← Complete research archive
Architecture researchPreregistered184 lines

Shohin Transferable Temporal Revision Contract

The target is a transferable temporal reasoning architecture that improves an existing pretrained model using only model-owned computation. One state of a backbone writes a complete draft; a separately trained state of the same backbone reads the source and that draft, then commi…

docs/research/SHOHIN_TRANSFERABLE_TEMPORAL_REVISION_CONTRACT.mdOpen original Markdown ↗

Shohin Transferable Temporal Revision Contract

Status: current evidence contract, updated 2026-08-09. TTR1 established strong aggregate dense cross-family transfer but failed executable-code retention. MTR1 shared-attention and RCR1 static-router transfer both failed on small OLMoE development. The current critical path is MoE route/error attribution followed by one genuinely MoE-native successor; scratch pretraining and a larger-MoE campaign remain unauthorized.

Mission

The target is a transferable temporal reasoning architecture that improves an existing pretrained model using only model-owned computation. One state of a backbone writes a complete draft; a separately trained state of the same backbone reads the source and that draft, then commits one coherent final trajectory. The final system contains the backbone, both role states, and a deterministic two-phase controller in one deployable package.

At claim time there is no external proposal model, verifier, answer router, tool, host solver, correctness bit, or benchmark-specific route. Training may use independently verified targets. Evaluation never exposes those targets.

Existing Boundary

HostDevelopment treatment/controlHoldout treatment/controlBoundary
Qwen3.5-0.8B323/236 (+6.749 points)328/242 (+6.724 points)aggregate transfer; holdout code 8/9 fails conjunctive retention
Qwen3.5-4B529/371 (+12.26 points)554/380 (+13.61 points)every source-disjoint domain positive; protected product mixed fail
Qwen3.5-9B589/464 (+9.70 points)625/495 (+10.16 points)every attribution domain positive; original promotion floor missed
SmolLM3-3B469/358 (+8.61 points)sealedaggregate cross-family transfer; code 4/9 fails conjunctive retention

These results establish a real learned revision effect across Qwen and SmolLM families at aggregate level. They do not establish reliable all-domain family transfer, protected-product promotion, or a MoE mechanism.

TTR1 Host And Inputs

The shortest cross-family test uses the already-pinned HuggingFaceTB/SmolLM3-3B causal model:

  • revision: a07cc9a04f16550a088caea529712d1d335b0ac1;
  • local config SHA-256: c72b1031274ff4626e434d0019e88e95a767460135db9ee492eb80652b786af1;
  • draft warm start: existing SmolLM3 B1 final-four-layer rank-8 checkpoint;
  • checkpoint SHA-256: c4af49e88d31ef751d7ab8697f12481b645801963e1ecb71174a7da645df6a35;
  • exact source bank: 4,096 MATH, 4,096 logic/science, and 200 execution-verified MBPP identities;
  • bank SHA-256 values: e0ede832...dbe5, 5a96859f...017, and 0b6d068b...398;
  • split and evaluator: the existing source-identity-disjoint 9,655/1,289/1,279 revision contract and unchanged exact-answer/execution evaluators.

No Qwen draft or adapter may enter TTR1. Every draft and final response is produced by a SmolLM3 role state.

Changed Factor

Let G(theta, x, b) be one bounded generation from pretrained backbone state theta, input x, and decoding budget b.

  1. Draft: d = G(theta + delta_draft, source, 768).
  2. Revision: y = G(theta + delta_revision, source || d, 768).
  3. delta_revision is trained on complete verified final trajectories while theta and delta_draft remain frozen.
  4. Commitment is whole-trajectory: the system emits exactly one revision and never averages incompatible answer fields.

The frozen treatment geometry is the prior IDR schedule: final four decoder layers, rank 8, alpha 16, 256 AdamW updates, batch 1, accumulation 8, context 4,096, learning rate 2e-5, and the existing data/order seeds. The legacy B1 metadata field unfreeze_layers=null must be normalized to numeric zero by the already-tested bitwise-preserving migration before use.

Matched Controls

All arms use the same SmolLM3 revision, source identities, prompts where applicable, evaluator, and recorded generated-token and FLOP accounting.

  1. Unchanged second pass: B1 reads the exact treatment draft through the exact treatment revision prompt and receives the same 768-token final budget. This is the primary causal control.
  2. Standard self-refinement: unchanged B1 receives the same draft under a generic review-and-correct instruction and the same final budget.
  3. Long single generation: B1 receives source only and at most 1,536 new tokens, matching the maximum two-pass generated-token budget.
  4. Best-of-two/majority: two source-only B1 attempts share the same total generated-token ceiling; exact-answer agreement commits, and deterministic first-attempt tie-breaking is reported.
  5. Independent commitment: an equal-geometry, equal-update LoRA is trained on the same final targets without informative draft content. Input shapes and attention FLOPs are matched by masking the draft span rather than deleting it. This isolates learned finalization from trajectory-conditioned revision.

Prompt tokens, generated tokens, wall time, peak memory, trainable parameters, and estimated attention/MLP FLOPs are reported per arm. Any unmatched control is labeled diagnostic and cannot satisfy the matched superiority claim.

Release Gates

First run one mechanics-only smoke on 24 balanced examples. It must prove generic SmolLM3 loading, exact checkpoint restoration, draft serialization, masked-span control behavior, and identical treatment/control prompt tokens. It is not a capability result.

Development opens only once. TTR1 advances to holdout only if:

  1. trained revision exceeds unchanged second pass by at least five absolute points overall;
  2. MATH, logic/science, and executable-code correct-count deltas are each nonnegative;
  3. trained revision exceeds the strongest fully matched standard control by at least three absolute points overall;
  4. all 1,289 identities and all compute receipts are complete.

Holdout applies the same conjunctive conditions independently to all 1,279 identities. Only a development-plus-holdout pass opens one protected product evaluation. Product promotion requires at least +0.05 five-domain macro accuracy over the strongest matched control and nonnegative correct-count deltas in every main domain. A miss closes exact TTR1 without seed, rank, layer, duration, prompt, decoding, or threshold rescue variants.

Frozen Result

TTR1 development completed all matched arms on all 1,289 identities:

ArmCorrectAccuracy
trained draft revision46936.3848%
unchanged second pass35827.7735%
self-refinement39830.8766%
long single generation42032.5834%
best-of-two33926.2995%
independent commitment37128.7820%

Treatment exceeds unchanged by +111 answers / +8.6113 points and the strongest matched control, long single generation, by +49 answers / +3.8014 points. Domain correct-count deltas against unchanged are MATH +62, logic/science +54, and executable MBPP -5 (9 -> 4). The first, third, and coverage conditions pass; the all-domain condition fails. Holdout was not opened.

Comparison SHA-256 is a20cdd7567aa2502a62ad5591748bc9241371b3f7fe4620c29a1de71e90c9cd7. Treatment and independent-report SHA-256 values are 98d4a4b4f7b37f9369fe7c81ec3a43f60531c804a4d8b055c27a3b51772c5475 and a6974226297da951eca5abf98bd513cd5b04e451b2ff5158dfa8b522847f4dc7. Complete execution charged 27.793 H100-hours, including canceled unsharded work and node failures. Exact TTR1 is closed without a nearby retry.

Advancement

SmolLM3 established aggregate cross-family transfer but exposed a capability preservation failure. OLMo2-7B then supplied the larger dense non-Qwen test: direct revision reached only 259/1,289, while selective commitment and error-syndrome revision both failed their frozen gates. More local OLMo variants are not authorized.

MTR1 on small open OLMoE-1B-7B-0125-Instruct is complete and closed. Final-four-layer rank-8 shared-attention revision reaches 204/1,289, only +13 answers / +1.0085 points over unchanged, while mean all-layer route-count L1 drift is only 0.002018. RCR1 then directly trains a bounded low-rank residual on the final four router logits and reaches 194/1,289, only three answers over matched rank-1 attention and unchanged and below MTR1's 204. Both holdouts remain sealed.

The next operation is read-only attribution of corrected, broken, persistent-wrong, and preserved-correct cases against per-layer route changes. Only then may one draft-conditioned multi-token MoE controller be frozen. The leading candidate jointly controls bounded routing deltas and small revision-specific expert-side low-rank capacity, with router-only, expert-only, equal-budget attention, and draft-shuffled/masked controls. docs/research/SHOHIN_MOE_FRONTIER_CONSULTATION_BRIEF_20260809.md contains the self-contained problem statement.

Larger-MoE capability work, including the mechanics-qualified Qwen3.6-35B-A3B path, remains prohibited until a small-MoE development and sealed-holdout pass.

Scratch Shohin training is optional later efficiency evidence. It is not the critical path and is not authorized by TTR1.