DIVERGE-VCR1: Verified Temporal Correction Reactor

Status: frozen bounded gate, no result yet.

Capability hypothesis

The failed LTM1/VMT1 objectives assigned semantic roles to exchangeable parallel trajectories. Their exact objectives admit a collapsed barycenter. VCR1 removes that symmetry by assigning roles through time:

1. the protected generator emits one autonomous first-pass draft;
2. a model-owned reactor reads the original problem and that draft through distinct segment channels;
3. a learned validity state controls a persistent correction prefix; and
4. the same protected generator emits one final response conditioned on the prefix.

The hypothesis is that explicit temporal roles let a compact module learn wrong -> right repair without erasing right -> right behavior. VCR1 is not credited for merely fitting teacher-supplied drafts. It must improve autonomous drafts from the exact protected generator.

Frozen source

• Backbone: HuggingFaceTB/SmolLM3-3B revision a07cc9a04f16550a088caea529712d1d335b0ac1.
• Protected generator checkpoint: baseline_late2_v12u1000_rollout_replay_u400_r1/checkpoint_0000400.pt.
• Checkpoint SHA-256: 34c82454e0c53609bc1ac6a9f127437080e431f147e28ae63b4080c413d9a82e.
• Verified pair source: rest_rollout_preferences_freshv3_k4_p2_r1.jsonl.
• Pair-source SHA-256: 38d582fa6f5626f34d7f390e243bc4f17c4114ce06d83fe0feca755bd7b88ba6.

The pair source contains 7,452 same-generator math/science pairs. Every pair has one independently verifier-correct response and one wrong response. A deterministic identity split prevents any problem from crossing train and teacher-draft development sets. The builder keeps each complete autonomous draft and derives one concise terminal target from the verifier-correct response using the same task-specific answer extractor used for scoring. It admits a pair only when both the wrong-draft correction example and the correct-draft no-op example fit in the 4,096-token contract without truncating the draft or terminal target. The first full-response admission attempt is a mechanical pre-training negative: it left only 211 train-math identities and produced no board or model result.

The correction instruction matches this target: the model verifies or repairs the draft internally and emits exactly one terminal boxed-answer line. The complete draft remains available to the correction reactor; no source trace is truncated or replaced by a gold trace.

Architecture

The protected product generator is loaded exactly and then frozen. VCR1 adds only a correction workspace:

• eight 384-wide persistent slots;
• four applications of one tied recurrent correction block;
• separate question and draft cross-attention channels;
• a signed discrepancy fusion over question context, draft context, and their difference;
• slot self-attention and a gated recurrent update;
• a terminal draft-validity head; and
• an eight-token soft prefix projected into the frozen 2,048-wide language state.

Let q and d denote the original-problem and draft token sets. At recurrent step t, the shared block computes

Q_t = Attn_q(S_t, q)
D_t = Attn_d(S_t, d)
C_t = MLP([Q_t, D_t, Q_t - D_t])
S_{t+1} = GatedUpdate(S_t, SelfAttn(S_t) + C_t)

The terminal validity probability p_valid scales the emitted correction prefix by 0.05 + 0.95 * (1 - p_valid). Correct drafts therefore retain a small learnable no-op channel while wrong drafts can activate the full patch. No verifier, gold answer, alternate response, or candidate bank exists at inference.

Matched controls

Three systems consume the same autonomous first drafts:

1. one-pass source: the protected generator's original draft;
2. plain two-pass: the protected generator receives the same correction prompt but no learned prefix; and
3. role-blind control: the exact VCR1 parameters, recurrent steps, losses, and FLOPs, but both cross-attention channels receive the union of problem and draft tokens.

The treatment differs from the role-blind control only in preserving the question/draft fault line. This is a stronger control than a smaller MLP or a zero-parameter prompt baseline.

Training contract

• seed 2026080603; data seed 2026080603;
• one wrong-draft and one correct-draft presentation per selected pair;
• identical verifier-accepted terminal answer as the target in both presentations;
• 4,096 total positions including eight correction slots;
• 200 updates, eight pairs per update through gradient accumulation;
• AdamW, fused, LR 3e-4, cosine decay, gradient clip 1.0;
• language loss plus 0.20 * BCE(validity) and 0.10 * max(0, 0.25 - correction_wrong + correction_correct);
• protected generator parameters remain bit-identical.

Mechanical smoke uses two updates and cannot qualify capability. Treatment and role-blind control start from byte-identical correction-module weights.

Autonomous gate

The fixed development gate uses one deterministic first-pass draft per prompt from the protected generator on 100 MATH-500 and 100 held-out OpenScience short-answer problems. The same draft bank is consumed by plain two-pass, role-blind, and VCR1 correction. Report for each arm:

• correction generation is deterministic greedy decoding, capped at 256 new tokens inside one common 4,096-position admission contract;

• total and per-domain exact accuracy;

• wrong -> right, right -> right, right -> wrong, and wrong -> wrong;

• net correction (wrong -> right - right -> wrong);

• generated-token use and exhaustion; and

• validity calibration on the independently scored first drafts.

VCR1 advances only if all are true:

1. finite execution and byte-identical protected source tensors;
2. at least five net autonomous corrections over one-pass source;
3. at least three more correct answers than role-blind over 200 prompts;
4. positive net correction in both math and science;
5. no domain loses more than two answers versus source;
6. right -> wrong is at most half of wrong -> right; and
7. swapping question/draft masks or resetting the prefix removes at least two treatment solves.

If plain or role-blind two-pass equals or beats VCR1, temporal correction may still be a useful product technique, but this architecture has not earned promotion. Failure closes VCR1 without a nearby seed, width, depth, loss, duration, or threshold repair. The exact reducer is pipeline/score_diverge_vcr1_gate.py; it additionally fails closed unless all five correction arms consume the same complete ordered 100-row draft bank in each domain with zero admission skips.