s42chen/sci-run

Science Speedrun: 42 completed task histories

Reproducible submitted solutions, native evaluation reports, and screened agent logs from GPT-6 Astra (Codex, xhigh reasoning; 22 histories) and from Claude Code (Fable 5.2 requested, xhigh effort; 20 histories, see "Claude Code runs" below). These are file artifacts, not a tabular training dataset. The histories contain 910 inference-distinct candidates; each has a completed sealed evaluation. Existing additional measurements of duplicate packages are retained. reviews/fairness-20261031/ holds the operator-side fairness review of every first-round trajectory (per-task verdicts and the cross-cutting infrastructure findings); each first-round index carries its verdict under operator_fairness_review, and a fresh trial's index names the first trial it repeats under fresh_trial_of.

Task	Agent · model	Distinct candidates	Artifacts
atomistic-materials	Codex · gpt-6-astra	3	Files
atomistic-materials	Claude Code · claude-fable-5-2	11	Files
brain-computer-interface-decoding	Codex · gpt-6-astra	127	Files
brain-computer-interface-decoding	Claude Code · claude-fable-5-2	14	Files
crystal-structure-generation	Codex · gpt-6-astra	20	Files
crystal-structure-generation	Claude Code · claude-fable-5-2	10	Files
earth-observation-pretraining-transfer	Codex · gpt-6-astra	11	Files
earth-observation-pretraining-transfer	Claude Code · claude-fable-5-2	5	Files
fluid-flow-control	Codex · gpt-6-astra	11	Files
fluid-flow-control	Claude Code · claude-fable-5-2	12	Files
fruit-fly-neuron-reconstruction	Codex · gpt-6-astra	7	Files
fruit-fly-neuron-reconstruction	Claude Code · claude-fable-5-2	8	Files
genome-sequence-to-function	Codex · gpt-6-astra	33	Files
genome-sequence-to-function	Claude Code · claude-fable-5-2	26	Files
mass-spectrometry-molecule-retrieval	Codex · gpt-6-astra	21	Files
mass-spectrometry-molecule-retrieval	Claude Code · claude-fable-5-2	17	Files
molecular-pretraining-transfer	Codex · gpt-6-astra	14	Files
molecular-pretraining-transfer	Claude Code · claude-fable-5-2 requested, claude-opus-5 fallback (105 / 504 requests)	15	Files
natural-disease-progression	Claude Code · claude-fable-5-2	29	Files
natural-disease-progression	Codex · gpt-6-astra	13	Files
pde-surrogate	Codex · gpt-6-astra	60	Files
pde-surrogate	Claude Code · claude-fable-5-2	24	Files
probabilistic-global-weather	Codex · gpt-6-astra	9	Files
probabilistic-global-weather	Claude Code · claude-fable-5-2	4	Files
protein-design	Codex · gpt-6-astra	1	Files · superseded by the corrected trial
protein-design	Codex · gpt-6-astra	3	Files
protein-design	Claude Code · claude-fable-5-2 requested, claude-opus-5 fallback (16 / 1149 requests)	21	Files
protein-folding-prediction	Codex · gpt-6-astra	18	Files
protein-folding-prediction	Claude Code · claude-fable-5-2	12	Files
protein-sequence-modeling	Codex · gpt-6-astra	58	Files
protein-sequence-modeling	Codex · gpt-6-astra	9	Files · fresh trial after the fairness review
protein-sequence-modeling	Claude Code · claude-fable-5-2 requested, claude-opus-5 fallback (789 / 421 requests)	16	Files
single-cell-perturbation-prediction	Codex · gpt-6-astra	17	Files
single-cell-perturbation-prediction	Claude Code · claude-fable-5-2	17	Files
surface-code-decoding	Codex · gpt-6-astra	45	Files
surface-code-decoding	Claude Code · claude-fable-5-2	16	Files
target-based-drug-screening	Codex · gpt-6-astra	47	Files
target-based-drug-screening	Claude Code · claude-fable-5-2 requested, claude-opus-5 fallback (10 / 876 requests)	23	Files
timeseries-forecasting	Codex · gpt-6-astra	53	Files
timeseries-forecasting	Claude Code · claude-fable-5-2	33	Files
tokamak-magnetic-control	Codex · gpt-6-astra	7	Files
tokamak-magnetic-control	Claude Code · claude-fable-5-2	10	Files

Layout

tasks/<task>/<run-id>/ contains submissions/, evaluations/, logs/, finalization/, and one index.json. The index maps every original snapshot and research timestamp to its canonical package and actual evaluation, and records file hashes, executable permissions, source commit, images, resources, and deviations. plots/ contains the complete-history figures and their metric/reference CSV files. reproduction/files/ preserves the frozen verifier/adapter source files referenced by each evaluation. Scientific native reports are retained without rewriting scores. Fruit Fly evaluations also keep the final SHAPE stage's per-node predictions (evaluations/<snapshot>/disjoint-pair-shape/node-assignments.npz: skeleton node_id to predicted segment prediction_id, with that stage's metrics.json), so other segmentation metrics can be computed without rerunning inference.

Disease submissions include their required Python/CUDA runtime. Identical large weights and bundled dependencies are shared under shared/package-files/, as plain files, so repeated checkpoint contents need only be downloaded once. Small entry-point files remain browsable in each submission folder. Indexes map stored files back to their original paths and executable permissions. Reconstruct a package before execution. From a complete downloaded repository:

python restore_submission.py tasks/<task>/<run-id>/index.json S001 /path/to/new/submission

To fetch just one candidate, download its index.json and the helper first (keeping the directory layout), then add --repo s42chen/sci-run to the command above. With huggingface_hub installed, the helper downloads only that candidate's missing files. Use --revision <dataset-commit> to pin the download to a specific dataset version.

Use GRACE_FINAL for the separate post-budget package. The helper copies files into a new directory, restores permissions, and checks the exported names/modes/content fingerprint. It never executes the submitted code. Do not execute submissions merely to inspect this archive; use the benchmark's isolated evaluation environment.

Interpretation and limitations

All runs used one accessible H100 and 48 hours of charged research time. All tasks except Tokamak, Disease, Earth, and BCI were run with reduced CPU/RAM/GPU allocations relative to their pinned eight-GPU task configurations; those four match their tasks' one-GPU research requests at their recorded source commits. Runs were interrupted and resumed (operator pauses, Beaker preemptions, one shared-storage outage); preserving the clock does not reproduce uninterrupted computation. The fairness review under reviews/ quantifies the operator-caused losses per task and names the trials judged materially unfair. Fresh trials of those tasks were run from scratch (new session, same images, data, and budget) and are added as separate run directories next to the retained first trials, as each fresh history is fully scored: Protein Sequence (run 20261031a) is included; Weather, Genome, Atomistic Materials, and Mass Spectrometry will follow. Brain-computer interface decoding (BCI), the nineteenth task, ran once at its native budget; its five tracks are scored and plotted separately. Beaker execution uses documented isolation/resource adaptations and may grant more RAM than requested. Published baseline lines are reference measurements, not equal- budget human trials. Missing or incompatible baselines are explicitly omitted.

The first Protein Design trial (run 20261022a) ran without its pinned pretrained inference assets mounted; it is retained for the record but superseded by the corrected trial (run 20261027a), which is the Design result shown in the plots. finalization/ is an additional submission-only window of at most five minutes; it is not backdated into the 48-hour curves. Purple markers denote those reviewed post-budget results; orange marks denote reused scores for duplicate snapshots.

Two results carry operator rulings (2026-11-02). Tokamak is scored under the task's metric v3 that the trial was pinned to (35 sealed episodes, timed-out episodes voided, so per- candidate denominators vary from 32 to 35); the task specification moved to v4 (105 episodes, no voiding) after the trial, and no rescoring was performed. The PDE surrogate final package, from its sixteenth candidate onward, routes most datasets by case metadata to independently implemented equation operators declared in its manifest; this is permitted by the task contract and the scores stand, but from that point the reward mostly measures re-derivation of the governing equations rather than a learned surrogate.

Claude Code runs

Run folders named <task>-fable52-xhigh-1h100-48h-2026111* are Claude Code trials (agent version 2.1.319, launched 2026-11-10): one fresh session per task with no prior work reused, on the same images, data mounts and 48-hour one-H100 research budget as the Astra trials. 20 of the 20 trials are included; a trial is added once every distinct snapshot of its history has been scored. Target-based drug screening is the twentieth task; both agents ran it on one H100 against the eight the task declares, like the other eight-GPU tasks.

The model requested was claude-fable-5-2 at xhigh effort. In 4 included trials, API safeguard refusals of Fable 5.2 requests made Claude Code fall back to claude-opus-5 for the rest of the session: Molecular Pretraining (within minutes of launch; 105 requests answered by Fable 5.2, 504 by Opus 5); Protein Design (within minutes of launch; 16 requests answered by Fable 5.2, 1149 by Opus 5); Protein Sequence (about 11 hours after launch; 789 requests answered by Fable 5.2, 421 by Opus 5); Drug Screening (within minutes of launch; 10 requests answered by Fable 5.2, 876 by Opus 5). Read a run answered mostly by Opus 5 as an Opus 5 result and the others as mixed-model runs; each index records the counts under model_fallback and the table above shows them. Each index also carries agent, model_requested, scoring_hardware and caveats. Differences from the Astra runs are stated there: PDE surrogate snapshots were scored on one A100 (allowed by the task; one snapshot scored on both differed by 1.4e-5); Protein Folding and Molecular Pretraining snapshots were scored on one H100 after a matching eight-GPU replay; Fruit Fly evaluations mount the H100 build of the trusted point-sampling library. Two trials carry a completion correction, with the original failure records kept under logs/completion-correction/: the launcher wrongly marked the finished Molecular trial failed, and the Weather trial was stopped by its supervisor 430 s before the cutoff (its last snapshot is accepted as final; the clock was not reset). Evaluations that were replayed after a cluster interruption are listed per snapshot under evaluation_replays. logs/sessions/ holds Claude Code's own session transcripts, screened like the Codex logs. plots/claude-score-over-time.* shows score over research time for the scored Claude Code histories, and plots/claude-fluid-flow-control.* shows Fluid Flow's flow cases, which have no single score; the CSV files beside them hold the plotted values and the baseline references.

Provenance and excluded material

Submissions were produced by the evaluated agent; reports by the native scientific scorers under the recorded adapters; logs by the agents (Codex, Claude Code) and the harness. Organization, indexes, plots, and the reconstruction helper were created by the operator. Logs are screened for credentials, with redaction counts and original/exported hashes recorded in the indexes. Authentication stores, private test inputs/labels, raw research workspaces and unsubmitted checkpoints are excluded. Dataset inputs and separately distributed pretrained dependencies remain at their upstream locations. Existing source and dependency licenses continue to apply; no blanket relicensing is asserted here.

Benchmark source: https://github.com/sunnweiwei/science-speedrun

Lean export

Mass Spectrometry packages omit unused research backups and unselected generations. The selection follows the same reviewed inference footprint used to deduplicate evaluations. Every retained inference file, mode, and relevant verifier metadata was checked against the recorded evaluated identity; required notes and adjacent licenses are also retained. No model code, weights, paths, or scores were rewritten. These trimmed packages have their own export fingerprint, with the original full-package fingerprint and evaluated inference identity recorded separately in each index. Other tasks retain their exact submitted package contents, with identical files shared. Original full run artifacts remain preserved on the operator's storage.