CTSR1: Causal Temporal State Routing

Status: closed negative, 2026-08-09. Development only. Holdout and large-MoE transfer remain sealed.

Hypothesis

MTR1, RCR1, ECR1, SER1, and RME1 show that static native expert identity is not predictive of revision success. The missing variable may be temporal: revision requires a diagnosis that survives the draft prefill and changes as the model emits a replacement trajectory. CTSR1 tests whether a small causal state, updated across actual tokens and persisted through cached generation, can make later sparse computation useful.

At each of all 16 sparse layers, a single GRU shared across layers reads the pre-MoE hidden sequence. Each layer owns only a 64-dimensional initial code and persistent inference state. Two tied 64-to-32-to-64 heads read the causal state. Every layer also owns the already useful rank-18 shared post-MoE residual.

s[l,t] = shared_GRU(h[l,t], s[l,t-1])
u1, u2 = tied_heads(s[l,t])

treatment:
  router_logits = frozen_logits + tanh(u1)
  residual_gate = P[l] tanh(u2)

matched temporal control:
  router_logits = frozen_logits
  residual_gate = P[l] tanh((u1 + u2) / 2)

delta = B[l] (A[l] h * (1 + residual_gate))
output = frozen_selected_experts(h, router_logits) + delta

P[l] is a deterministic fixed projection, never trained. During training, the GRU is reset for each complete source+draft+target sequence and is causal over tokens. During greedy generation it consumes the left-padded prompt once, stores the final state separately at every layer, then advances that state on every generated token. No frozen-backbone prepass, pooled draft vector, source KV side channel, or host state exists.

Exact accounting

Geometry is fixed: hidden 2,048, state 64, head width 32, residual rank 18, 16 layers, 64 native experts, native top-8.

• shared GRU: 405,888 parameters;
• 16 layer initial codes: 1,024;
• two tied state heads: 8,192;
• 16 rank-18 residuals: 1,179,648;
• total: exactly 1,594,752 trainables in either arm;
• adapter MAC/token/layer: exactly 488,576 in either arm.

All native attention, routers, experts, embeddings, norms, and LM head remain frozen. Both arms use identical heads and execute identical GRU/head/residual work. The changed factor is whether u1 changes expert routing or supplies a second generic residual-state signal.

Development gate

One-update mechanics must prove exact base parity before optimization, exact parameter/MAC receipts, finite gradients, left-padding equivalence, streaming prefill/token equivalence to one-shot causal recurrence, protected hashes, and complete 4,096-token sequence custody.

Only then train treatment and temporal control once for the existing 256 updates, LR 2e-5, seed/data seed, 9,651 complete rows, and 338,620 charged target tokens. Reuse static shared rank-18 248/1,289 read-only as an external floor. Promotion requires all of:

1. treatment at least 280/1,289;
2. treatment at least 26 answers above the newly trained temporal control;
3. treatment at least 13 answers above static shared rank-18;
4. domain counts at least math 55, logic/science 180, code 5;
5. finite state/route receipts, all 64 native experts used, load entropy at least 0.80, and top-1 route-change rate at least 0.01.

Only a conjunctive pass opens evaluation-only state reset, route-head zero, state transplant, and true draft-unavailable controls. Normal must beat each by at least 13 answers before one holdout. Failure closes CTSR1 and the small-OLMoE MoE-native route without width/depth/rank/duration/seed variants.

Result

Mechanics job 747342 passed exact base-parity, parameter, compute, memory, sequence, and protected-weight receipts. Treatment 747343 and the valid temporal-control retry 747345 each completed the exact 256 updates and 338,620 target tokens. The initial control allocation 747344 failed before training because evc31 lacked node-local disk and is retained as an infrastructure failure.

Autonomous comparison 747420 scores treatment 249/1,289 = 19.3173%, temporal shared 245 = 19.0070%, and immutable static shared rank-18 248 = 19.2397%. Treatment domains are math 59, logic/science 185, and code 5. Domain floors and load entropy pass, but treatment misses 280, beats the temporal control by only four, and beats static shared by only one. Mean top-1 route change is 0.0002485, far below the frozen 0.01 causal gate. Comparison SHA-256 is 34af28ab62ebb5214dbadca144f1ef81df8403a8807b489ca464affbfac1c203.

Persistent token-causal state therefore does not establish control over expert computation in this host and budget. CTSR1, its controls, holdout, and large-MoE scale-up are closed. The small-OLMoE MoE-native lane is retired without width, depth, scale, rank, duration, or seed rescue.