← Complete research archive
Architecture researchAudit122 lines

DIVERGE-SDR1: Source-Only Verified Reasoning Distillation

VCR1 reached 643/1,279 on source-disjoint holdout and 72.302% product macro, but requires two complete 4B candidate solutions at inference. SDR1 tests the shortest path to a standalone reasoner: can the identical verified repair supervision teach the pinned 9B host to solve direc…

docs/research/DIVERGE_SDR1_SOURCE_ONLY_DISTILLATION.mdOpen original Markdown ↗

DIVERGE-SDR1: Source-Only Verified Reasoning Distillation

Status: completed once and closed on 2026-08-08. Development and holdout both fail the frozen retention gate, so the product board remains sealed and SDR1 receives no variants.

Capability Question

VCR1 reached 643/1,279 on source-disjoint holdout and 72.302% product macro, but requires two complete 4B candidate solutions at inference. SDR1 tests the shortest path to a standalone reasoner: can the identical verified repair supervision teach the pinned 9B host to solve directly from the source problem, without either candidate?

This is not a VCR1 rescue variant. It is a matched causal control and a different deployable system boundary. If SDR1 retains VCR1 capability, the verified target curriculum rather than candidate-conditioned revision is the main transferable mechanism. If it collapses, the candidate trajectories are causally necessary and must remain inside the architecture.

Single Changed Factor

SDR1 matches VCR1's:

  • exact pinned Qwen3.5-9B B1 warm start;
  • 5,824/1,289/1,279 source split and row order;
  • 9,655 training presentations and every response target;
  • outcome-class weighting through deterministic duplicate presentations;
  • 256 updates, batch 1, accumulation 8, context 4,096, LR 2e-5;
  • four LoRA layers, rank 8, alpha 16, optimizer, schedule, and seeds;
  • greedy source-disjoint evaluator and assessor.

The only changed factor is question: VCR1 receives the source plus Candidate A and Candidate B; SDR1 receives only the original task prompt. Candidate text, correctness, task label, outcome class, gold answer, execution result, and assessor fields are unavailable to the runtime. An independent audit proves exact response-target/order parity for all 9,655 presentations and finds zero Candidate A/B markers in any SDR1 prompt. Training prompt bytes fall from 40,823,625 to 6,059,006.

Frozen Data

ArtifactRowsSHA-256
Train9,655a97344d4339d2687b6ed861844294efd14c575962811811ee74b97f70a7771d6
Development1,289aa0e700069a4f066cc16514d7db4f53d95ebaf6911cd4118f55cd8a09fecf5f0
Holdout1,2793e623da7b11a0bc36adbbfd0be845e3310b0125744eae09d37e8575ba2f696ce
Report-62c4eb5d9a1d1b19d9ea8b954dec4d5a2c09ea4c1791e4c176e4605d2bded97b

Immutable runtime runtime/sdr1_d11b231_r1 has manifest SHA-256 bffddf1ef4d8eb25b97065ccddd81f4190fd633ed56d280c34e8bcdb47351b6b.

Frozen Gates

Development must satisfy all:

  • at least 550/1,289 overall;
  • MATH at least 224/623;
  • BBH logic at least 294/637;
  • MBPP at least 19/29;
  • at least 75 correct cases from the both-source-wrong stratum.

Holdout must satisfy all:

  • at least 618/1,279, retaining VCR1 within two percentage points;
  • MATH at least 255/621;
  • BBH logic at least 327/625;
  • MBPP at least 24/33;
  • at least 100 correct cases from the both-source-wrong stratum.

Both splits must pass before one unchanged 568-row product evaluation may open. A qualified standalone product model must retain code >=30/40, reach at least 70% five-domain macro and 350/538 solved, improve at least three domains over QPT1, and regress no domain by more than two points versus QPT1.

Failure closes exact SDR1 without prompt, target, seed, LR, duration, rank, layer, context, decoding, or threshold variants. A result below VCR1 but above QPT1 may be recorded as a standalone specialist, but does not establish that candidate-conditioned reasoning was removed without capability loss.

Budget

Projected total charge before launch is 2--4 H100-hours for one fit, two source-disjoint evaluations, and one conditional product evaluation. Current monthly use is 409.9/2000.0 H100-hours.

Result

Training job 745625 completes all 256 matched updates in 00:12:33, charging 365,028 target tokens at 511.70 target tok/s. Checkpoint and report SHA-256 values are 8947a593793d42321e76f671cefd8c28ad34c94cc005ad91ac5ce8c756697195 and c15ff1dbe6ab7fbcfa2062f925efbbb0d3fb765518e9cb294c54e6361309fa7d. The shorter source-only context makes fit 1.72x faster than VCR1 but does not preserve its capability.

SplitQPT1VCR1SDR1SDR1 both-wrong solved
Development453/1,289575/1,289448/1,289126
Holdout471/1,279643/1,279490/1,279137

On holdout, SDR1 scores MATH 142/621, BBH logic 321/625, and MBPP 27/33. Code and the both-wrong floor pass, but overall, MATH, and science retention fail. Development independently agrees: MATH 109/623, logic 319/637, MBPP 20/29, and overall 448/1,289. Development/holdout report SHA-256 values are 64282b672a7fe3c644f823bbfd8933594c600f2aaa200fc3a665c0e75ca756f3 and 516f244896dee9c1ad1e08049c55bf3f43a507d085c0d606413ba7268a1cc481.

This is a clean causal result. The verified target curriculum alone does not explain VCR1: source-only holdout is 153 answers below candidate-conditioned VCR1. Candidate trajectories are especially important for MATH, where VCR1 solves 273 and SDR1 only 142. SDR1 still demonstrates some direct transfer over QPT1 on holdout (+19) and strong code, but it is not a standalone replacement. The prebuilt product board stays unopened. The next architecture must generate a draft internally and revise it in a later model-owned pass; one-pass source-only distillation is closed.

Training plus both evaluations consumed 0.676 H100-hours, below budget.