← Complete research archive
Architecture researchPreregistered57 lines

RSOT1: Role-Separated On-Policy Transduction

OCET1 closes at 1690/1908. Its aligned trained owner generates only 1,304 correct mandatory-rewrite proposals, then improves those same proposals to 1,690 after the commit transition. The on-policy objective therefore learned a causal repair policy (+386 within the aligned trajec…

docs/research/SHOHIN_RSOT1_ROLE_SEPARATED_ON_POLICY_TRANSDUCER.mdOpen original Markdown ↗

RSOT1: Role-Separated On-Policy Transduction

Status: prospectively frozen after OCET1 development failure, before RSOT1 output, 2026-08-10
Scope: opened source-disjoint development board; holdout sealed

Evidence and hypothesis

OCET1 closes at 1690/1908. Its aligned trained owner generates only 1,304 correct mandatory-rewrite proposals, then improves those same proposals to 1,690 after the commit transition. The on-policy objective therefore learned a causal repair policy (+386 within the aligned trajectory), but sharing that adapter between proposal and commit damaged the previously qualified proposal role. The controls remain far lower: permuted 940, hidden 623.

RSOT1 tests one structural correction, not a training variant: preserve the immutable qualified DSET proposal owner and assign the OCET-trained owner only to commit. The two roles share the same frozen Qwen3.6-35B-A3B base but use separate 1,179,648-parameter adapter states. No new weights or updates are authorized.

Frozen execution

  1. Reuse the exact eight immutable aligned FRET1 proposal shards from the qualified DSET owner. They cover all 1,908 development presentations.
  2. Run three commit owners over the byte-identical proposals:
    • aligned OCET owner;
    • within-pair label-permuted OCET control;
    • exact draft-hidden OCET control.
  3. Use the same prompt, generic edit executor, greedy decode, 32-token budget, source board, tokenizer, and evaluator as OCET1.
  4. Report proposal, commit validity, final correctness, choice/numeric, clean/fault, exhaustion, hashes, wall time, and compute.

The runtime receives no verifier, answer label, solver, or host repair. The generic executor only applies the model-emitted KEEP or REPLACE_LAST script.

Frozen gate

All conditions are conjunctive:

  • aligned final >=1874/1908;
  • choice final >=220/256;
  • clean final >=945/954;
  • fault final >=859/954;
  • valid commits >=95%;
  • aligned is at least 13 answers above ISET1 (1838), permuted, and hidden;
  • zero aligned decode exhaustion.

One pass opens one sealed holdout. A miss closes this exact role-separated system without adapter rank/layer/LR/update/seed/prompt/decode variants.

Claim boundary

Separate proposal and verifier/editor roles are not novel. RSOT1 is a practical architecture test of whether Shohin's model-owned on-policy repair capability becomes operational once destructive cross-role parameter sharing is removed.