DIVERGE-NVE1 Natural Variable Evidence Result

Status: PASS on the one frozen run, 2026-08-06.

Question

Can one learned natural-evidence compiler bind instruction ordinals, signed rational values, and source-owned named registers, then drive exact delayed recovery through the protected TFS1 factorized runtime across swaps, guarded predicates, and noncommuting state updates?

NVE1 is a controlled interface/mechanism gate. It is not a public benchmark, continuation-pretraining result, or claim of unrestricted language reasoning.

Frozen lineage and data

• preregistration commit: 5ca88f6;
• implementation commit: ada830d;
• pre-data audit correction: da2adcd;
• scientific launcher commit: b06bae1;
• strengthened evaluator commit: fbb2f33;
• unchanged TOL3 checkpoint SHA-256: b8b9dfe54b7ab4a31a74739625b8650fa4ee93a41221ab5d82610ebc1c030328.

The deterministic training set contains 50,000 natural evidence statements over six balanced layouts. Training-data SHA-256 is 7eb27276332a14dcbf57c651c8823393fa7eb18b4122cff30c07b961db285b35.

The fresh confirmation board contains 256 episodes, 12 binary semantic fault lines per episode, and 4,096 coherent programs per episode: 3,072 evidence items and 1,048,576 represented programs in total. Three held layouts each contribute 1,024 statements. Exactly four confirmation occurrences are duplicates of another confirmation sentence; this is reported and was never a frozen exclusion. Training/confirmation exact sentence overlap is zero; episode identities are unique, and no model score enters selection. Board SHA-256 is 23e06c65ed89ce9136d3a504384d5be7996a76cddef089a92698e1b01b990add; data-report SHA-256 is 54aaf1d5b40dc8200970dcbd83286ff81d3b9f2ec0011943cc0cce9f4bdd40e7.

Learned interface

The only newly trained component is a 435,076-parameter, two-layer, bidirectional byte GRU with width 192, no position embeddings, and zero dropout. A lexical scanner proposes exactly two complete numeric/rational mentions. One hard permutation assigns them to STEP/VALUE. Source-owned register occurrences are quotient-grouped by identity, including repeated mentions, and a second hard permutation assigns two distinct groups to TARGET/DISTRACTOR.

Only after the hard assignments does exact code parse the one-based ordinal and rational. The sealed receipt binds packet, source, evidence sentence, compiler, step, target, distractor, value, and span/group provenance. It is either converted to the unchanged TFS1 typed receipt or rejected; no field is silently repaired.

The frozen schedule is seed 2026080610, 1,000 AdamW updates, batch 256, learning rate 0.003 with cosine decay, and class-balanced loss over both hard role assignments. CPU training consumes 25,618,773 source bytes in 2,582.580 seconds. Training numeric, symbol, and joint assignment are each 50,000/50,000. Checkpoint SHA-256 is 1610815471c695b0d2d198922dd99369e1f45a5dabc1b1c5d8e986b30fd200ff; model-state SHA-256 is ab321d501dd81d29317b1253f72761bd8f98f31cfab62c828a7f373b39485fcd; training-report SHA-256 is d6bc7bb05cc86f3513500966e9d33a7eb24bc2dc39c721840e2885824b7bf67c.

Confirmation result

The source and evidence components are exact:

• TOL3 complete programs: 256/256;
• two positive operation options: 3,072/3,072;
• TOL3 gold support: 256/256;
• learned natural receipts: 3,072/3,072;
• held renderer 0/1/2: 1,024/1,024 each;
• fully sealed evidence sets: 256/256;
• initially wrong highest-support top-1: 256/256.

The matched end-to-end results are:

Learned natural evidence recovers all 256 episodes whose initial top-1 program is wrong. Factorized state maps match complete independent enumeration with and without evidence on every episode. Oracle and learned evidence preserve exactly the gold program and terminal state. No-evidence support answers all 256 invariant queries, and withholding the last receipt abstains on all 256 underdetermined queries.

Causal, integrity, and resource controls

• shuffled complete evidence: 0/256;
• forced target/distractor role swap: 0/256;
• forced step/value role swap: 0/256;
• declaration-state reset: 0/256;
• fault-operation semantic shift: 0/256;
• packet/query swap: 256/256 reject;
• post-seal evidence poison: 256/256 bit-invariant;
• wrong source, packet, sentence, step, target, distractor, or value accepted: zero;
• false commitments, malformed accepted packets, gold deletion, and overflow: zero.

Canonical complete-particle storage is 11,320,697,243 bytes versus 4,788,501 bytes for factorized packet/state/natural-evidence accounting, a 2,364.142x ratio. The equal-memory control funds one highest-ranked complete particle per episode. Factorized execution performs 7,639,040 logical instruction applications through 9,472 unique state-group applications, an 806.486x sharing ratio. Peak surviving state groups is two. Evaluation runs in 1,432.164 CPU seconds.

Evaluation SHA-256 is 2cda05808f0e5230fd285c345a7431b6791a22b3c35c62b227df35c0f622d6c3.

Decision and boundary

NVE1 passes the intended boundary. A learned source-disjoint natural interface can compile variable identity, step provenance, distractor provenance, and rational state evidence, then cause exact delayed recovery across a typed program with swaps, guards, predicates, and noncommuting updates. This extends NFE1 beyond scalar equations while leaving TOL3 and TFS1 mechanics unchanged.

The claim remains bounded. The sentence grammar, numeric scanner, declaration- owned register table, operation vocabulary, rational executor, and evidence verifier are engineered. The evidence describes verified state after each ambiguous update; the model does not yet induce an unrestricted fault-line ontology from arbitrary prose. NVE1 is not general reasoning and does not authorize continuation pretraining.

Freeze NVE1 with no width, duration, seed, optimizer, renderer, or loss variants. The passed gate authorizes one integrated trainable DIVERGE module that jointly learns source semantics, a compact coherent version-space packet, evidence refinement, execution, and late query readout. That successor must retain TFS1 and NVE1 as protected controls and earn any scaling step through held natural-program capability rather than synthetic fit alone.