DIVERGE-CRP1: Causal Revision Packet

Status: closed after the one frozen neural gate. The packet causes substantial causal revision on scalar and register programs, but misses the absolute, matched-control, localization, joint, and cross-family promotion thresholds.

Capability hypothesis

VCR1 proved that a small recurrent prefix can turn incomplete SmolLM3 drafts into concise answers, but it did not revise a complete explicit wrong derivation. CRP1 tests a materially different mechanism: preserve a bounded version space over the location of the first invalid step, evaluate every location with tied recurrent weights, and commit one whole guarded repair packet before generating a correction and causally replayed suffix.

The packet has one categorical fault-line variable with values NO_ERROR and STEP_1 ... STEP_N. Candidate e receives separate source masks for the valid prefix before e, the proposed fault at e, and every dependent step after e. All candidates share parameters and remain separate through recurrence. The generator receives only the prefix from one hard-selected candidate; candidate fields are never averaged.

The exact matched control instantiates the same parameters, candidate identities, recurrent updates, training rows, targets, and FLOPs, but every candidate sees the complete trace in all three channels. A gain over this control is evidence for the causal guard structure rather than candidate width or recurrent compute.

Board

The deterministic board contains scalar arithmetic, noncommuting two-register programs, and symbolic string transformations. Train/development depths are 4--6. Evaluation depths are 7--9 and also hold out the renderer and value/width band. Every wrong trace is complete and contains exactly one verifier-created first error. Every later draft step is recomputed from that corrupted state, so the suffix is locally coherent but globally wrong. The correct answer is never equal to the wrong draft answer. Every trace also has a correct no-op twin.

Supervisor program objects and error certificates are stored for independent assessment but are never rendered to the model beyond the ordinary problem, complete draft, and final draft answer. The model must emit the first error, one corrected step, the replayed dependent suffix, and one boxed answer.

Frozen one-seed experiment

1. Build and hash 4,800 train, 480 development, and 480 OOD evaluation rows.
2. Run two-update guarded and unguarded mechanical smokes from byte-identical packet initialization.
3. On mechanical success, train each arm for exactly 200 updates, accumulation 8, at learning rate 3e-4. The SmolLM3-3B product generator remains frozen.
4. Evaluate the exact same 480 OOD identities under prompt-only, guarded, unguarded, reset, shifted-selection, and packet-swap conditions. Evaluate all three trained/plain arms on the 480 correct no-op twins.

Frozen promotion gate

All conditions are conjunctive:

• guarded wrong-trace exact answers at least 240/480;
• guarded beats prompt-only by at least 48 solves;
• guarded beats equal-compute unguarded by at least 24 solves;
• guarded autonomous packet localization at least 360/480;
• guarded joint localized-and-correct revision at least 192/480;
• guarded correct-twin answer and no-error packet preservation each at least 432/480, and no more than 12 solves below prompt-only preservation;
• guarded beats unguarded by at least five solves in each family;
• reset, shifted selection, and packet swap each cost at least 24 solves, and at least two cost at least 48;
• no arm exhausts the 384-token correction budget.

Failure closes this exact first-error packet without a seed, width, duration, loss, depth-band, renderer, or threshold repair. A pass qualifies only a bounded model-owned causal-revision mechanism. It then requires transfer to natural verified math/code/science traces before any broader DIVERGE claim.

Result

The exact board passed construction with 4,800 train, 480 development, and 480 evaluation rows, balanced across the three families. There were zero identity overlaps, malformed rows, or tokenizer truncations; the longest admitted row used 824 of 4,096 positions. Train, development, evaluation, and board-report SHA-256 values are:

• 33dacc5e0f9a72ad01eaa58bbee627c534437db95752b7e6a76a8488af0b2ede;
• ad27c4ca8f89a158fc1c4516e1cfce67f792bd5015f0b68e9c30ba39b62cfde5;
• db0bde0c22afe3d25f4f1f578249bf67156f4115d38a094c07f1ea36f6be6849;
• b9ead3db62c91622231cb310fd5cc6f48d814e69cf26b40ad195c8306a689d8d.

Both arms completed 200 updates from packet SHA-256 dc5f273d3fb629fccadbd9ac2209a6b897f2be48d00c9aef71b23116b6a5966e. Each trained 3,556,353 parameters on 191,201 charged response tokens, retained an identical frozen source, and used 12.45GB peak allocated CUDA memory. Guarded and unguarded training reports / checkpoints are:

• guarded: ff0419e26295125a5451603de078a8b25b7c2ebaa8ad902585b6d5f958000bc7 / 588dce4f608fde47516a8b29feedc40bf7ee58d2ff2aa8b344848915dcacb5ce;
• unguarded: 0a0353c450c245acd6f44aaa34974e5ff1caa718ea54c371056aad80e84f0f85 / 93e7e71db74e1f4efe68d13b028157cdaea449bb34e2f8e8bdc528af40ce4ced.

The autonomous OOD result is:

Guarded exact answers by family are 135/160 scalar, 78/160 register, and 0/160 symbolic, versus 115/160, 79/160, and 0/160 for unguarded. The net answer advantage is therefore only +19, concentrated entirely in scalar arithmetic. The causal-guard effect is much larger on coherent attribution: guarded creates 129 joint wins not achieved by unguarded and loses only six. Reset, location shift, and cross-example packet swap cost 213, 177, and 113 answers, respectively.

Transcript inspection confirms a real but bounded capability. Of 183 guarded joint successes, 181 exactly match the canonical complete correction target after whitespace normalization and 182 include the complete correct dependent suffix verbatim. Examples explicitly replace a wrong arithmetic/register step and replay every later state to the correct terminal value. This is not VCR1's truncation/finalization effect. It nevertheless does not generalize to the held-out symbolic family and does not create a large enough answer advantage over equal-compute generic recurrence.

The formal gate fails. Guarded misses 240 wrong answers, 360 packet localizations, 192 joint repairs, the +24 unguarded margin, and the positive advantage in every family. Some destructive controls and the prompt baseline also exhaust generation, failing the all-arms exhaustion check. Correct-trace preservation and all causal-drop checks pass. Gate report SHA-256 is cdd5a717e55cb3c589fecdefc7455a83903f9ef68376b822c3846f9da7573e8c.

Decision

Close exact CRP1 without a nearby seed, width, duration, loss, renderer, threshold, or prompt repair. Preserve its demonstrated first-error packet and the 182 audited complete replay successes as causal evidence. A successor must change the execution/readout substrate materially: packet localization must drive one persistent model-owned state replay rather than merely conditioning a frozen language generator, and it must transfer across numeric, register, and symbolic state domains before natural-trace promotion.