R12 Typed Controller Internalization
Status: ACTIVE EXPERIMENT. Capability improvement attempt from the strongest current foothold (source-scheduled raw-260k confirmation: 115/256 scheduled vs 9/256 whole / 16/256 direct). Not an R12 novelty claim. Documents a training protocol that steals the SSC taxonomy controller shape and ACW's discrete register discipline.
Protocol: R12-TYPED-CONTROLLER-v1
1. Starting from current best
Immutable confirmation 689542 established:
- local arithmetic executor exists under an external schedule (115/256);
- whole free-form decode scores 9/256 but reaches the answer in 45/256 before continuation/parser destruction;
- 214/256 whole responses loop; 0/1920 calls emit EOS;
- updater-likelihood for naive internalization is negative.
Taxonomy implication (verbatim intent): train a typed controller carrying
(state, next_operation, operand, cursor, done) with one op per transition and
an explicit DONE policy, evaluated as one uninterrupted model call.
2. Mechanism (stolen pieces)
| Source | What is reused |
|---|---|
| SSC taxonomy | controller fields, DONE/EOS policy, family set |
| ACW | hard discrete registers, one-write discipline, source-deleted state |
| DRS | single-transition competence as the executor substrate |
| Problem/Work renderer | frozen SSC format that worked |
Not reused: free-form CoT, latent soft tokens, broad teacher SFT mixes.
3. Interface
Single-line register file:
state=<int>; ops=<op> <arg> | ...; cursor=<int>
Model completion (one or more lines, then EOS):
<op> <arg> -> <next>; cursor=<int>; done=<0|1>
...
answer=<int>
Training mixes:
- atomic — one transition from gold state (keeps DRS/SSC executor);
- rollout — full multi-step episode from initial state to
done=1+ answer; - resume — mid-cursor continuation (forces cursor use, not bag-of-ops).
4. Gates (locked before scores)
From immutable best_step200000.pt, one epoch, isolated output:
| Metric | Floor |
|---|---|
| Typed held-out final accuracy | ≥ 0.35 |
| Typed − raw-direct on same held-out | ≥ +0.10 |
| Typed on frozen SSC question set (re-rendered) | ≥ 0.30 |
| Atomic step exactness on held-out | ≥ 0.70 |
Fraction of typed rollouts that emit done=1 before cap | ≥ 0.80 |
Fail any → reject. No threshold shopping.
5. Explicit non-claims
Passing shows only that a discrete controller contract can be internalized enough to compose the already-present executor without an external scheduler. It does not establish general reasoning, language bridging, or ACW Track C.
6. Artifacts
artifacts/r12/typed_controller_v1/
train.jsonl
heldout.jsonl
audit.json
sft/ # isolated checkpoint
eval_*.json
decision.json