DIVERGE-EIC1: Equivariant Identity Commit

Status: confirmed positive after one frozen matched development gate and one source-disjoint confirmation. EIC1 qualifies a typed semantic transaction owner; it is not yet an end-to-end or open-domain reasoning result.

Capability hypothesis

CGL1 proves that terminal-outcome labels causally control the fitted decision, but a standard candidate-entailment scorer can still represent a different role convention for each renderer. Applying an orbit product only after CGL1 training recovers SmolLM2 (768/768) but not Shohin (512/768). The missing operation is therefore not merely another inference sample. The identity equivariance must be the trainable model's forward computation.

EIC1 projects every two-candidate score into the sign representation of the candidate-swap group before the loss or commit can observe it. For raw candidate scores r(x) and the exact mention transposition g, EIC1 uses

e(x) = 0.5 * (r(x) + g^-1 r(gx)).

The same shared LoRA owner produces both terms. The model can fit only evidence that survives this projection in the correct physical-identity frame. It cannot commit one renderer-local convention on x and a different convention on gx. This is a changed computation boundary, not a CGL1 seed, width, duration, prompt, threshold, or post-hoc loss retry.

Matched arms

Run four independent arms with the immutable CGL1 public/supervisor bytes, order, seed, optimizer, one-pair epoch, LoRA geometry, parent, and evaluator:

1. Shohin EIC1 involution projection.
2. Shohin duplicate-forward control, which evaluates the normal view twice.
3. SmolLM2 EIC1 involution projection.
4. SmolLM2 duplicate-forward control.

The duplicate arm matches two backbone forwards and all trainable parameters without receiving the transposed view or candidate mapping. Frozen CGL1 is an additional historical control. The SmolLM2 treatment is a capacity ceiling; Shohin is the decisive protected-backbone arm.

Development gate

The already-open CGL1 development board rejects EIC1 unless the Shohin treatment satisfies every condition:

• normal assignment at least 765/768;
• every mode at least 254/256 and every renderer at least 127/128;
• mapped mention-swap assignment at least 765/768;
• context deletion loses at least 250 exact assignments;
• entity-renamed prompts, scores, and predictions are bit-identical;
• frozen-parent tensors are bit-identical; and
• mapped-swap accuracy exceeds the equal-FLOP duplicate control by at least 200 assignments without losing more than three normal assignments.

The evaluator reports the exact projection identity residual. A development miss closes EIC1 without seed, width, duration, layer, loss, prompt, renderer, or optimizer variants. A Shohin pass admits exactly one independently built, source-disjoint EIC1 confirmation board. SmolLM2-only success records a capacity floor and does not promote Shohin.

The confirmation generator is frozen before training at seed 2026080713. It builds 256 new exact programs from a 32-symbol opaque entity bank, six new query families with both clause orders, 768 balanced transactions, and 1,048,576 represented worlds. It must report zero source, query, identity, and entity overlap against the CGL1 training corpus, the opened development board, the prior PQI board, and the still-unopened CGL1 confirmation board. Its bytes may be staged read-only, but no EIC1 model may access them unless the Shohin development treatment and its matched-control comparison both pass.

Claim boundary

EIC1 is an exact typed identity commit mechanism, not open-domain reasoning. It may qualify a semantic transaction owner for later DIVERGE execution and verified branch-local plasticity. It does not authorize continuation pretraining, access the unopened CGL1 confirmation board, or rescue CGL1.

Frozen result

All four single-H100 arms complete exactly 1,563 updates over the same 50,000 paired sources. Every arm fits 100,000/100,000 true training assignments. The decisive development results are:

The Shohin treatment therefore gains 384 swapped assignments over its parameter- and forward-matched control while losing no normal assignments. Assessor job 744629 passes all six frozen conditions; receipt SHA-256 is 556a85e07d5a00ab2a50c3fafe2188017c4fc61fbcee5e4947d61e63a6d31acd.

Exactly one Shohin confirmation is opened after a separate hash-bound CPU authorization. Job 744633 scores 768/768 normal and 768/768 mapped swap on the independently generated board. Every mode is 256/256, every renderer is 128/128, arbitrary entity renaming is prompt/prediction/score bit-exact, source scrub is balanced chance at 384/768 with zero mean margin, the frozen parent is exact, and the projection identity error is 0.0. Confirmation result SHA-256 is 268d2b0e7fc263c0a2d2db990d040a25f87612698bf4a90f5dc6fef857441a5c. The selected checkpoint SHA-256 is cd86f7335b9b7df078acc4b17c4dbe6b8bf944daee40f5397b452a0b5fae8849.

This result rejects the renderer-local shortcut without increasing backbone capacity: the changed factor is the exact identity involution inside the trainable commit. The next admissible test is one end-to-end natural composition using this confirmed QUERY owner, the protected NVE1 EVIDENCE owner, the existing structural WORLD parser/executor, and the already-positive branch-local verified plasticity mechanics.