DIVERGE-IDR1: Internal Draft Revision

Status: exact training, source-disjoint evaluations, and the matched no-revision attribution control are complete. Trained revision is causally positive, the conjunctive promotion gate failed narrowly, and product remains sealed.

Hypothesis

VCR1 proves that candidate-conditioned revision can exceed its two source attempts. SDR1 proves that the same verified targets without candidate trajectories lose 153 holdout answers. IDR1 preserves temporal computation while removing external proposal models: one pinned Qwen3.5-9B B1 owner first generates a complete draft from the source, and a later 9B revision owner sees source + internal draft and emits the final solution.

The model family owns both passes. No 4B B1, 4B QPT1, task router, gold answer, correctness bit, evaluator feedback, or external tool is available at inference. The proposal and revision owners share the exact pinned 9B backbone, but use distinct small LoRA states. This is a deployable two-pass reasoning architecture rather than source-only SFT.

Frozen Draft Collection

The source bank is the exact 8,392-row CVG1 bank already partitioned by the NUL-delimited source identity split. Draft generation uses:

• pinned Qwen/Qwen3.5-9B@c202236...337b9a;
• exact 9B B1 checkpoint SHA-256 854a7cc44fbc2b54418f4e5bd09b7efeed0da44fc9ce217b0bb6b1997b722971;
• one greedy sample, no thinking mode, maximum 768 new tokens;
• reasoning prompt style, prompt batch 4, seed 2026080818;
• eight 512-row math shards, eight 512-row science/logic shards, and one 200-row execution-verified code shard;
• 17 independent one-H100 jobs with isolated atomic outputs.

Every shard must bind the immutable runtime, model, checkpoint, and source bank. Merge must prove 8,392 unique identities, exact bank coverage, one draft per identity, and unchanged source scoring. No draft score is a gate by itself; drafts become model-visible training inputs only after the complete receipt passes.

The prelaunch draft-generation estimate was 3--6 H100-hours. At 05:19 EDT, early measured progress from 16 concurrent shards corrected that estimate to 16--21 H100-hours. The jobs remain scientifically valid and already hold the allocations, so they continue unchanged. The later revision fit and source-disjoint evaluation are separately expected to consume 2--4 H100-hours. Both the original estimate and correction are preserved.

Draft And Data Result

All 17 jobs 745628--745644 completed for an exact charge of 17.837 H100-hours. The complete source-owned first pass scored:

Merged drafts SHA-256 is 509b114e42773dbe8d14536ff5ce7e5a2f92b45b3acd57a37a035fce43505d73; the complete receipt is 2154a49811dece70f25ccb309c8ef1fdc08d4513880477b500a4cbfba156a852. The frozen revision data contain 9,655/1,289/1,279 train/development/holdout rows. Their SHA-256 values are 6df32045...ac6c, 0c52dd35...4224, and df347d59...36f2; report SHA-256 is 8e4b2817...c80e.

CPU build 745645 failed before output on a CLI receipt keyword mismatch. Replay 745648 completed and preserved the valid merge, then failed before revision data on an omitted transitive runtime import. Data-only recovery 745650 reused the hash-bound merge and completed. These were packaging failures only: no scientific input or setting changed.

Revision Training Result

Job 745652 completed all 256 frozen updates in 19m50s. It charged 365,028 target tokens at 320.55 target tokens/s. Checkpoint SHA-256 is df3c264d426941fef8ba9c10a90fe9fab304ec2864738209a4d79f9f81e0c473.

Source-Disjoint Gate Result

Development missed the frozen MATH and code floors by one and two answers. Holdout passed overall, logic, code, and both-wrong floors but missed the MATH floor by seven. The exact gate is therefore closed and the product board remains sealed. Development and holdout report SHA-256 values are 0880e59c03460f3ab0f5c8da87136fd9f60961f3fa16a762225188ac998b7a40 and 74834cad3ee4c32e1e263d968bbb2f5b1f4dfeb6eca91b124e1a4f5a03148b53.

The result is nevertheless a large, independently reproduced capability gain. On holdout, revision improves the internal first pass from 265 to 625 correct, exceeds one-pass source-only SDR1 (490) and the 9B QPT1 expert (471), and finishes 18 answers below externally proposed VCR1 (643). Of 363 cases repaired after an incorrect draft, 356 drafts had exhausted the 768-token budget and 327 contained no explicit candidate answer; only three initially correct drafts were lost. This identifies a major learned continuation/finalization effect, especially for long truncated trajectories.

Matched Attribution Control Result

The matched control ran the same frozen internal drafts, source-plus-draft prompt, evaluator, generation budget, and split using the original exact 9B B1 adapter rather than the trained IDR1 revision state. Slow long-form generation made monolithic jobs 745657/745658 unable to fit their scheduler windows. They were replaced by four contiguous batch-aligned shards per split, 745659--745667, with no model, prompt, ordering, batching, decoding, seed, or scoring change. One failed CUDA-invisible allocation is replaced exactly. Complete identity coverage was proven before merge and scoring. The result is:

Holdout gains are MATH +83 (165->248), logic +46 (305->351), and code +1 (25->26). Development gains are +73/+50/+2. Trained IDR1 fixes 176 control errors while losing 46 control-correct answers on holdout. On the 914 token-exhausted drafts, control/trained score 286/402; on the 365 non-exhausted drafts they score 209/223. The result therefore proves a learned revision and termination policy, not just the generic benefit of providing a draft to an additional inference pass.

Development and holdout control reports hash to 08f5e12f006426df2d9a9a71dc3a69a9a0825a508121309e1fa386054be60d1a and 37bb0c5300d5b8dec45a6e8b3a07b20b708a17cdff2ce5b22481c002702ccb9a. Valid sharded evaluation cost 10.220 H100-hours. Superseded monolithic and failed/canceled evc33 allocations add 1.104, making the exact attribution control cost 11.324 H100-hours. Including draft collection, training, and the primary source gates, complete IDR1 cost is 30.124 H100-hours.

The attribution result cannot reopen the exact IDR1 gate or authorize the sealed product board. It does establish IDR1 as the strongest measured model-owned same-family architecture in this campaign.

Frozen Revision Gate

After complete draft custody, IDR1 will use the exact SDR1/VCR1 verified targets and split. The only new input is the frozen model-owned 9B draft. Training geometry remains the 256-update pinned 9B B1 LoRA schedule. Holdout must retain VCR1 within two percentage points (>=618/1,279), meet the same fixed MATH/logic/code and both-wrong floors, and reproduce on development before the product board opens.

Failure closes exact IDR1 without draft sampling, prompt, seed, LR, duration, rank, layer, context, decoding, or threshold variants. The measured failure is now final. The matched no-revision control is attribution, not a rescue.