R12 Cursor Readout / Actuation Diagnostic

Status: frozen development-only diagnostic. This is not a selector confirmation, architecture promotion, reasoning result, or novelty claim.

Question

Cursor-action v1 failed because its final-block, head-zero Q intervention moved the five action logits by about 0.10 while the median gap to the full-vocabulary winner was about 2.54. The next experiment must distinguish two possibilities without tuning on the exposed v1 confirmation:

1. the frozen representation does not make operation order jointly readable with the supplied cursor; or
2. the representation is readable, but the Q-only path cannot write the result into vocabulary logits.

Allowed data and custody

Only the immutable train and development splits from the v1 canary may be used. A deterministic builder projects them into a separate hash-bound artifact. An independent auditor rederives every projected field from the source canary and rejects any confirmation key. The H100 process receives only this confirmation-free artifact and its audit; it is not given the source canary, source audit, or tokenizer paths.

The development-view and audit SHA-256 values are respectively 24abd93737be57c6792a1d44c8f2e3a28d7c5fbc1666b083383350f410ce6ec9 and 33fb4792ed0a8027d49de157c295cb9ba651cdd9c59ab5cfa04a71e99af8ea25.

This diagnostic may choose a pre-final versus post-final feature surface on development. A later score-bearing experiment must withhold operation-order permutations as well as use fresh operands and renderers after one implementation is frozen. V1 development contains all 24 order permutations, so a pass here establishes only oracle-cursor-indexed linear separability across new operands/renderers, not compositional extrapolation.

Readout family

For frozen source feature h in R^d and cursor c in {0,...,4}, the joint readout is

s_a(h,c) = <W[c,a], standardize(h)> + b[c,a].

It is a tensor-product linear probe with 5 * 5 * (d + 1) trainable scalars. It is deliberately favorable and diagnostic. It is a hard-gated linear mixture of experts with an externally supplied cursor, not evidence that Shohin represents a cursor internally and not an R12 primitive.

Controls are:

• source-only: one linear five-action readout shared across all cursors;
• cursor-only: a five-by-five bias table independent of source features.

The shortcut theorem bounds their deterministic development accuracy at 20% and 40%, respectively.

Readout and calibrated write requirement

1. Restricted readout: cross-entropy over the five probe scores. This tests oracle-cursor-indexed linear separability without vocabulary competition.
2. Frozen-readout calibration: freeze each restricted readout, then fit only two scalars in delta_action_logits = softplus(raw_alpha) * restricted_scores + beta. The loss is exact full-vocabulary cross-entropy. The non-action log-sum-exp and maximum are cached from the frozen base. This does not independently prove an internal actuation path: any positively separated finite readout can eventually beat a finite vocabulary margin by scaling. It instead measures the exact gain, common action bias, and intervention magnitude a direct write path would require.

The calibrated write is active only at the selector position in this diagnostic. It does not update a cursor, execute arithmetic, emit COMMIT, emit DONE/EOS in one call, or modify the flagship.

Frozen optimization

• base: immutable raw 260k;
• seed: 2026071601;
• 100 epochs;
• deterministic epoch shuffle;
• batch size 256;
• AdamW, learning rate 0.03, betas 0.9/0.95, epsilon 1e-8;
• zero weight decay;
• gradient clip 5.0;
• per-feature standardization from training sources only, standard deviation clamped to at least 1e-5.

Pre-final and post-final frozen features are both measured. No rank, layer, threshold, seed, or optimizer search is allowed under v1.

Development decision

Representation availability passes only if at least one joint restricted arm has all of:

• at least 99% train cell accuracy;
• at least 95% development cell accuracy;
• at least 90% exact five-action development source groups;
• source-only development accuracy at most 21%;
• cursor-only development accuracy at most 41%.

The calibrated full-vocabulary fit is reported on the same feature surface and must have the following before it can inform a v2 design:

• at least 95% full-vocabulary development cell accuracy;
• at least 90% exact five-action development source groups;
• at least 95% full-vocabulary cell accuracy on each development renderer;
• at least a 10 percentage-point cell-accuracy advantage over both valve calibrated controls.

The report must also preserve the frozen readout parameters, standardization statistics, scalar calibration, per-example development predictions, base and adjusted margins, delta magnitudes, runtime identity, and immutable input/code bindings. The gain and bias are descriptive measurements, not an actuation gate.

If representation fails, stop this branch. If it passes, the only authorized next action is to use the measured write requirement to design a matched direct-write v2 comparison of orbit/interchange training versus ordinary CE, with held-out permutations and fresh confirmation. No result here authorizes a reasoning, compositionality, novelty, or internal-cursor claim.