, all six topologies plus two references. 0 unsafe releases in every topology.
From Round 40 · tests H40.4, H40.5
16 exhibits from the Round 39 track with accession numbers. Five topologies 16/16; role pipeline 15/16 (false hold on X24).
From Round 40 · tests H40.4, H40.5
Exact enumeration of 112 states (7 resolution channels x 4 claim pairs x 2 relevance x 2 base verdicts; 108 determinate, 4 hold-required, 56 with conflicting records) across seven gates, plus the exact 3-of-4 lens composition analysis and a resolver check that reproduces 12/12 authored packet labels. G0-G3 21.43% unsafe; G4 17.86%; G5 0.00% unsafe, 3.70% false holds; G6 0 unsafe and 0 false holds by construction. No model calls.
From Round 39 · tests H39.1, H39.2
96 live witness evaluations across 12 packets (K01-K12), two lenses (presumptive, precautionary), and Witness A (Max reasoning), Witness B, Witness C, Witness D; 96/96 parsed. 48/48 correct on resolvable conflicts; 0/96 HOLD; K08 verdict equaled the lens in 8/8; K08 held in 6/6 balanced and released in 10/10 unbalanced lens assignments; CEV-39 gate 0/12 unsafe, 2/12 holds.
From Round 39 · tests H39.3, H39.4, H39.5, H39.6
24 executed exhibits (amendments, consents, merger agreements) filed on SEC EDGAR, frozen in 4 strata before any model call, each with accession number and archive link. Witness C 22/24 matched the structural label; Witness B 9/9 on a frozen 12-exhibit subset (3 not run); 8/9 paired agreement; deterministic prevalence scan over 109 fetched exhibits.
From Round 39 · tests H39.7
Exact enumeration of 512 raw states (464 feasible: 16 positive, 448 negative) across seven gates, plus five resolvers on V01-V12. G0 and G1 FAR 15.18%; G2 (TDR-37) FAR 9.15%, FRR 81.25%; G3 FAR 1.56%; G4 FAR 2.68%; G5 FAR 0.45%, FRR 6.25%; G6 (CIV-38) 0 FP and 0 FN by construction. No model calls..
From Round 38 · tests H38.1, H38.2, H38.3, H38.4, H38.5, H38.6
96 live witness evaluations across 12 lifecycles, two views, and Witness A (Max reasoning), Witness B, Witness C, Witness D with assigned evidentiary lenses. View R: 9/12 split 3-1; A1, A3 and A4 8/12 false releases; A6 CIV-38 0 false, 9 holds. View V: 48/48 correct; A5 TDR-37 10/12 holds; A7 CIV-38 12/12 correct..
From Round 38 · tests H38.1, H38.2, H38.3, H38.4, H38.5, H38.6
Exact enumeration of 1,536 feasible states (16 positive, 1,520 negative) across seven gates, plus four resolvers on L01-L12. G0 FAR 15.79%; G1 (CCS-36) FAR 13.68%; G2 FRR 50.00%; G3 FAR 2.11%; G4 FAR 1.58%; G5 FAR 0.53%; G6 (TDR-37) 0 FP and 0 FN by construction. No model calls..
From Round 37 · tests H37.2
96 live witness evaluations across 12 lifecycles, two views, and Witness A (Max reasoning), Witness B, Witness C, Witness D. View B: 7/12 unanimous wrong, ungated and CCS-36 7/12 false releases, TDR-37 0 false and 11 holds. View F: 48/48 correct, all arms 12/12..
From Round 37 · tests H37.1, H37.3, H37.4
48 live witness evaluations with the most recently executed instrument removed from packet and register extract. 20/20 wrong on the 5 decisive lifecycles; packet-register TDR-37 5/12 false releases; gate-side register 0 false, 11 holds..
From Round 37 · tests H37.5, H37.6
Exact enumeration of 4^7 = 16,384 multi-document reference-graph configurations (8 positive CCS_CERTIFIED, 16,376 negative compromised) plus an exact multi-hop reference-truncation surface. G0 FAR 24.96% (4,088 FP); G1 local CIS-35 FAR 0.77% (126 FP) and FRR 75.00% (6 FN); G2 FAR 0.34% (56 FP); G3 FRR 75.00% (6 FN); G5 (CCS-36) TP = 8, FP = 0, TN = 16,376, FN = 0. Canonical.
From Round 36 · tests H36.1, H36.2, H36.3, H36.4, H36.5
108 live witness evaluations across 12 multi-instrument enterprise transaction packets (M01-M12) and four distinct witness architecture families (Witness A, Witness B, Witness C, Witness D). Local single-document CIS-35 (Arm 2) and ungated 4-family quorums (Arm 3) both suffered 6/6 unanimous 4-0 false consensus on cross-document dangling overrides (M01-M06, 50.00% false certification rate); CCS-36 intercepted 6/6 fail-closed before reopen (Arm 4, 0/12 false releases) and achieved 12/12 (100.00%) certified correct releases with 0/12 false releases and 0/12 holds on the closed multi-document bundle (Arm 5). Canonical.
From Round 36 · tests H36.1, H36.4, H36.6, H36.7
Exact enumeration of 4^6 = 4,096 post-GTP-34 witness-panel configurations (8 positive, 4,088 negative) plus an exact Eckhardt-Lee correlated-quorum failure surface. G0 structural-only FAR 100.00%; G1 and G2 FAR 49.90%; G5 (CIS-35) FAR 0.00%, FRR 0.00%. Canonical. Combinatorial facts, not prevalence estimates.
From Round 35 · tests H35.1, H35.3, H35.4, H35.5
100 live witness evaluations across 12 frozen contract/policy audit cases (C01-C12) and four distinct witness architecture families (Witness A, Witness B, Witness C, Witness D). Ungated monoculture and ungated 4-family quorums both suffered 4/4 unanimous false consensus on shared retrieval-slice truncation traps (33.33% false certification rate); CIS-35 on the canonical bundle reduced false certified releases to 0/12 with 2/12 conservative fail-closed holds (C03, C12). Canonical. Twelve cases bound the empirical claim; no production system changed.
From Round 35 · tests H35.2, H35.6, H35.7
Exact enumeration of 279,936 finite synthetic graphs only; not a production PKI audit, security certification, model benchmark or prevalence estimate. Result. Configured anchor authenticity and independence are inputs, not inferred.
From Round 34
Not run. Independently authored trust-graph fixtures at equal verification budgets, followed by matched-compute collaborative reasoning only when A3, I1, R4, R5 and C6 are present. Readiness remains BLOCKED.
From Round 34 · tests H34.1, H34.2, H34.3, H34.4, H34.5, H34.6, H34.7
Proposed only. Preferred: provision three independently administered witnesses with one external binding per witness per critical layer, run AIC-33 before action, and adjudicate hidden shared-root and stale-binding cases separately. Fallback: basis-completeness red team. The matched-compute model experiment activates only after A3, I1, R4 and R5 are present. No experiment in Round 33 Section 12 was run.
From Round 33 · tests H33.1, H33.2, H33.3, H33.4, H33.5, H33.6
Proposed only. Harden independently administered witness signing, ephemeral isolation and external receipt registration; keep model calls disabled until critical infrastructure controls pass. No live E30-B or answer benchmark was run.
From Round 31 · tests H31.1, H31.2, H31.3, H31.4, H31.5, H31.6
No E30-B results are reported. Blocked until critical readiness controls pass in a sandbox and the receipt verifier is tested independently. Proposed 720 paired cases, four arms, matched compute and independent exposure records.
From Round 30 · tests H30.2, H30.3, H30.4, H30.5
72 structural worlds, 36 paired response behaviors, six policies and four budgets yield 1,728 traces. Full content probes identify behavior in 100% and provenance in 0%. Exact only for the declared diagnostic. No live benchmark ran.
From Round 29
No results reported. Preregistration-ready design for 720 paired cases; permissioned replay, independent receipts and fixed compute required.
From Round 29 · tests H29.1, H29.2, H29.3, H29.4, H29.5, H29.6
Model-set intervals were conditionally valid in-model, but passive adequacy signals failed on the hidden-common-origin alias. TSAR did not universally dominate a scalar. E28-A is an exact constructed trace replay of 9,744 frozen histories and 29,232 weighted rows, not a production, language-model, or Encevra answer benchmark. Round 29 remains a proposed study of active dependency interventions that can reveal omitted dependencies.
From Round 28 · tests H28.1, H28.2, H28.3
E27-A is exact only for the declared scenario set, deterministic observation rules, authored cost tables, priors, and policies: 58 worlds per dependency variant, a 116-scenario planning set, 9,744 policy-world runs, and 504 aggregate evaluation cells. Four model-set gate policies produced no wrong categorical verdict in 3,712 in-model runs; the out-of-model hidden_common_origin stress produced false-clean rates up to 33.68 percent at budget 9. Not a production trial, not a language-model reasoning benchmark, and no live Encevra benchmark was run. E27-B and H27.1 to H27.7 are proposed. Architecture impact: None. Production impact: None.
From Round 27
Not run. Requires fixed-compute shadow trial with hidden independent adjudication, frozen lineage graphs, costs, and policies, and separate natural and sandbox strata.
From Round 27 · tests H27.1, H27.2, H27.3, H27.4, H27.5, H27.6, H27.7
E26-A is exact only for its declared finite model: 2, 16, 58, and 128 worlds at fault budgets 0 through 3, respectively. Seven failure roots and eight observable types were enumerated, with eight executable tests passed. Root independence is an assumption. The proposed Observation Sufficiency Certificate (OSC) is research-only, not installed or production-validated. No production incident was sampled, no language model was benchmarked, and no new Encevra answer benchmark was run. E26-B and H26.1 to H26.7 are proposed. Architecture impact: None. Production impact: None.
From Round 26
Not run. Hidden independent adjudication, fixed budgets, and separate natural and sandbox strata required.
From Round 26 · tests H26.1, H26.2, H26.3, H26.4, H26.5, H26.6, H26.7
D25-A is an authored-fixture conformance diagnostic, not production evidence. 80 traces: 60 positive gaps, 20 negative controls. SACP TP=55/60, FP=0/20, trace recall 0.9167, claim-weighted recall 0.8944, with five intentional misses.
From Round 25
Not run.
From Round 25 · tests H25-1, H25-2, H25-3, H25-4, H25-5, H25-6, H25-7, H25-8, H25-9, H25-10
D24-A is a deterministic authored-fixture conformance diagnostic, not an estimate of production precision, recall, reliability, prevalence, or cost. No Encevra answer benchmark, external-service experiment, production observer, or deployment was run.
From Round 24
Not run.
From Round 24 · tests H24-1, H24-2, H24-3, H24-4, H24-5, H24-6, H24-7, H24-8
Result boundary: one post-freeze real release interval; same-process post-hoc impact labels; eight authored conformance fixtures; zero model, external-service, production-database, and production-write calls; no product reliability, answer-quality, security, latency, or cost claim; no production CTAG or ABMG implementation.
From Round 23
Not run.
From Round 23 · tests H23-1, H23-2, H23-3, H23-4, H23-5, H23-6, H23-7, H23-8
Seven internal consistency checks passed. D22-A is deterministic routing against a manually audited actual diff plus authored mutation fixtures. It made zero model/external-service/production-database calls. It is not a model-quality benchmark, automatic dependency-discovery result, production integration test, security audit, product reliability claim, production change, or deployment.
From Round 22
Not run. Preregistered selectors, independent two-reviewer gold set, prospective real commits plus controlled mutations.
From Round 22 · tests H22-1, H22-2, H22-3, H22-4, H22-5, H22-6
34 of 34 named checks over constructed write states, envelope mutations, source-version pairs, service-response pairs and stream fixtures. Zero model, external-service network, or production-database calls. Conformance diagnostic, not an outcome benchmark.
From Round 21
Not run. Shadow mode at the external-service adapters with fault injection, an independent proxy oracle, ablations, and preregistered adoption thresholds.
From Round 21 · tests H21-1, H21-2, H21-3, H21-4, H21-5, H21-6
76 of 76 named checks across 96 constructed episodes, eight synthetic task families, twelve corruption families, and 6,912 mutated episode-arm evaluations; 480 valid reveals verified, 480 altered reveals rejected. Zero model calls; development diagnostic, not a security evaluation.
From Round 20
Not run. Compares MDIR against strong generic traces with independent immutable export on frozen synthetic trajectories, including unavailable openings and failed checkpoints.
From Round 20 · tests H20-1, H20-2, H20-3, H20-4
44 named checks across 15 constructed fixtures; static audit of 18 collector requirements (4 present, 3 partial, 11 absent). No model call or human annotation.
From Round 19
Not run. Tests the measurement contract, not model intelligence; specified to run before E18-B.
From Round 19 · tests H19-1, H19-2, H19-3
36 passing checks (24 across all six arrival orders, 12 semantic) and twelve developer fixtures. No language-model benchmark, model call, or human annotation.
From Round 18
Not run; gated on the existing Round 13 collection gate. Proposed 480 development episodes across twelve source families.
From Round 18 · tests H18-1, H18-2, H18-3
28 named checks, all 15 partitions of a four-world example, 60 deterministic recompressions, and twelve developer fixtures. No language-model benchmark or model inference.
From Round 17
Not run or preregistered; gated on the existing Round 13 collection gate. Proposed 288 core episodes across twelve source families.
From Round 17 · tests H17-1, H17-2, H17-3
6,912 Boolean function-observation pairs, 37 named checks, and twelve generated developer fixtures. No language-model benchmark or model inference.
From Round 16
Untested with language models; not preregistered or run. Proposed 24 instances and 192 assigned continuations across B0-B3, after the Round 13 collection gate.
From Round 16 · tests H16-1, H16-2, H16-3
32 mathematical checks on invented distributions. No model inference or human annotation.
From Round 15
Not preregistered or run; follows the Round 13 collection gate.
From Round 15
Not preregistered or run; oracle delivery cannot establish sender discovery.
From Round 15
30 checks on invented labels, reproducible from published files. No model inference or human annotation.
From Round 14
Proposed for preregistration, not registered or run; same frozen answers and matched annotation and judge budgets.
From Round 14 · tests H14-1, H14-2, H14-3
Offline replay and perturbation of existing exports. No new model runs.
From Round 13
Fixed-state branches; labels not produced by the system under test.
From Round 13 · tests H13-1
Worked calculations on invented episodes, reproducible from published files.
From Round 12
Preregistered; identical generation compute across arms.
From Round 12 · tests H12-1, H12-2, H12-3, H12-4
Three cost ceilings; harmful-revision denominator fixed in advance.
From Round 11 · tests H11-1, H11-2, H11-3, H11-4
Six arms including debate and independent baselines.
From Round 10 · tests H10
Changes made to Noetis because of a research round. These are safeguards and contracts; none is claimed to improve answer quality until tested.
Flag synthesis that omits supported risks or adds untraced claims, rather than presenting it as consensus.
Source: Round 12, 13
Return a null risk score until version-bound, externally labelled calibration exists.
Source: Round 12
Skip optional roles only on confident, valid routing; otherwise run every role.
Source: Round 11
