Shohin RCR1 Revision-Conditioned Routing Gate
Status: closed negative on development, 2026-08-09. Holdout remains sealed. A static final-four-layer router residual is not sufficient.
Decision boundary
MTR1 is closed on development at 204/1,289, only +13 answers over the
unchanged second pass. Its final-four-layer shared-attention adapter changed
the frozen OLMoE route distribution by only 0.002018 mean L1. RCR1 tests a
different causal hypothesis: revision may require changing which frozen
experts process the model-owned draft, rather than adding capacity only to
shared attention.
Frozen intervention
The host, revision data, split, prompts, target responses, optimizer, seed,
256-update budget, and exact development evaluator are inherited unchanged
from MTR1. The host is pinned
allenai/OLMoE-1B-7B-0125-Instruct@b89a7c4bc24fb9e55ce2543c9458ce0ca5c4650e.
Treatment replaces each of the last four frozen top-k routers with
logits' = logits + (alpha / rank) * tanh(B(A(h)))
where rank is 8, alpha is 16, A/B are the only trainable tensors, and the
base router and all experts remain frozen. Initial B=0, so update zero is
exactly the base router. The residual is bounded per expert logit.
The matched control uses rank-1 LoRA on the final four shared attention mixers. It has 65,536 trainable parameters versus 67,584 for routing, a 3.13% parameter difference. Both arms receive the same 342,896 target tokens and generation budget. No external solver, verifier, answer router, or teacher runs at inference.
Development gate
Development contains all 1,289 frozen MTR1 identities. Holdout remains sealed. RCR1 passes only if all conditions hold:
- router treatment exceeds unchanged second pass by at least 5 percentage points;
- router treatment exceeds the matched rank-1 attention control by at least 3 percentage points;
- correct counts do not regress on MATH, logic/science, or executable code;
- both arms cover all identities and retain exact protected parameter rules.
A miss closes this exact router intervention. No seed, rank, layer, alpha, duration, prompt, or threshold rescue is authorized. Only a pass may open the existing holdout and authorize transfer to a larger MoE.
Frozen result
| Arm | Correct / 1,289 | Accuracy |
|---|---|---|
| revision-conditioned router residual | 194 | 15.0504% |
| matched rank-1 shared-attention control | 191 | 14.8177% |
| unchanged second pass | 191 | 14.8177% |
| prior rank-8 shared-attention MTR1 | 204 | 15.8262% |
The router treatment adds only three answers / 0.2327 points over both the
matched control and unchanged pass, and remains ten answers below MTR1. Broad
domain deltas versus unchanged are math +2, logic/science +1, and code
0; retention passes while both magnitude gates fail.
RCR1 rules out this narrow mechanism: a small, token-local residual on late router logits cannot by itself create a useful revision operator when all experts remain frozen. It does not distinguish whether the missing factor is persistent draft-level state, earlier routing, expert-side revision capacity, or the active-capacity boundary. That attribution is the next read-only step.
Comparison SHA-256 is
eb1b304adbf634e141bd497c5cf4b3d60d8616655ec1a674b7c556e399ae4c35.
No full larger-MoE campaign is authorized from this result.