R12 Result-Digit Motor Exploratory Evaluation Contract

Protocol: R12-RDM-DEV-v1-disclosed-patch

Frozen: 2026-07-17 19:45 EDT, while Newton job 692235 was still in feature extraction. No teacher-forced development metric, autonomous score, or cycle score had been printed or read. This contract does not change the running source bytes.

Status: exploratory rejection/advancement gate only. It can never promote an architecture or support a broad reasoning claim.

Bound execution

The motor's disclosed training-source manifest is dd94a00be3984d2f43d6342459ba9749e3164624bb9b96977bd19c11f23845f0. The evaluator records this separately from its own source contract and requires the explicit noncanonical --allow-evaluator-patch authorization. The output must be newly created at artifacts/r12/result_digit_motor_r2_eval/eval.json.

Parameter contract

The frozen base has 125,081,664 unique parameters. The 576 -> 4096 -> 4096 -> 10 motor has 19,185,674, for a system total of 144,267,338 and 5,732,661 parameters of strict headroom below 150,000,000.

Locked interpretation

Pool rates over all reported regimes by summing numerators and denominators. All percentage-point differences below are absolute.

The exploratory treatment advances to a new hardened confirmation evaluator only if every gate passes:

1. Mechanical: job exit 0:0; all bound hashes match; output is complete, canonical JSON; dead teacher-forced metrics equal base; non-DWS checks report no motor firing and decoded identity on every listed prompt.
2. Held-out actuator: treatment teacher-forced digit top-1 is at least 99%, at least 3 points above base, and at least 10 points above shuffled.
3. First autonomous transition: treatment exceeds both base and shuffled by at least 5 points.
4. Full autonomous state loop: treatment exceeds both base and shuffled by at least 2 points.
5. Frozen-cycle first transition: treatment exceeds both base and shuffled by at least 5 points.
6. Routing: treatment and shuffled fire exactly once for every reported grammar-site opportunity; dead is behaviorally equal to base on every aggregate autonomous and cycle count.

Failure of any gate rejects this wide motor as an autonomous actuator despite its training-board fit. There is no threshold retuning after the report opens.

Mandatory confirmation after an exploratory pass

Even a complete pass above is not promotion evidence because this evaluator:

• rolls only the normal autonomous branch rather than a paired normal and counterfactual intervention;
• retains only bounded transcript samples rather than complete raw rollout evidence; and
• checks off-site preservation by decoded strings, not exact token IDs and per-boundary full-vocabulary logits.

Any successor must freeze a new secret confirmation board before extraction, score both branches and paired intervention, retain complete row-level evidence, require exact token and logit trajectory identity off-site, include parameter-matched generic and shuffled controls, and remain host-ALU-free. A positive result would demonstrate a grammar-gated local serialization actuator, not general reasoning or a new computational primitive.