DIVERGE-EWC1: Equivariant WORLD Compiler

Status: development FAIL. Confirmation remains unopened; EWC1 is closed without local variants.

Pre-score implementation note: the first deterministic build rejected one depth-28 row because the original 512-byte tensor cap was too short. It wrote no train, development, or confirmation file. The cap alone is corrected to 768 bytes before rebuilding; seeds, rows, renderer assignments, depths, models, training schedule, evaluator, and gates remain unchanged.

The next pre-score overlap audit rejected revision 1 because one opaque alias (kaquyekeko) collided between training and confirmation; source and identity overlap were both zero. That rejected corpus is preserved. The deterministic nonce length alone is increased from 10 to 14 letters for revision 2, before any model training or score. Seeds, semantic examples, split assignments, and all gates remain unchanged.

Capability hypothesis

DIVERGE-NPL2 reaches 85.6104% source-deleted late-query exactness, exactly matching its oracle-semantic PL1 ceiling, but its WORLD transaction is still produced by one exact regular expression. EWC1 changes only that boundary:

A learned byte-level compiler can bind source values to episode-local register identities and select the ordered operation mentions from natural WORLD text, while exact permutation equivariance prevents renderer-local or absolute-name conventions.

NPL2, NVE1 EVIDENCE, EIC1 QUERY, the 64-scalar plastic policy, executor, and verifier remain unchanged and hash-protected. EWC1 is not a continuation- pretraining experiment.

Model-owned boundary

A generic lexer exposes only:

• unsigned integer spans;
• exact occurrences of the eight already-declared operation aliases; and
• exact occurrences of the two already-declared register names.

The lexer does not know which numbers initialize state, which alias mentions are executable rather than distractors, their operation meaning, or the final answer. A two-layer bidirectional byte GRU must:

1. assign one distinct numeric mention to each declared register; and
2. classify every alias occurrence as executable or irrelevant.

Selected alias occurrences are emitted in source order and mapped to the episode-local alias table. Register-value scores are pairwise functions of numeric and register mention representations. Reordering the register table therefore permutes the output rows exactly, with no learned absolute register embedding in the treatment path. Alias occurrence scoring shares all weights and pools same-identity occurrences without alias-index parameters.

The matched absolute-role control has the same encoder, heads, parameter count, data, updates, seed, optimizer, and batch size. It replaces the two source-derived register keys with two learned absolute keys. This preserves normal capacity but removes exact register-order equivariance.

Data split

The frozen generator contains eight initial-state clauses and eight ordered- program clauses. Their 64 Cartesian pairs are assigned by a fixed arithmetic partition to train, development, or confirmation. Every phrase family occurs across the campaign, but development and confirmation hold out complete clause compositions. Clause order, distractor placement, initial values, alias sequences, depths, aliases, and register names vary independently.

• training: 50,000 rows, seed 2026080721, depths 3--20;
• development: 4,096 rows, seed 2026080722, depths 3--28;
• confirmation: 4,096 rows, seed 2026080723, depths 3--28.

Confirmation bytes are generated and overlap-audited before model scoring but remain unopened unless the single development gate passes. Exact source, identity, alias, and register overlap between every pair of splits must be zero.

Frozen controls

Evaluation reports complete typed-WORLD exactness normally and under:

• register-table transposition, mapped back to the original physical frame;
• a fixed eight-alias table permutation, mapped back;
• deterministic unseen entity renaming;
• context scrub preserving only candidate numbers and declared symbols; and
• the equal-parameter absolute-role control.

The model never receives initial_state, symbols, numeric_targets, or operation_targets in its forward call. Those fields are assessor-only.

Development gate

The one treatment and one matched control are trained for exactly 1,000 updates, batch 256, AdamW learning rate 0.003, cosine decay, and seed 2026080721. Development passes only if all conditions hold:

1. treatment train typed-WORLD exactness is at least 99%;
2. normal joint exactness is at least 99%, initial state at least 99.5%, and operation sequence at least 99%;
3. every held-out renderer pair is at least 95%;
4. mapped register-order, mapped alias-order, and unseen-rename exactness are each at least 99%;
5. source scrub joint exactness is at most 20%;
6. treatment/control parameters, data, seed, schedule, and optimizer match;
7. control train exactness is at least 99% and normal development exactness is at least 95%; and
8. treatment mapped register-order exactness exceeds control by at least 25 percentage points.

A development pass opens the frozen confirmation exactly once. Confirmation must independently satisfy conditions 1--5 and all receipts. A miss closes this exact EWC1 mechanism without width, duration, seed, prompt, threshold, renderer, loss, or optimizer variants.

Conditional integration

Only a confirmed EWC1 checkpoint may replace parse_program_surface inside the protected NPL2 runtime. The integration must compile every acquisition and transfer WORLD before any hidden outcome is observed, delete source bytes, and reuse the unchanged NVE1/EIC1/PL1 path. It passes only if WORLD semantics remain at least 99.5%, late-query exactness remains at least 80%, stays within five points of oracle PL1, and all existing NPL2 destructive controls remain qualified.

Even a complete pass establishes controlled learned natural-WORLD reasoning, not unrestricted prose, public benchmark reasoning, or authorization for a long pretraining run. The exact executor and verifier remain the next engineered boundary.

Result

Matched one-H100 jobs 744648/744649 complete in 7m29s. Both 582,530- parameter arms fit all 50,000/50,000 typed WORLDs exactly, use 1.159 GB peak allocated VRAM, and process about 261k source bytes/s. Treatment and control checkpoint SHA-256 values are 0816ed1c65d51951cf11625f505ea7886a176c008cc42ac3ac821f369788d1ed and 208ec08c0d4a4d26853724e0bbf384b64252ca75d4c193350f92d5691d4c8829.

Development job 744650 obtains:

Every held-out renderer is 100%, and every frozen gate passes except source scrub, whose ceiling was 20%. EWC1 is therefore a strict FAIL. It learned a highly accurate and exactly equivariant structural compiler, but operation selection still recovers 84.9% when all language context is destroyed. The punctuation/position geometry remains a shortcut, so the result does not qualify as language-grounded WORLD semantics. Development report SHA-256 is 2e173969609a4d50f7fb169b18d56611033a28a03c7c71d618c873a9ce1e5760.

Confirmation 744651 and conditional NPL2 jobs 744652--744657 never open and are canceled. Preserve EWC1 only as a frozen structural extraction baseline. The successor must make two or more complete candidate worlds positionally symmetric and use counterfactual semantic role language to commit one coherent world; changing EWC1's threshold, renderer, duration, width, seed, or loss is not authorized.