SER1: Selected-Expert Revision Residual

Status: closed negative on OLMoE, 2026-08-09. Holdout and larger-MoE transfer remain sealed.

Causal hypothesis

ECR1 failed because one shared transform dominated and its low-dimensional expert codes became causally decorative. SER1 removes that parameterization. The untouched frozen OLMoE router selects complete expert-owned residual transforms, so expert identity changes the computation rather than merely scaling a shared basis.

For each of all 16 sparse layers, hidden width h=2048, 64 native experts, native top 8, and expert rank r=1:

(logits, p, S) = frozen_router(h)
y_native = frozen_experts(h, S, p)
q = stopgrad(p / sum(p))
delta = sum_(e in S) q_e B_e A_e h
y = y_native + delta

Each A_e is [1,2048] and B_e is [2048,1]; every B_e starts at zero. The base router and experts remain frozen. SER1 has exactly 4,194,304 trainables total, but only eight rank-1 transforms are active per token. Its adapter compute is 32,768 MAC/token/layer.

Matched controls

1. Active-FLOP control: one shared rank-8 post-MoE residual at the same 16 locations, 524,288 trainables and 32,768 MAC/token/layer.
2. Total-parameter control: one shared rank-64 residual at the same locations, 4,194,304 trainables and 262,144 MAC/token/layer.

All three fits use the same pinned OLMoE revision, exact model-owned drafts, 9,651-row complete-context corpus, seed/order, 256 AdamW updates, LR 2e-5, 338,620 charged target tokens, and unchanged autonomous evaluator. No source, draft, or target truncation is allowed. Routers and native experts must expose zero trainable tensors.

Staged gate

One treatment mechanics update must be finite, preserve protected hashes, and match the exact parameter/FLOP receipt. Then Stage 1 trains all three arms and evaluates the full 1,289-row development board. Promotion requires jointly:

1. SER1 at least 280/1,289;
2. SER1 at least 26 answers above the stronger new shared control;
3. no regression below math 55, logic/science 180, or code 5;
4. complete parameter, token, latency, memory, load-entropy, mean token- entropy, expert-use, and residual receipts.

Only a Stage-1 pass may run the same checkpoint under zero, mean-bank, and within-layer expert-bank permutation interventions plus a true draft- unavailable fit. Normal must beat every intervention and draft-unavailable by at least 13 answers. These Stage-2 controls may authorize one sealed holdout; Stage 1 alone cannot. If Stage 1 fails, close SER1 without rank, depth, duration, seed, or loss variants.

Interpretation is fixed. Beating the active-FLOP control but not the parameter-matched control means sparse storage has not earned a capability claim. Beating both controls and collapsing under expert-bank permutation supports useful expert-specific revision computation. Failing either score gate closes the mechanism.

Result and decision

Mechanics passed exactly. Treatment has 4,194,304 trainables and 32,768 MAC/token/layer with zero trainable native router/expert tensors. All three fits completed the same 256 updates and 338,620 charged target tokens.

Full development scores are selected-expert treatment 201/1,289 = 15.5935%, active-FLOP shared control 236 = 18.3088%, and total-parameter shared control 241 = 18.6967%. Treatment domain counts are math 56, logic/science 139, and code 6; logic fails its floor. Exact comparison SHA-256 is a44fc09f3e46db195afc0bf32f58e05db0a5e7e030510678c4dbe80bcb7d1fb6.

SER1 is therefore closed without Stage 2, holdout, or nearby rescue. Native expert assignment fragments revision learning and is substantially worse than both shared controls. This rejects post-MoE expert-owned residual banks; it does not test new dedicated revision experts or adapters inside native expert nonlinearities.