← Complete research archive
Architecture researchClosed / no-go137 lines

ECR1: Expert-Conditioned Revision Residual

DREM1 is preserved as an unmatched upper-bound diagnostic, not a capability gate. Its controller is recurrent across controlled layers but static across generated tokens; its old draft-masked arm removed only a pooled draft vector; it added a full frozen-backbone prepass; it had …

docs/research/SHOHIN_ECR1_EXPERT_CONDITIONED_REVISION.mdOpen original Markdown ↗

ECR1: Expert-Conditioned Revision Residual

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

Why DREM1 is not decisive

DREM1 is preserved as an unmatched upper-bound diagnostic, not a capability gate. Its controller is recurrent across controlled layers but static across generated tokens; its old draft-masked arm removed only a pooled draft vector; it added a full frozen-backbone prepass; it had about 6.84M trainables versus MTR1's 524,288; and its 1,024-token default risked left truncation. DREM1 job 747085 was canceled before allocation and charged zero H100 time.

Fixed evidence and hypothesis

The host, data, drafts, prompts, decoding, and evaluator remain pinned to OLMoE-1B-7B-0125-Instruct at revision b89a7c4bc24fb9e55ce2543c9458ce0ca5c4650e. Unchanged scores 191/1,289; MTR1 shared-attention rank 8 scores 204; router-only RCR1 scores 194. At least 11 of MTR1's 16 repairs are serialization-only. Holdout is sealed.

ECR1 tests whether frozen native expert identity contains useful information for a small same-pass correction that generic post-MoE capacity cannot recover. It neither changes native routing nor adds a model prepass.

For each of the final four sparse layers, hidden width h=2048, 64 experts, native top 8, and rank r=31:

(logits, p, S) = frozen_router(h)
y_native = frozen_experts(h, S, p)
q = stopgrad(p / sum(p))
c = sum_(e in S) q_e tanh(C[e])
delta = B((A h) * (1 + c))
y = y_native + (alpha/r) delta, alpha=r

A is [31,2048], B is [2048,31], and C is [64,31]. B=C=0 initially, so initial behavior is bit-exact with the frozen base. Only A/B/C train. Total trainables are exactly 515,840. Additional work is 127,224 MACs (254,448 FLOPs) per token per controlled layer.

Matched arms

  1. ECR1 treatment: equation above, final four sparse layers.
  2. Shared residual: same post-MoE location, delta=B A h, rank 32, exactly 524,288 trainables and 131,072 MACs/token/layer.
  3. Draft-unavailable ECR1: identical ECR parameters, input IDs, positions, and training rows; draft-token keys receive zero attention in the complete second pass. This is a full-model causal key mask, not pooled-vector zeroing. Generation passes explicit original prompt position IDs so internal zero attention entries cannot compress positions.

Every arm receives 256 AdamW updates, LR 2e-5, seed 2026080901, and data seed 2026080814. Exhaustive pre-output custody found four copies of one duplicate source identity require 4,332 tokens, beyond pinned OLMoE's hard 4,096 context. Those four occurrences are deterministically excluded from all new arms before training; the remaining 9,651-row population is otherwise unchanged and its longest row is 2,616 tokens. No source, exact draft, or target token may be truncated. The admitted corpus and exclusion report are immutable and hash-bound before fits. This four-row difference from historical MTR1 is a declared limitation; the decisive ECR-versus-shared and draft-causality arms use identical admitted rows.

Frozen gates

Development promotion is conjunctive:

  1. treatment at least 256/1,289;
  2. treatment at least 39 answers above the stronger newly trained control;
  3. no regression versus unchanged in MATH (>=40), logic/science (>=145), or MBPP (>=5);
  4. treatment at least 13 answers above draft-unavailable treatment;
  5. at least 25 repairs not certified serialization-only, and possible-semantic repairs minus semantic breaks at least 20;
  6. normal expert codes beat each of zero, mean, and deterministic within-layer permutation by at least 13 answers. If permutation does not hurt, expert identity is decorative and the expert-conditioned claim fails;
  7. complete protected hashes, source/draft retention, parameter/FLOP, peak memory, latency, generated-token, native route, and residual receipts.

Metrics distinguish aggregate load_entropy from mean per-token entropy and also report top-8 set/order overlap, top-8/top-9 margin, later-layer route drift, residual norms by outcome class, expert-code cosine/effective rank, and expert utilization. Routers and experts must have zero trainable tensors.

Only a complete development pass opens one sealed holdout comparison. Holdout requires at least +5 points over unchanged, at least +3 points over shared, nonnegative domain deltas, and the same causal/semantic conditions.

Conditional depth follow-up

One follow-up is predeclared now: only when both final-four ECR1 and shared residual improve unchanged by at least 13 answers but miss promotion, test all 16 layers at rank 8. The receipts are fixed at 532,480 ECR trainables and 524,288 shared trainables. Stage 1 contains only these two fits and their exact development evaluations. It must score at least 256/1,289, beat shared by at least 39 answers, and preserve the frozen 40/145/5 math/logic-code floors. Only a conjunctive Stage-1 pass may spend compute on draft-removal and expert-code interventions; it never directly opens holdout. Otherwise it is unauthorized. If Stage 1 fails, ECR1 closes on OLMoE and does not scale to 35B.

Interpretation is fixed: shared passing without an ECR margin supports generic MLP-side correction and rejects expert identity; ECR passing with its matched margin and causal code interventions supports sparse expert conditioning; neither passing closes the mechanism.

Implementation

Core: train/ecr1_moe_revision.py. Trainer: train/train_ecr1_product.py. Slurm entry point: train/jobs/train_ecr1_product.sbatch. Autonomous loading and exact draft-key masking are integrated into train/hf_product_reasoning_eval.py.

Results and decision

Final-four development scores are ECR 221/1,289, shared 223, and true draft-unavailable 224, versus unchanged 191 and MTR1 204. Zeroing, averaging, or permuting the learned expert codes leaves ECR exactly unchanged at 221; expert identity is decorative in this parameterization. The exact comparison SHA-256 is 0b10336326f87b0e1b07114b98a6dde6cd02c0a06140d43502199f408a92ab72.

The predeclared all-layer follow-up was eligible and completed. Rank-8 ECR uses exactly 532,480 trainables and scores 240/1,289 = 18.6191%; rank-8 shared uses 524,288 and scores 239 = 18.5415%. ECR domain counts are math 55, logic/science 180, and code 5. Domain floors pass, but ECR misses the 256 floor and beats shared by one answer rather than 39. Exact comparison SHA-256 is 320076637af912e194b315ab6e7589240602126b8ab505184bdf2f499c602166.

Therefore Stage 2, holdout, and 35B scaling are blocked. ECR1 is closed on OLMoE without rank, layer, seed, duration, or loss rescue. The supported boundary is generic post-MoE correction capacity, not expert-conditioned sparse correction.