Dynamical-Systems/sdl2-development-investigations

SDL-2 Development Investigations

Complete records of 28 simulated investigations from SDL-2, Dynamical's benchmark for AI agents that investigate an unfamiliar materials process. In each, DeepSeek V4.1 Flash searches for a sintering route for a new NdFeB magnet powder lot in a digital twin of the process, or concludes that none exists. Each case is the same lot and noise run twice, once from the starting checkpoint and once after twelve reinforcement-learning updates on whole investigations.

The 14 cases are every matched run on the degassing-decides development lots where exactly one of the two agents reached a correct decision. In 12 only the trained agent did, and in 2 only the starting checkpoint did. They cover 12 lots, so two lots appear twice. Case 1 is the example in the release, and every step can be read in the investigation replays.

Contents

Only what the agent saw and wrote. Lot identities, seeds and evaluator fields are removed.

• investigations has one row per investigation, with id, case, lot, run, pair (trained_only or starting_only), agent (starting or trained), model, decision (nominate, restrict or none), correct and steps.
• steps has one row per agent turn, with investigation_id, case, agent, step, the model's reasoning and message, and actions. actions is a JSON list of tool calls. Each is a design (new batches, specimens, anneals and measurements, with the agent's stated purpose) or a decision (a stage revision, a nomination or a restriction, with its rationale and forecast), together with the simulated results the agent received.

correct is the simulated world's verdict on the final decision. The task the agent received is described on the task and environment page.

Leaderboard

The leaderboard config holds the SDL-2 frontier panel as aggregates. Each row is one model, stratum and score (Both, Outcome or Justified) over the 29 sealed test worlds, with 95% world-bootstrap and exact binomial intervals. It contains no world ids, routes or transcripts.

The config also holds one row labelled Ours, our trained agent. It is the best of three DeepSeek V4.1 Flash checkpoints from our training run in the digital twin, chosen by Both count on the same 29 worlds, with ties going to Outcome and then to the pre-registered primary checkpoint. Choosing it on this panel biases the row upward by about one to two worlds, and all three checkpoints are reported in the post's training section. Ours ran in the full agent harness and sampled trained weights on Tinker, while the frontier DeepSeek V4.1 Flash row ran through the provider API.

The file policy_arms.json holds the pre-registered results for all three checkpoints on the same worlds. It contains Both, Outcome and Justified with intervals, paired counts against the starting checkpoint, the pre-registered behavior counts, and the check of the prediction made before the runs. Like the rest of this dataset, it holds aggregates only.

Scope

These are simulated results on development lots, which informed development, recorded on the lean interface used for the training evaluation. The 29 SDL-2 test worlds stay sealed and are not included.

Citation

@article{barnes2026sdl2,
  author  = {Barnes, Jarrod},
  title   = {Teaching AI to Explore the Physical World},
  journal = {Dynamical Systems},
  year    = {2026},
  note    = {SDL-2 v1.0},
  url     = {https://dynamicalsystems.ai/blog/sdl-2}
}