Dynamical-Systems/sdl2

SDL-2

SDL-2 measures whether AI agents can decide, from experiments they choose, whether a process will reliably meet a stated performance requirement. Each investigation takes place in a digital twin of magnet processing. The investigator chooses experiments and either nominates a reliable processing route or concludes that none works.

This dataset contains all 32 evaluation worlds with their hidden parameters and the true outcome of every route, and all 256 investigations with their scores and released traces. Each of six language models and two classical methods ran once on every world.

The research post explains the findings. The code repository provides the harness, grader, baselines, digital twin and reproduction commands.

Start here

• Use investigations.parquet to compare decisions, scores and experiment spending, and to locate each trace.
• Use worlds/worlds.jsonl to inspect each world and worlds/routes.parquet for its route outcomes.
• Follow the reproduction instructions to regenerate the published results from the released records.

Contents

File inventory

Path	Contents
investigations.parquet	One row per investigation, with investigator, world, decision, Outcome, Supported, Both, credits, model cost and trace path
worlds/worlds.jsonl	One row per world, with its class, whether a working route exists and hidden parameters
worlds/routes.parquet	One row per world and route, with the true probability that magnets from a fresh batch meet the requirement
worlds/records/, worlds/truth/	The world records read by the grader and their complete route outcomes
worlds/board-package/	The physics package needed to run and grade the evaluation worlds
worlds/board_panel.json, worlds/ranking/	World setup records and the supplied model’s route ranking
traces/<episode>.tar.zst	Each investigation’s released record
board/	Frozen grades, results, sensitivity analyses, paired comparisons and investigation list
post/	The research post’s data, regenerated by sdl2 reproduce
annotations/blind_variation/	Annotations, codebook and identity key for the 66 near-requirement language-model claims
environment/	The experimental record, supplied physics models and initial investigation state used by the harness, with the exclusions below
images/	The aarch64 container images used by the harness
replaced/	The original replaced investigation and its grade
MANIFEST.sha256, WITHHELD.json	File hashes and the record of omissions and redactions

A trace contains the agent’s transcript, provider requests and responses, readable reasoning, laboratory requests and deliveries, retained workspace files and grade inputs. The shared initial laboratory state is stored once in environment/base-investigation/.

Scoring

The investigator has 570 experiment credits, at most 12 new magnets and 8,064 candidate routes. A route works when three magnets made from a fresh batch all meet 10 kOe coercivity and 11 MGOe energy product with probability at least 0.95.

• Outcome means the decision is true in the hidden world and within the resource limits.
• Supported means the investigator’s own measurements meet the fixed evidence rule.
• Both requires a correct, supported decision.

The grading rules define the evidence needed to support either decision.

Collection and provenance

The evaluation used 32 simulated worlds, balanced between 16 with a working route and 16 without one. Worlds were drawn from the supplied physics model’s uncertainty and selected on hidden properties, with quotas fixed before any agent ran. Language models used the same Pi 1.4.0 harness, at high reasoning effort, offline and text-only. Each model was pinned to one provider through OpenRouter.

The wall_seconds column in investigations.parquet records elapsed time from the original run summary. In board/FEATURES.jsonl, it uses summed event spans where available, excluding gaps between attempts, and otherwise retains the summary value.

The study ran on Linux aarch64. The release’s grading and reproduction commands have also been verified on Linux x86_64. Running a new investigation requires the released aarch64 container images. Platform requirements and evaluation issues are documented in the repository README.

The experimental record under environment/ comes from our experimental campaign with Rensselaer Polytechnic Institute (RPI). RPI cleared its data for this release. Its correspondence is withheld.

Release scope

The following material is omitted or redacted. File hashes and redaction markers preserve the connection to the original records.

• Held-out worlds. The 96 held-out worlds are excluded. The released physics package reproduces every route outcome in the 32 evaluation worlds.
• RPI correspondence. Messages, meeting transcript excerpts and message fields in delivery records are withheld. Quoted measurements and their methods or conditions remain.
• Article texts. Full texts of Wu et al. and Singh et al. are withheld. DOIs and transcribed numbers remain. Ikram et al. is included under CC BY 4.0.
• Provider-encrypted reasoning and signatures. Encrypted payloads and reasoning signatures are withheld. Readable reasoning and summaries remain as returned.
• Intermediate workspace files. The release omits 300 agent scratch files larger than 5 MB and 613 pickle files, listed by hash in WITHHELD.json. These are not needed to grade investigations or reproduce the published results.

The original agents had the correspondence and article texts listed above. New runs use the released context, so they do not recreate every information condition of the recorded investigations.

Redaction and world-package details

Correspondence is replaced with [withheld:correspondence sha256=…], article text with [withheld:article:… sha256=…] and encrypted reasoning or signatures with [withheld:provider-encrypted sha256=…]. In traces, attributed withheld quotes and verbatim runs of 64 or more characters are replaced. Shorter paraphrases in the agent’s own words remain.

The world package keeps the original file shapes. Posterior and anneal rows unused by the evaluation worlds are set to NaN, and the anneal refit’s seed is removed. A row used by an evaluation world can also be sampled by a held-out world, but the held-out world configurations are not released. The package reproduces the frozen evaluation outcomes exactly.

AI-generated content

Assistant messages, reasoning and tool calls in the 192 language-model traces are AI-generated. Each trace has an AI_GENERATED.md label, and investigations.parquet identifies its model. Tool results, laboratory deliveries and grades come from the harness. The 64 classical-baseline traces contain no model output.

Uses and limits

Use this dataset to evaluate agents on SDL-2, study their investigations and check the published findings. The evaluation measurements are simulated. The twin’s density model is calibrated on the experimental campaign and its magnetic model on earlier samples. The experimental records used for calibration are included separately under environment/.

License and citation

The dataset is released under CC BY 4.0. Model outputs were produced under their providers’ terms. The code is released separately under Apache-2.0.

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