R12 Cursor-Conditioned Token-Tape Diagnostic

Status: frozen development-only candidate. It is not a confirmation, architecture promotion, novelty result, internal-cursor result, or reasoning claim.

Motivation

The final-token linear diagnostic memorized train cells but reached only 43.13% pre-final and 41.67% post-final on new operand/renderers, versus a 40% cursor shortcut and 0/192 exact groups. This rejects a simple renderer-invariant code at the final selector position. It does not test whether operation order remains distributed over the prompt token states.

Frozen data boundary

Use only the immutable confirmation-free development view and audit with SHA-256 values:

• view: 24abd93737be57c6792a1d44c8f2e3a28d7c5fbc1666b083383350f410ce6ec9;
• audit: 33fb4792ed0a8027d49de157c295cb9ba651cdd9c59ab5cfa04a71e99af8ea25.

The H100 process receives no source canary, source audit, tokenizer path, or confirmation row. Development contains all 24 operation permutations, so this diagnostic cannot establish unseen-permutation extrapolation.

Readout arms

All arms use frozen raw-260k token states and the externally supplied cursor.

1. Shared token tape: one query per cursor attends over valid prompt-token states; one shared linear decoder maps the attended state to five actions.
2. Cursor-specific token tape: the same attention, with an independent action decoder per cursor. This is a favorable upper bound.
3. Mean joint: masked mean pooling followed by the same cursor-indexed linear family used in the failed final-token diagnostic.
4. Source-only tape: one cursor-independent query and one shared action decoder.
5. Cursor-only: a five-by-five table with no token state.
6. Embedding-only shared tape: the identical 5,765-parameter shared probe over frozen input embeddings, separating lexical parsing from deep-state use.
7. Position-only shared tape: the identical probe over deterministic sinusoidal positions plus the real length mask.
8. Source-deranged shared tape: the identical probe receives a seeded derangement of source tapes while retaining each original label/cursor.
9. Raw and token-RMS pre-final shared tapes: fixed preprocessing controls for the primary train-standardized surface.

The shared arm has 5d + 5d + 5 = 5,765 trainable scalars at d=576. The cursor-specific upper bound has 5d + 25d + 25 = 17,305. These are ordinary attention/gating mechanisms. A pass means token-tape readability with an oracle cursor, not internal cursor state or autonomous reasoning.

Source-only and cursor-only are structural leakage bounds, not parameter-matched competitors. Embedding, position, source-deranged, raw, and RMS arms are the matched controls for the shared probe. Attention weights are descriptive and do not establish causal clause retrieval; no clause-swap claim is authorized by v1.

Optimization

• seeds 2026071602, 2026071603, and 2026071604 for the shared and cursor-specific deep-state families; all controls use 2026071602;
• 100 epochs, deterministic epoch shuffle;
• batch 256;
• AdamW, learning rate 0.03, betas 0.9/0.95, epsilon 1e-8;
• zero weight decay, gradient clip 5;
• primary train-valid-token feature standardization, plus frozen raw and per-token RMS controls;
• exact masks exclude right padding.

No output valve or full-vocabulary calibration is fit in this experiment. Model states, query norms, preprocessing statistics, attention entropy/peaks, and per-example development predictions must be preserved. The 1,152 train-derived mean/std scalars are reported as preprocessing state.

Decision

A replicate passes only with at least 99% train cell accuracy, 95% development cell accuracy, 90% exact five-action development groups, at least 95% on every development renderer, and at least 95% separately at each of the four non-DONE cursors. Source-only must be at most 20% and cursor-only at most 40%. The shared or cursor-specific family passes only if at least two of three fixed replicates pass. If fewer than two replicates reach 99% train accuracy, the family is optimization-inconclusive rather than a representation no-go.

A deep shared-state claim additionally requires its median development cell accuracy to exceed the embedding-only, position-only, and source-deranged matched controls by at least 10 percentage points. Raw/RMS results determine whether any apparent effect depends on train-derived feature scaling.

• A deep shared-tape pass supports only a shared attention-bottleneck v2 design.
• An embedding-only pass supports only a lexical controller design.
• Only a cursor-specific pass supports at most a hard-gated upper-bound v2.
• Only a mean-joint pass shows distributed aggregation without retrieval.
• If fitted deep families fail, stop the external-cursor/token-tape branch; if they do not fit train, report optimization-inconclusive.

Any v2 must be a new preregistration with held-out permutations and fresh operand/renderers. No v1 confirmation reuse is authorized.