R12 Cursor-Conditioned Token-Tape Result

Decision: external_cursor_token_tape_no_go

Claim boundary: At raw step 260k, pre-final token states are not renderer-invariantly readable by the frozen single-query attention plus linear decoder family under its specified optimization recipe. This does not close all token-tape or external-cursor mechanisms: post-final, multi-query, nonlinear, and order-level access were not tested. It does not authorize a reasoning, internal-cursor, retrieval, compositionality, actuation, or novelty claim.

Custody

• implementation commit: 2778d7999ce3539866c2df80d6a5dc4f975361af;
• raw 260k base SHA-256: 91d5288f184fc5230516add9851ac1a8815d3369ffd816cd7d0c03d8bafc741d;
• confirmation-free development view SHA-256: 24abd93737be57c6792a1d44c8f2e3a28d7c5fbc1666b083383350f410ce6ec9;
• independent view-audit SHA-256: 33fb4792ed0a8027d49de157c295cb9ba651cdd9c59ab5cfa04a71e99af8ea25;
• runtime SHA-256: af7da54fd23ac1f7a64766438ba72d14591ae96d495da2306cc535da875d7f7c;
• Newton job 689976, completed on evc44 in 87 seconds with exit code 0;
• preserved Slurm log: logs/r12_cursor_token_tape_689976.out, 964 bytes, mode 0444, SHA-256 246976f5177e9e790ebf3e3208e9ad7aafb57c298a550e13b77bf379f0a58b42;
• immutable result: artifacts/r12/cursor_token_tape_dev_v1.json, 9,815,027 bytes, mode 0444, SHA-256 7065401b13fd83b8a5b514be9a9b2a8cd5158af39abfa5464df43b368bd825e1.

The H100 process received only the 5,760 train and 960 development cells. It received no source canary, source audit, tokenizer path, or confirmation row. The job reconstructed all source from the frozen commit, staged every input node-locally, re-hashed it before load, removed ambient Python and preload variables, and recorded the actual Python, PyTorch, dependency, and model module paths. The result preserves preprocessing vectors, probe states, query norms, attention summaries, and every development prediction.

Scores

The shared deep family fits train in all three seeds, so this is not an optimization-inconclusive result. Its median development accuracy is 48.33%, only 8.33 percentage points above the best matched control and below the frozen 10-point margin. No shared or cursor-specific replicate approaches the 95% cell, renderer, and per-cursor gates or the 90% exact-group gate. Source-only and cursor-only stay at their nominal ceilings, while source derangement falls below the cursor shortcut.

The source-only ceiling is not a valid leakage detector here. A cursor-independent unique prediction necessarily matches exactly one of five cells because every source contains each target once. Its 20% result is therefore mathematically automatic. Cursor-only and the parameter-matched embedding, position, and source-deranged shared arms remain informative, but the cursor-specific deep family lacks cursor-specific matched controls.

Mechanistic diagnosis

The cursor-specific family's descriptive 56.98--57.81% score is concentrated in the first instruction and deterministic DONE: cursor-zero accuracy is 72.40--81.25%, DONE is 100%, but cursor one is 36.46--37.50%, cursor two is 21.88--29.69%, and cursor three is 44.79--47.92%. Only 3--11 of 192 sources recover the full four-operation order in any seed. Because the shared family misses its frozen ten-point matched-control margin and the cursor-specific family has no parameter-matched controls of its own, this lift is descriptive; it is not an authorized deep-representation finding.

The observed asymmetry is consistent with partial first-clause access and poor later-clause routing, but the present experiment cannot distinguish that story from unmeasured cursor-specific lexical or positional shortcuts. The 99% train fit and poor renderer-held-out development performance do reject a stable factorization through this exact probe family.

Raw and per-token-RMS controls reach only 58.07% and 96.30% train accuracy, so neither satisfies the 99% train-fit condition. They are optimization- inconclusive and do not resolve whether the standardized-arm behavior depends on train-derived feature scaling.

Decision and next boundary

Under the frozen preregistration, this pre-final single-query probe family stops. Do not tune it on the exposed development result or present its lift as a reasoning mechanism. Development includes all 24 permutations, so it cannot support unseen-permutation extrapolation in any case.

A branch-wide representation conclusion would require a newly preregistered fresh-data diagnostic that includes post-final tapes, a 24-way order readout, cursor-specific embedding/position/deranged controls, and train-only renderer holdouts for optimization selection. Gold-span layerwise paired-transposition readout is the smallest diagnostic that can separate operation erasure from failed routing. It cannot reuse this development split as confirmation.

Independently, any capability mechanism must create and causally update its own sequential state rather than assume an oracle cursor. Before architecture code or H100 training, the R12 invention charter still requires an explicit state-transport theorem, an equivalence dossier, an exact collapse test, and a finite synthetic falsifier with held-out permutations and fresh renderers.