DIVERGE-IEM1 Integrated Epistemic Machine Result

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

Question

Can one shared trainable whole-mention encoder, one learned latent-to-primitive transport, and one optimizer replace the separately qualified TOL3 source, NVE1 evidence, and typed-query interfaces while retaining exact factorized execution and delayed recovery?

IEM1 is a controlled integration gate. It is not a public benchmark, continuation-pretraining result, or test of unrestricted language reasoning.

Frozen lineage and data

• frozen specification commit: 6f75eca;
• implementation commit: 058d2ad;
• hash-bound Newton launcher commit: 8fd48c5;
• corrected immutable TOL3 artifact paths: 112f5d7;
• empty-evaluation-batch fail-closed repair: fb23f6b.

The repair adds only an empty-input return to the query batch compiler. It was made after training, changes no model, data, prediction, denominator, or gate, and allowed the already-trained checkpoint to report a valid component miss.

IEM1 reuses 50,000 NVE1 evidence statements and adds 50,000 deterministic natural-query statements. The fresh confirmation board contains 256 programs, 3,072 evidence items, 768 late queries, 12 binary fault lines and 4,096 worlds per program, or 1,048,576 represented worlds. Training and confirmation have zero exact evidence, query, or program-identity overlap. The immutable hashes are:

• query training data: 8dc6085f7632b56563416fa75c13ff611344ad53c7c2ca4350c54bbcc17301fa;
• confirmation board: 4294799cace78330eabdbba369d383feac870329edd00e15ba53d3a24cc9ce8a;
• data report: 2e516549ce5d7fb4ece0967c2255d8c272cad0e98750bb33710f0f83e87d8f6c;
• inherited evidence data: 7eb27276332a14dcbf57c651c8823393fa7eb18b4122cff30c07b961db285b35.

Model and training

The 550,343-parameter IEM1 starts from the exact NVE1 byte encoder. The same encoder serves source operations and predicates, evidence roles, and query roles. Four operation channels and six comparator channels are transported through learned doubly normalized matrices into fixed rational primitives. Only the composed semantic distribution is supervised; hard inference commits to one latent channel and one primitive.

Newton job 744075 ran on one H100 for exactly 1,000 AdamW updates at batch 256 and learning rate 1e-3 with cosine decay. It consumed 43,176,637 source bytes in 119.504 seconds, sustained 361,298 source bytes/second, used 497,242,624 peak allocated bytes, and recorded no skipped or non-finite updates. Training fit was:

• evidence numeric, symbol, and joint roles: 50,000/50,000 each;
• natural query role: 50,000/50,000;
• local operation phrases: 48/50;
• local comparator phrases: 11/18.

The final operation and comparator train accuracies were 96.0% and 61.1%. The final total loss was 0.294324; evidence loss was zero and query loss was effectively zero. The learned transport remained diffuse. Direct hard/soft inspection showed SET at 0/2 and LT/GE at 0/3 each, while the remaining direct operations were exact.

Artifact hashes are:

• checkpoint: c7560eb5c0bb08a8bbe0f15b1127790859fb2a12bf80d5a89f292be30789e84a;
• model state: 6552b6fbd6e736338ae2b6a583ab11f36ea141a66f5b925454791fe7f2f21738;
• training report: 1bb8861ce8567215034936860c1ce39c376b27b12f5b3be2bc55c9d4339519dd.

Confirmation result

Official evaluation job 744089 completed cleanly on H100 evc31. The learned source compiler produced 0/256 exact fresh programs. Every failure was declaration operation differs: the failed learned SET mapping invalidated every program before a sealed packet existed. In contrast, the immutable separate source ceiling compiled 256/256 programs.

Because no IEM1 packet was valid, no fresh evidence receipt, query binding, factorized execution, or protected joint recovery received an evaluable denominator. The evaluator therefore reports all matched arms as zero:

The table's denominator is the frozen board size; internally the end-to-end loop evaluated zero valid IEM1 packets. The separate source compiler's independent 256/256 result and the previously frozen NVE1 256/256 result remain the valid modular ceilings. Fresh IEM1 evidence/query generalization must not be described as passing: their training sets fit exactly, but source failure prevented confirmation evaluation. Vacuous zero-denominator integrity flags are not positive evidence.

Joint-training regression controls also fail closed: the IEM1 source interface is 0/1,024 on protected TOL3 semantic programs, and the IEM1-composed path is 0/256 on the protected NVE1 board. No invalid receipt, false commitment, malformed packet, gold deletion, or overflow is accepted because execution is never entered.

Official H100 evaluation SHA-256 is afe52c7ec3f68da8aa01b5999a1479ff13aed50f268aec1efb884a309e9804bf. An independent CPU run gives identical component counts, matched arms, controls, protected regressions, and promotion conditions; its diagnostic SHA-256 is 9ee1baef29cd53e11b8817dd3750f3d9e46a2cabdd3bf4a63b8a0a8986cef674.

Read-only stage-ownership attribution

One post-result diagnostic changes no weight or gate. It replaces only the failed IEM1 source path with the immutable TOL3 source owner, then applies the unchanged IEM1 evidence and query paths to those qualified packets:

• source programs: 256/256;
• IEM1 evidence receipts: 3,072/3,072;
• fully sealed episodes: 256/256;
• IEM1 natural query transactions: 280/768 exact;
• sensitive end-to-end answers: 256/256.

This does not rescue IEM1. It localizes the catastrophic end-to-end failure to shared source semantic ownership and independently shows that the general query owner is not qualified. The evidence owner and sensitive late-read path remain useful initialization evidence for a stage-owned successor. Diagnostic SHA-256 is d4b8f208836d9aa6742fb0ee663b936d6a9f8d4e520a349ed52e0f043e239a3c.

Decision and architectural implication

IEM1 fails the first component condition by the maximum possible margin. The frozen conjunctive promotion gate fails. Per preregistration, there will be no IEM1 width, duration, seed, transport, renderer, optimizer, or loss variant.

The negative is specific and useful. The factorized runtime did not fail; execution was never reached. Evidence and query ownership were learnable in isolation and fit jointly, while anonymous shared semantic transport lost a single mandatory structural primitive and thereby destroyed every complete program. A universal encoder plus symmetric latent transport is therefore too brittle for transaction-critical semantic ownership.

The next architecture must preserve complete, qualified specialists and join them through a typed, provenance-bound transaction bus. Source declarations, state updates, predicates, evidence receipts, and late queries need distinct owners with explicit contracts, while one model-owned controller coordinates their recurrent use. This is not an IEM1 repair: it rejects universal shared encoding and anonymous transport as the integration mechanism. The next gate must first prove that one composite checkpoint retains the immutable TOL3 and NVE1 component ceilings before measuring autonomous composition. No continuation pretraining is authorized by this result.