← Complete research archive
Architecture researchPreregistered72 lines

R12 Separated Controller–Executor Bus (SCEB)

On Shohin-scale (~125M), multi-step exact arithmetic fails when a single autoregressive head must jointly (a) select the next operator/cursor/DONE and (b) emit the integer result. Evidence:

R12_SEPARATED_CONTROLLER_EXECUTOR_BUS.mdOpen original Markdown ↗

R12 Separated Controller–Executor Bus (SCEB)

Status: theory + implementation authorized under creative directive. Protocol R12-SCEB-v1.

Conjecture

On Shohin-scale (~125M), multi-step exact arithmetic fails when a single autoregressive head must jointly (a) select the next operator/cursor/DONE and (b) emit the integer result. Evidence:

  • SSC external schedule: 115/256 finals (executor works under control).
  • Typed v1: DONE 86%, rollout 16%, atomic 27% (controller format learned; value emission weak).
  • Typed v2 mixture: DONE 0%, rollout 0.8% (mixing native+typed destroys the controller).
  • Post-DRS residual: digit directions +31 Δlogodds (value lives in residual but is not reliably consumed across steps).

Conjecture C-SCEB (amended). Separating controller from executor yields a useful systems upper bound and localization clue. Host arithmetic does not count as internal model reasoning. Claims of Shohin execution require in-model serialization (residual motors) without external apply_op.

Realizations (ordered)

A. Host-executed typed bus (runtime; no new weights)

Model emits one typed step line. Host parses (op, args), ignores the model’s -> next integer, applies apply_op in Python, rebuilds the register prompt, repeats until done=1 or depth cap. Primary metrics: final exact, op-match rate vs gold schedule, host-corrected vs joint-LM baseline (v1).

Gates (256 heldout rollouts, v1 ckpt):

GateThreshold
Host-exec final exact≥50%
Advantage vs joint typed v1 (16.4%)≥+20pp
Op-match rate (emitted op matches schedule[cursor])≥80%

B. Stateful Residual Register (SRR) — architecture

Frozen GPT backbone. New module StatefulResidualRegister:

  • n_reg digit registers (default 8 decimal digits + sign), each embedding in d_model.
  • At write layer L_w (default 17): last-token residual → per-register digit logits; Gumbel-softmax / straight-through into register ids.
  • At read layers L_r ⊆ {17…29}: add Σ_i E[digit_i] to residual (broadcast last token or all tokens).
  • Aux loss: CE on register digits vs teacher state integer.
  • Optional: keep LM loss on typed step lines with host-corrected next values in the target (model not graded on arithmetic tokens).

Gates: beat host-exec-only if SRR adds ≥5pp on heldout without host at inference; else SRR is a representation aid, not an autonomy claim.

C. Dual-head (later)

Separate linear heads for {op,cursor,done} vs digit value; out of scope for v1 score unless A/B underperform.

Comparator / integrity

  • Init: immutable sft_typed_controller_v1_200k_r1/sft_ep1.pt or best_step200000.pt as stated per arm.
  • Never write train/flagship_out/.
  • Do not collide with ACW Track S custody.
  • No threshold shopping after scores land.