R12 Atomic COMMAND-Binding Oracle-Initial Diagnostic

Status

Completed negative on H100 job 725517. This remains an isolation experiment, never a native-reasoning claim.

Hypothesis

The frozen Shohin WORLD compiler emits an imperfect hard packet. Canonical initial-to-terminal edit labels therefore mix two causes: legal COMMAND edits and attempted repairs of WORLD information absent from the packet. The atomic editor receives only that packet plus COMMAND bytes. A repair whose source fact was erased is unidentifiable and encourages memorization or invariant fallbacks even when the fixed edit algebra is exact.

Treatment

Keep commit fb368f2, atomic architecture, fixed executor, tokenizer, protected checkpoint, reader, training rows, architecture seed 31, data seed 11, 1,000 updates, LR 3e-4, clip 1.0, action weight 1.0, delta weight 1.0, and 32-batch evaluator unchanged. Change only training initial state from autonomous to exact oracle WORLD packet. Canonical action labels then represent COMMAND effects without unrecoverable WORLD repairs.

Inference interfaces remain hard and source-deleted. No QUERY, answer, oracle program, host solver, verifier, or candidate selection reaches the editor. Oracle state is a training-only diagnostic intervention and disqualifies the run from native-reasoning promotion.

Decision Gate

The useful diagnostic is held-out oracle-initial terminal transport: value, relation, exact-packet, and COMMAND-sensitive movement must materially exceed the autonomous-trained editor. Fully autonomous scores are recorded but cannot be promoted. If oracle-initial transport remains invariant, reject the frozen COMMAND residual-to-edit interface. If it improves materially, retain atomic algebra and replace the frozen WORLD bottleneck with a jointly trained, stage-separated WORLD compiler before retesting native composition.

Decision: separate_command_edit_learnability_from_irrecoverable_world_packet_errors_before_redesigning_the_whole_stack.

Result

The exact runtime completed all 1,000 updates on evc49 in 11:55. Report SHA-256 is 071c1978ae7462e225fd016734f03022765613fa54aa9379db9ec9fd2bda204a. With oracle initial state, held-out terminal value accuracy fell from 74.344978% before optimization to 71.506550% afterward; exact terminal packets remained 0/512. With the autonomous initial state, value accuracy fell from 58.542576% to 55.431223%; exact packets also remained zero. Fully autonomous WORLD and COMMAND strict and margin-1 gates remained zero.

The strong autonomous_program_oracle_state readout is an unchanged control: it feeds the exact oracle terminal packet to the reader and was identical before and after training. It is not evidence that the atomic editor learned. The diagnostic therefore rejects the frozen layer-19 COMMAND residual as a sufficient edit-binding interface. The fixed atomic algebra is still exact, but its selector needs a less lossy source representation.

Decision: retain_the_exact_atomic_algebra_and_replace_frozen_residual_only_command_binding_with_a_direct_lexical_rail.