DIVERGE-NLS1 Neural Episode-Law Synthesizer

Status: frozen after EAL2 confirmation and before NLS1 data materialization or neural scoring.

Capability hypothesis

EAL2 confirms the full observable-semantic path but still uses exact support intersection over 25 coefficient rows to induce each episode-local law. NLS1 changes exactly that owner:

A permutation-invariant neural synthesizer can consume three complete before/after transactions from the qualified EAL2 reader, infer each unseen episode-local law, and recurrently execute held matrix combinations without an exact support solver.

The confirmed 397,250-parameter EAL2 reader and its checkpoint remain frozen. NLS1 adds one shared value embedding, one demonstration encoder, sum pooling, and two categorical row heads. Demonstration order cannot identify a law. The synthesizer exposes the bounded 25-row output vocabulary but contains no support intersection or oracle transition. The already-qualified typed EAL2 executor remains exact and frozen; NLS1 changes only episode-law induction.

This is a scaffold-removal test, not a novelty claim for set encoders, hypernetworks, categorical program synthesis, or recurrent execution.

Frozen data and schedule

Training contains 100,000 typed three-demonstration law episodes under seed 2026080781. Matrices come only from the existing EAL2 training partition; all individual coefficient rows occur in both train and held-out partitions, while complete 2x2 matrices are disjoint. Development uses 256 fresh EAL2 episodes under seed 2026080782. Five conditional confirmation boards use seeds 2026080783--2026080787. Sources, aliases/registers, and episode identities must be disjoint from EAL2 training/development/confirmation and between every new board. Every artifact must regenerate byte-for-byte before training.

Two matched 216,946-parameter synthesizers start from the same state and train for exactly 500 AdamW updates, batch 2,048, peak learning rate 0.003:

• treatment receives the true three before/after transactions;
• shuffled-outcome control receives the same before states and labels, but after states are shifted between examples inside every identical minibatch.

Both arms receive the same sampled row order, optimizer schedule, update count, parameter count, and charged examples. NLS1 does not retrain EAL2.

Frozen development gate

All conditions are conjunctive:

1. inherited EAL2 normal and temporal-counterfactual complete reading are at least 99%, and temporal scrub is at most 30%;
2. treatment coefficient-row, terminal-state, and late-query exactness are each at least 99%;
3. the treatment terminal-state floor is at least 95% at every held depth from 12 through 32;
4. temporal-counterfactual terminal-state exactness is at least 99%;
5. shuffled-outcome-model and after-value-scrub terminal-state exactness are each at most 5%;
6. one-example terminal-state exactness is at most 20%, because one transition does not identify both coefficients;
7. temporal-scrub terminal-state exactness is at most 10%;
8. initialization, parameters, data, update count, batch, and optimizer schedule are matched; checkpoint/report custody is exact; and
9. source deletion and runtime-source audits pass.

A development miss closes NLS1 without width, embedding, update, seed, learning-rate, renderer, threshold, or duration variants. A pass opens the five already-built confirmation boards exactly once with the same frozen reader and synthesizer checkpoints. NLS1 does not authorize continuation pretraining or an open-domain reasoning claim.

Development result

NLS1 is closed as a conjunctive FAIL. Jobs 744720 and 744721 trained the matched treatment and shuffled-outcome arms from identical initialization. Treatment reached 100% on its fixed 10,000-matrix training sample; the shuffled control reached 0.53%. Development job 744722 then measured:

• treatment coefficient rows, terminal states, late queries, and every held depth: 100%;
• shuffled-outcome model, outcome scrub, one example, and temporal scrub: 0% terminal-state exactness;
• inherited normal and temporal-counterfactual readers: 100% complete; inherited temporal scrub: 28.4668%;
• temporal-counterfactual compiled execution against the unchanged original- world assessor: 0%, failing its frozen 99% condition.

The failed condition was specified incorrectly for this intervention. The counterfactual renderer swaps BEFORE and AFTER semantics around fixed numeric values, thereby defining reversed demonstrations and generally a different or unrepresentable inverse law. Exact reading therefore cannot imply the original terminal state. This diagnosis does not change the preregistered result: confirmation remains unopened and no NLS1 variant is authorized. The exact normal-path result remains development evidence that a neural set synthesizer can replace support intersection on this bounded carrier, not a qualified claim.

Development report SHA-256 is f5500c41ef71e30644031aae31f2e07041b5e712ad68aeafa11ad5e581bff241. Treatment/control checkpoint SHA-256 values are dfbabe8c5993f0cf0dc9a0d1f370cf3292d1ce8bc0b7bd4215205c923506f99f and 14184513a2a01be6f4091886fb1d53350a966aa87b1e7f295623621731b60d41.