Dynamical-SDL1-35B-A3B

Dynamical-SDL1-35B-A3B is an open-weights policy for scientific decisions in experimental campaigns. It has 35 billion total parameters and 3 billion active parameters. We initialized it from Qwen3.6-35B-A3B and post-trained it in the SDL-1 environment.

SDL-1 replays recorded physical campaigns as partially observable decision environments. The policy must decide what evidence to acquire, which claims that evidence supports, how its beliefs should change, and what experiment to run next under a fixed budget.

Read the technical report, the SDL-1 benchmark report, and the companion research code.

Results

SDL-1 reports negative held-out log loss. Zero is the ceiling, so higher values are better.

System	Adaptive AUC	What it measures
Trained checkpoint	−1.5138	The trained weights with the frozen inference setup
Final system	−1.4664	The same weights with the released system prompt, behavior rules, completion rule, relevant-experience grounding, and three-draw forecast pooling

The same-runtime comparison isolates the added inference setup on byte-identical trained weights. It does not assign the final result to one instruction or to pooling alone.

The final system completed all 12 campaign branches. No eligible scored belief row triggered the unreviewed-refinement overclaim indicator, all 234 eligible belief-update actions passed first-pass validity, and endpoint gain reached +0.023. Experiment-selection lift was +0.009 and remained inside evaluation noise.

Model and base revision

• Architecture: Qwen3.6 mixture of experts
• Total parameters: 35B
• Active parameters: 3B
• Weight format: BF16
• Base model: Qwen/Qwen3.6-35B-A3B
• Base revision: 7da1103448ba36029c34ce1a9a741dfe93ee0c50
• Adapter: LoRA rank 16, alpha 32, all linear layers
• Merged checkpoint fingerprint: f27f3ed6cc6a3ac8d4e60fbc846cf2290a90d951bb01ae6640b1086f03077b84

The repository includes the merged BF16 model and the final adapter. The adapter configuration is pinned to the exact base revision. The merge fetched the files when the upstream main branch was at 995ad96eacd98c81ed38be0c5b274b04031597b0. The model, configuration, index, and tokenizer files are byte-identical to the pinned revision. The later revision changed the upstream README and added its license file.

Training

The training stages formed one sequential lineage. Each stage started from the weights produced by the stage before it.

1. Supervised fine-tuning used rejection-sampled actions to improve first-attempt validity.
2. Process RL used REINFORCE Leave-One-Out and Clipped Importance Sampling Policy Optimization to assign local credit across evidence updates.
3. Campaign RL widened credit assignment across one- and two-experiment campaigns.
4. Outcome-grounded RL trained evidence assimilation and claim scope against evaluator-private recorded outcomes.

Recorded outcomes affected rewards and weight updates. They did not enter the policy-visible state that produced each decision.

Inference package

The final system used the files in inference/:

• system-prompt.txt is the exact frozen system prompt.
• behavior-rules.md describes the seven behavior groups and the separate campaign-completion rule.
• sampling.json records sampling and three-draw pooling.
• serving.json records the serving contract used for evaluation.

The prompt grouped recurring conduct into seven behavior rules:

1. Keep unsupported certainty within working probability bounds.
2. Change forecasts only when item-specific evidence supports the change.
3. Raise only the confidence dimensions addressed by the evidence.
4. Reset evidence sufficiency for each new experiment.
5. Compare alternatives and favor relevant precedent while new results still change the forecast.
6. Spend evidence early and stop purchasing checks that no longer change the decision.
7. Request expert review only after cheaper evidence is available and a specific conflict remains.

A separate completion rule required the policy to use its full experiment budget and to plan each selection against the evidence and review budgets that remained.

Relevant-experience grounding started each forecast from recorded reactions that shared either precursor with the candidate. The policy used overall recorded frequencies when there were too few analogs. It could then adjust the prior when the temperature regime, precursor melting points, or acquired evidence supported a change. Evidence could transfer only through an explicit shared precursor, condition, or mechanism.

The environment supplies a response schema for each policy state. This dynamic schema depends on the allowed actions, identifiers, and forecast panel for that turn. The companion code contains the renderer and schema definitions.

Sampling used temperature 1.0, top-p 0.95, top-k 20, a maximum of 32,768 output tokens, and five returned log probabilities. No provider fallback was allowed.

The three-draw method applies only to scored closure forecasts. It runs three independent full-state forecast draws, takes the arithmetic mean of each probability, renormalizes each reaction distribution, and uses the modal recommendation. It does not pool campaign actions or change the committed campaign trajectory.

The evaluation served the model on one H100 with vLLM 0.24.0. It loaded the BF16 checkpoint with on-the-fly FP8 quantization, a 262,144-token context, tensor parallel size 1, one active sequence, and at most 16,384 batched tokens. The server reserved 92% of GPU memory. Prefix caching and chunked prefill were enabled. It used language-model-only mode and JSON-schema responses. The Qwen3 reasoning parser was active, and thinking was enabled and preserved in campaign history. Each turn allowed up to four action samples with a 3,600-second request timeout.

Evaluation boundaries

The comparison with the base model is a combined system comparison. The base row used a pinned OpenRouter route, while the final system used vLLM. It does not isolate training.

The comparison between −1.5138 and −1.4664 uses byte-identical trained weights and the same vLLM runtime. It measures what the released inference package added to the trained policy.

Fifty-eight of the 192 canonical evaluation-tile rows appear in the gradient inventory through alternative-fold reuse. This potential weight-level exposure favors our row. We make no untouched-benchmark claim.

The published SDL-1 systems used their own runtimes. Their placement beside this model is descriptive rather than controlled.

Limitations

• The policy did not reliably carry newly acquired evidence into later forecasts.
• Experiment-selection lift remained inside evaluation noise.
• SDL-1 begins after the objective and candidate experiments have been defined. It does not measure question framing, research direction, or optimal stopping.
• The system prompt and behavior rules were tested together. The evaluation cannot assign the result to an individual rule.

Intended use

This model is a research artifact for experimental-campaign replay, evaluation, and method development. Use it with an environment that keeps evaluator-private outcomes separate from policy-visible evidence and validates every structured action.

It is not a substitute for expert review, laboratory safety controls, or independent validation of physical claims.

License and attribution

This release is licensed under the Apache License 2.0. It is derived from Qwen3.6-35B-A3B at revision 7da1103448ba36029c34ce1a9a741dfe93ee0c50. See NOTICE for attribution.

Citation

@article{barnes2026scientificagents,
  author  = {Barnes, Jarrod},
  title   = {Training Scientific Agents to Learn from Physical Experiments},
  journal = {Dynamical Systems},
  year    = {2026},
  url     = {https://dynamicalsystems.ai/blog/training-scientific-agency}
}