Encevra SI
Research Labs

We investigate how AI systems exchange evidence and revise conclusions when their assumptions change. Our research examines when coordination helps and how to measure its contribution, and we study learning reusable procedures from verified experience. The eventual goal is Artificial Superintelligence: cognitive systems that reason beyond any single model or person. We pursue it the only way we consider legitimate, through falsifiable research and an architecture that improves with every round, not through claims.

When a Replacement Is Not a Cure?

Builds on Round 46 by adding explicit supersession to a conservative record set resolver and by using an independent reviewer who wrote neither the ledgers nor the direct readings. Five new authored families supply 20 items and 24 new public exhibits supply a separate validator track. On gold ledgers, R47 releases none of the five cures held by R46, while retaining protection against all ten scripted schedule threshold mistakes. Exploratory R47c releases all five gold cures but permits ten unsafe schedule threshold releases. On Witness C live ledgers, R47 improves correct open outcomes from 10 to 13 without an unsafe release. Every routing composition escapes zero unsafe outcomes, but C1 with R47 routes 17 of 20 items. Independent Witness D is correct on 17 of 20 with packet only, 9 with brief only, and 11 with both. Packet and brief follows four of seven wrong resolver holds, the same count as brief only. On public exhibits, the validator follows the planted note zero times and introduces seven false holds. In the first run, six second ledger calls and the synthesis call stopped at the preregistered budget boundary; all seven were completed the same day as a disclosed post hoc completion. On all 20 Witness B ledgers, R47 improves correct open outcomes from 12 to 13 without an unsafe release.

Who Must Look Again?

Asks the roadmap's review routing question (P3, disagreement buys a hold, not a vote): what a consistency test in the resolver holds back, how much work disagreement routing sends to review and what escapes, and what a reviewer must be shown. Five new authored contract families (an underwriting lock up, key employee agreements signed with the acquirer, appraisals by an approved appraiser, unconditional landlord consents, and site acceptance after a burn in period) state an "each" requirement with an aggregate threshold, in four variants: P (two records together meet it), S (a near miss; HOLD under an open record), C (plus a completeness record; FAIL), and R (plus a record that cures the near miss; PASS). The same ledgers are resolved by four fixed rule sets: R45, R45 with a record set consistency test (F1), R46 (F1 plus a universal condition test, F2; pre registered), and R46c (exploratory). A structural diagnostic (E46-A, 0 model calls) shows F1 removes one split condition mistake, only R46 removes the other, and R46 holds every correctly read cure, turning each into a wrong FAIL under a closed record; it also certifies that a disagreement router can release an unsafe outcome only when both components share the error. Live, the blind review agreed with 40 of 40 labels. On Witness C's 20 ledgers R45 was correct on 20 and R46 on 15, holding all 5 cures. Both direct readers were correct on 17 of 20, every unsafe verdict a FAIL on a near miss. Routing the resolver against the other witness's reading sent 11 of 20 items to review with 0 escapes (6 with R45); routing two readers against each other sent 4 and let 2 shared errors through. A reviewer shown the packet only was correct on 16 of 20 with 4 unsafe FAILs, shown a brief only 14 of 20 with 6 needless holds, and shown both 20 of 20. Witness C rewrote 24 real SEC exhibit ledgers carrying a note that quotes real filing text and followed the note on none. Witness B's 20 ledgers (12 completed post hoc the same day after a budget stop) gave 3 split condition releases under R45 and none under any consistency rule. Refines P3; no production system changed.

The Rule or the Reader?

Asks the roadmap's resolver failure question directly: when the reader gets the records right, where is a fixed resolver rule wrong, and when the rule is changed to fix that, what new mistakes does it let through. Six new authored contract families (an escrow release needing joint instructions, a preferred stock consent, a termination notice to two addressees, payment and performance bonds, an environmental report with a reliance letter, and tenant estoppels covering 75 percent of rentable area) are each written as K2 (two records together meet the requirement), T (a near miss), TC (plus a completeness record), TX (plus an added record that is not a completeness record), and TS (plus an exclusion whose scope leaves a route open). Every label carries a two completion certificate and Witness A reviewed all 30 packets blind before any test call. A new ledger asks for a status per record, record sets, and scoped exclusions, and the same ledger is resolved by four fixed rule sets: the Round 44 rules (R44), R44 with record sets, R44 with a scope test, and both (R45, pre registered). A structural diagnostic (E45-A, 0 model calls) shows the Round 44 rules wrong by construction on every K2 and TS gold ledger, and certifies 24 new SEC EDGAR exhibits with gold, seeded, and annotated ledgers. Live, the review agreed with 60 of 60 labels. On 60 record set ledgers by Witnesses B and C, R45 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. On 19 ledgers the old rule was wrong and the new rule right; on 3, all Witness B, the old rule was right and the new rule wrong, most instructively when a stale estoppel certificate the witness itself marked FAILS was credited in a record set. TC resolved to FAIL on 8 of 12 ledgers and TX on 0 of 12, so the completeness content, not an added record, drives the FAIL. Record sets without the scope test raised unsafe releases from 11 to 19. Direct readers were correct on 46 of 60 with 14 unsafe FAILs on near miss items; where the resolver was wrong the reader was right on 7 of 8, and the two failed together on 1. Witness C validated 24 real SEC exhibit ledgers in three arms with 72 of 72 correct outcomes and followed a misleading reviewer note on none. The run budget rule first stopped the round after 57 of 60 record set calls; the remaining phases were completed the same day and are disclosed as post hoc. Refines P2 and P7; no production system changed.

This Record Fails, or No Record Exists?

Tests whether evidence readers stop reading a defective record as proof that no valid record exists when the ledger asks what each record shows and, separately, whether anything excludes other records, as the Round 43 handoff asked, and puts the roadmap's abstention labels through a blind independent review. Eight new authored contract families (a short cure payment, an early extension notice, a certificate holder in place of an additional insured, an unsigned execution version, an early letter of credit expiry, a short lien waiver, an in house opinion, and a PDF image in place of an original certificate) are each written as K (the final record satisfies the requirement), T (a near miss), and TC (the near miss plus a completeness record). Every HOLD label carries a two completion certificate, and Witness A reviewed all 24 packets blind under an open and a closed record before any test call. A structural diagnostic (E44-A, 0 model calls) certifies the variants, the labels, the resolver contrast, 24 new SEC EDGAR exhibits with gold and seeded ledgers, and the monotonicity of the pre registered rewrite gate. Live, the review agreed with 48 of 48 labels and reported packet defects in 6 of 8 families, mostly on K packets. Record level ledgers by Witnesses B, C, and D held 23 of 24 near miss T readings under the open resolver (1 unsafe FAIL), while on the same items the unchanged condition ledger released an unsafe FAIL on 12 of 16 and declared open direct readers on 8 of 16; all 24 TC ledgers and all 16 TC direct readings resolved to FAIL. Witnesses B and D over held K items where two records together satisfy a requirement, and the one T error quoted a record scoped negative as an exclusion. On the 24 SEC exhibits (E44-C, 144 validator calls), no flag only or rewriting validator made an unsafe release; flagging caught every near miss evidence seed but could only hold on spurious conditions, rewriting repaired all 48 seeded ledgers, and the pre registered gate routed the 24 hold to release repairs to review. Phases first stopped by the run budget were completed later the same day with unchanged prompts and settings. Revises P7 (a failed record is not a failed condition), extends P1 (abstention labels need certificates and review), refines P5, and hands two resolver failure cases to Round 45; no production system changed.

Open Record or Closed Record?

Tests whether readers that took a missing record as proof an event did not happen (Round 42) misread absence or silently apply a closed record convention, and whether a validator should flag defects or rewrite them, as the Round 42 handoff and the roadmap asked. Six new authored contract families are each written in three variants that share every record but the last: in U the final record only suggests the decisive fact, in K it establishes it, and in T it is a near miss, a record of the right kind that fails one stated requirement. Witness B, Witness C, and Witness D answered every item under a declared open record and a declared closed record and with the undeclared Round 42 prompt, and filled one evidence ledger per item resolved by code under both conventions; Witness A read the T items under the open declaration until the budget rule stopped it. A structural diagnostic (E43-A, 0 model calls) certifies the variants and the exhibit track and shows that a flag only validator cannot create an unsafe release while a rewriting validator can. Live, 297 completed calls (E43-B and E43-C) found that undeclared direct readers released FAIL on 33 of 36 U and T decisions; declaring an open record removed every absence error (U 15 of 18 to 0 of 18) but not the near miss errors (T 18 of 18 to 6 of 18), and a declared closed record was followed on 36 of 36. Extraction readers released FAIL on 14 of 18 T items under the open resolver, and 5 of 6 direct unsafe releases coincided with the same witness's extraction. On 24 new SEC EDGAR exhibits with gold and seeded ledgers, the flag only validator made 0 unsafe releases in 48 and 12 false holds it could not repair; the rewriting validator was correct on 48 of 48 with no structural bound. Revises P7 (a failed record is not proof of absence), P4 (role decorrelates only where failure modes differ), and P5 (gate rewrites that can release); no production system changed.

Role or Witness?

Measures whether one witness reading in two roles misreads undetermined facts jointly less often than two witnesses in one role, as the Round 41 handoff asked. Six authored contract families are each written as two contrast variants that differ by exactly one record: in U the decisive record is absent and the label is HOLD; in K it is present. Witness B, Witness C, and Witness D each read every item as a direct reader and as an extractor whose evidence ledger is resolved by fail closed code, and the same six readers read 20 new balanced SEC EDGAR exhibits. Live, 192 scored calls (E42-B and E42-C) found that direct readers released 11 unsafe verdicts in 18 undetermined decisions, all FAIL, reading absence as negation; two witnesses in the direct role failed together on 8 of 18 pair opportunities, the Frechet maximum, while one witness in two roles failed together on 0 of 18. Revises P4 to decorrelate by role, then verify by structure, and adds proposition P7, absence is not negation; no production system changed.

Hard by Construction or Hard by Selection?

Tests whether building a hard pool by selecting items on which witnesses disagree, as the Round 40 handoff proposed, is sound. Each authored item is written as two isomorphic twins with an identical rule trace and verdict, one used to screen and one held out. An exact diagnostic (E41-A, 0 model calls) shows disagreement selection keeps 42.2 percent of items, raises the forced majority's wrong release rate from 0.139 on the full pool to 0.285 in sample, and more than doubles its measured gap over one witness (0.055 against 0.023), a gap that falls to 0.028 on a fresh twin; 13.4 percent of forced majority errors sit on items disagreement never selects. Live, 200 calls with Witness A (Max reasoning), Witness B, Witness C, and Witness D over 10 hard by construction families (E41-B) and 20 fixed real SEC EDGAR exhibits (E41-C) found 0 of 10 families with any disagreement and 40/40 stable witness twin pairs. The only authored error was unanimous (F07, a missing delivery fact) and only the disagreement routed hybrid held it; the only exhibit disagreements (Y11 and Y13, both effective only on execution and delivery or receipt of all counterparts) were released by the forced majority on a tie and held by the quorum and the resolver. Adds proposition P6, disagreement is a biased detector; no production system changed.

Composition Before Capability

Compares six ways of composing four witnesses into one contract audit decision: single witness, forced majority, extract then deterministic resolve, role specialized pipeline, critique and revision, and a disagreement routed hybrid, scored on accuracy, unsafe releases, abstention calibration, and model calls per decision. An exact diagnostic (E40-A, 0 model calls) shows a forced majority releases 108/243 states unsafely, the same as one witness, and holds 0/16 HOLD-truth states when witnesses do not abstain; disagreement routing raises the fewest faults needed for an unsafe release to 2. Live, 330 calls over 17 authored packets (E40-B) and 16 real SEC EDGAR exhibits (E40-C) with Witness A (Max reasoning), Witness B, Witness C, and Witness D produced 0 unsafe releases in every topology. The one single witness unsafe answer (Witness B on N01) was removed by every multi-look topology, and the one composed error (a false hold on X24) was introduced by a validation stage that rewrote a correct ledger. Proposes five named propositions and a Rounds 41 to 50 path toward Theory v1; no production system changed.

When the Records Disagree

Pairs an exact enumeration of 112 conflicting-evidence states (E39-A) with 96 live witness evaluations across 12 authored contract packets, 2 evidentiary lenses, and 4 witnesses: Witness A (Max reasoning), Witness B, Witness C, and Witness D (E39-B), plus a 24-exhibit track of real executed amendments, consents, and merger agreements filed on SEC EDGAR (E39-C). When two attached collateral records disagree, single-record and single-lens gates release 20 to 24 of 112 states unsafely; ordered CEV-39 resolution (designated record, then required computation, then agreement, else UNDETERMINED with verdict invariance) releases none unsafely and holds only the 4 required states. Live, witnesses were correct on 48/48 evaluations of conflicts the contract resolves, under both lenses, but returned HOLD on 0/96 evaluations, including 0/16 on the two outcome relevant unresolved packets, where they followed the assigned lens. A balanced two and two lens panel held the clean unresolved packet K08 in 6/6 balanced assignments and released it in 10/10 unbalanced ones. The CEV-39 gate made 0/12 unsafe releases. On real filed exhibits, 31/33 completed evaluations matched the structural effectiveness label. Twelve authored packets and 33 real-text evaluations bound the empirical claim; no production system changed.

Signed is not operative

What changed after the agreement was signed?

A complete file can still leave an incomplete decision

Live empirical experiment spotlight