R12 Counterfactual Cursor-Action Mechanics Result

Status: persistent mechanics board admitted and independently audited. The optional final-block/head-zero Q-intervention path and separately serializable 192-parameter sidecar pass focused CPU and existing inference regressions. No score-bearing model fit or GPU job has been run.

The immutable board contains 600 cells, 120 sources, 180 adjacent-order pairs, and 24 renderer groups. The independent audit reproduced the frozen symbolic scores exactly:

It also passed all 96 finite-state transition assertions and all 320 exact query-folding assertions. Board SHA-256 is 02a202070efa45f14c4e53b7d7f532d98791c7eef9daf438b02d31cc0ec6ab95; audit SHA-256 is c64951a1369b3dd29ca7e651840e5644e8445c7236cf994c3e54f05ca4a844b2.

train/model.py accepts an optional Q delta at one named layer/head. With the argument omitted, old checkpoints still load strictly and the original path is unchanged. train/counterfactual_cursor_action.py supplies the frozen event FSM and a zero-initialized centered-three-bit projection. The production projection has exactly 3 * 64 = 192 trainable scalars; all base parameters are frozen. Focused tests cover strict-load/zero-delta parity, code geometry, event transitions, prompt-boundary alignment, gradient isolation, and cached decode/full-replay equivalence. Existing causal-KV, batched-generation, recurrent-inference, and masked-loss regressions also pass.

Neural-canary preflight

The disjoint generator and independent auditor now reconstruct these exact split geometries without loading a model:

Every four-pack block is a Latin rotation: each operand value appears exactly once under each operation, so operand magnitude has zero deterministic operation identity inside a split. Train, development, and confirmation use disjoint numbers and disjoint renderer IDs. The exposure contract separates prompt-row tokens from the cursor side-state and from gold-only labels. Seven generator/auditor tests mutate targets, tokenization, ordering, integer types, pair maps, exposure fields, hashes, evalgrams, and Latin balance; all pass.

The favorable controls now have executable mechanics: an eight-entry cursor table has 512 total and 320 reachable scalars, while the rank-one final-head text-cursor LoRA has 640 scalars. The 192-scalar source-only arm has only 64 gradient-active dimensions when its cursor is clamped, so total, reachable, and active capacity must be reported separately rather than called exactly capacity matched.

This data package remains a draft rather than frozen evidence. The exact typed loader, six-arm relation equations and matched forward counts, base-checkpoint and tokenizer binding, full-vocabulary plus restricted-label evaluator, and score-blind publication receipt are not implemented yet.

This is a mechanics result only. It does not show that Shohin uses the cursor, selects the correct action, executes arithmetic, carries state, or halts. A neural run remains forbidden until the matched-arm data generator, immutable split hashes, typed loader allowlist, six-arm trainer, independent evaluator, and score-blind result receipt are frozen in one committed implementation.