docs/README.md

STEM: eight domains, 40 tasks, 55 recorded Argon attempts

This package combines the latest five-task samples for Astronomy, Biology, Chemistry, Earth science, Engineering, Mathematics, Medicine, and Physics. All task folders share one tasks/ directory; all corresponding attempt folders share one trajectory/ directory. There are no domain-level wrapper directories, nested archives, duplicated runs/ trees, or original-ID mapping tables.

40 independent tasks · 55 recorded attempts · 19 passes · 34.55% pooled attempt pass rate.

The retained task and trajectory files are preserved byte-for-byte from the eight normalized five-task samples. Task IDs are unique, and each domain uses IDs numbered 1–5. This combines the later five-task Chemistry sample, not the earlier six-task Chemistry set whose IDs overlap. Shared grading utilities, the inventory under docs/, documentation, and checksums have been consolidated for this package.

Domain inventory and observed outcomes

Domain	Task IDs	Tasks	Recorded attempts	Passes	Attempt pass rate
Astronomy	ASTRO1–ASTRO5	5	5	2	40.00%
Biology	BIO1–BIO5	5	5	2	40.00%
Chemistry	CHEM1–CHEM5	5	10	2	20.00%
Earth science	EARTH1–EARTH5	5	5	2	40.00%
Engineering	ENG1–ENG5	5	5	2	40.00%
Mathematics	MATH1–MATH5	5	15	5	33.33%
Medicine	MED1–MED5	5	5	2	40.00%
Physics	PHYS1–PHYS5	5	5	2	40.00%
Total	Eight domains	40	55	19	34.55%

These were deliberately selected samples. The pass rates describe the included recorded attempts, not general model performance, independent estimates of task difficulty, or prospective success probabilities. A pass requires the complete answer to satisfy its included verifier; fractional field reward is not a pass.

The pooled rate weights every recorded attempt equally. It combines different capture modes and model labels and is not a common pass@1 or pass@k estimate. Separately, 19 of 40 tasks (47.5%) were solved at least once; that is a different measure from the requested 20%–40% per-attempt sample rate. Each domain's attempt pass rate remains within that range.

Directory layout

STEM-final.zip
├── tasks/
│   ├── ASTRO1/ ... ASTRO5/
│   ├── BIO1/   ... BIO5/
│   ├── CHEM1/  ... CHEM5/
│   ├── EARTH1/ ... EARTH5/
│   ├── ENG1/   ... ENG5/
│   ├── MATH1/  ... MATH5/
│   ├── MED1/   ... MED5/
│   └── PHYS1/  ... PHYS5/
├── trajectory/
│   ├── ASTRO1/ ... ASTRO5/
│   ├── BIO1/   ... BIO5/
│   ├── CHEM1/  ... CHEM5/
│   ├── EARTH1/ ... EARTH5/
│   ├── ENG1/   ... ENG5/
│   ├── MATH1/  ... MATH5/
│   ├── MED1/   ... MED5/
│   └── PHYS1/  ... PHYS5/
├── environment/
│   ├── __init__.py
│   ├── grade.py
│   └── stem_env.py
└── docs/
    ├── README.md
    ├── tasks.json
    └── SHA256SUMS.json

The ZIP opens directly to these folders; it does not contain an extra STEM/ wrapper. Each tasks/<TASK_ID>/ includes instruction.md, instance.json, task.json, verify.py, and solution/solution.json. Optional Mathematics reference solvers are retained where present. Per-task READMEs have been removed; docs/README.md is the single authoritative package guide.

Each trajectory/<TASK_ID>/attempt_<N>/ contains:

• trajectory.json: the available captured model/tool record in its original normalized format.
• submitted_action.json: the complete answer actually submitted for grading.
• result.json: the recorded pass/fail outcome, reward, and available diagnostics.

Chemistry has attempt_1 and attempt_2; Mathematics has attempt_1 through attempt_3; all other domains have attempt_1. Attempt numbering is local to a task. docs/tasks.json is the unified machine-readable inventory and contains each task's domain, paths, attempt count, and recorded outcomes. In that manifest, path and trajectory_path are relative to the package root. Within each task.json, instruction, instance, verifier, and reference-solution paths are relative to the task folder; trajectory_path, where present, is root-relative. Explicit file links inside a trajectory are relative to its attempt folder.

Grade answers offline

Use Python 3.10 or newer. All 40 numerical verifiers and the shared grading utilities use only the Python standard library. No network connection, private image, external dataset, or separately downloaded reference answer is required. The environment/ folder contains utilities, not a container image.

From the extracted package root:

# Grade all 40 bundled reference answers.
python environment/grade.py --all

# Regrade all 55 saved submissions and compare their recorded outcomes.
python environment/grade.py --check-trajectories

# Check every packaged file against the combined SHA-256 manifest.
python environment/grade.py --check-integrity

# Restrict bulk grading or trajectory checking to one domain.
python environment/grade.py --all --domain BIO
python environment/grade.py --check-trajectories --domain MATH

# Grade a new answer or one saved attempt.
python environment/grade.py --task CHEM1 --candidate candidate.json
python environment/grade.py --task MATH1 --attempt 2

# Run an individual task verifier directly, with an explicit candidate path.
python tasks/ASTRO1/verify.py tasks/ASTRO1/solution/solution.json

--domain accepts ASTRO, BIO, CHEM, EARTH, ENG, MATH, MED, and PHYS. Bulk checks use up to eight workers. The CLI returns JSON; bulk modes include a summary and per-entry results. For individual grading, exit code 0 means pass and 1 means fail. For --check-trajectories, exit code 0 means every saved result matches, including correctly reproduced failures. Invalid CLI arguments return exit code 2. With --task and no candidate or attempt, the shared grader checks that task's reference solution.

Read each task's full instruction and exact schema before submitting. Depending on the original normalized task, the schema appears as submission_schema or action_schema in instance.json. Keep required keys, units, numeric types, and any required problem_id exactly as specified. Mathematics answers can include arrays, integers, or rational strings; do not assume every task accepts only scalar floats. The per-candidate size limit is recorded in docs/tasks.json.

One-step environment

from environment import STEMEnv

# Fix a task; alternatively, use STEMEnv(domain="BIO") or STEMEnv().
env = STEMEnv(task_id="BIO1")
observation, info = env.reset(seed=0)

# Supply your answer as a JSON-compatible dictionary, JSON string, or bytes.
# observation, reward, terminated, truncated, info = env.step(answer)

One episode accepts one submitted answer. Call reset() before another episode. The environment is a convenience interface, not a security sandbox. For a blind assessment, expose only problem instructions and instances to the solving agent. Keep reference solutions, verifiers, saved trajectories, and result data in evaluator-side storage. Verifier feedback may reveal reference values.

Models, capture fidelity, and limits

Astronomy records identify the model as gemini-4-argon-high; the other domains identify it as gemini-4-argon. These labels come from the supplied records, not newly verified provider identities. No new model runs were generated.

Chemistry preserves one single-turn, no-tool attempt and one agentic Python-tool attempt per task. Mathematics preserves three recorded tool-use attempts per task. The remaining domains preserve one captured tool-use attempt per task. The trajectory.json schemas differ between source samples and are intentionally preserved rather than forcing heterogeneous capture records into a lossy format. The shared directory layout, complete submitted answers, and task-level inventory provide consistent access across all eight domains.

Medical capture limitation: the supplied medical records cap or truncate some tool-code and console-output snippets. These are not fully executable transcripts. Their complete saved submitted answers remain independently gradeable. The capture metadata identifies these limitations; missing code and conversation content have not been reconstructed.

More generally, captured reasoning summaries, tool traces, stdout, or stderr may be excerpts rather than complete API conversations. Available content, order, recorded usage, and timing are retained as provided. Regrading checks final answers; it does not rerun Argon, replay historical tool execution, or certify that the captured code is sufficient to regenerate an answer.

Optional Mathematics solution/solver.py files are reference implementations, not required by the grading commands. The MATH2 solver imports SciPy. Historical tool snippets may also depend on optional numerical libraries; those libraries are not required to grade the packaged reference or submitted JSON answers.

Selected task catalog

Outcomes below follow attempt_1, attempt_2, and attempt_3 where available.

Task	Area	Problem	Recorded outcomes
ASTRO1	Cosmology / recombination and CMB damping	Coupled Peebles Recombination with $s+p$-Wave Dirac Dark-Matter Annihilation & Silk Damping	Fail
ASTRO2	Strong-field gravity / black-hole superradiance	Kerr Superradiant $2p$ Scalar Cloud with Self-Gravity Shift, Boson Number & Annihilation GW Signal	Pass
ASTRO3	Compact stars / relativistic stellar structure	MIT Bag Model Strange Quark Star TOV Structure, ISM Dispersion & Bare-Surface Thermal Flux	Fail
ASTRO4	Interstellar medium / gravitational stability	Bonnor–Ebert Critical Stability Limit via Emden–Chandrasekhar ODE Turning Point & Submm Flux	Fail
ASTRO5	Exoplanets / orbital irradiation and radiative transfer	Eccentric Exoplanet Orbital Insolation: Flux-Averaged vs. Time-Averaged Temperatures & Periastron Flux	Pass
BIO1	Enzyme kinetics and biochemical networks	Theorell–Chance enzyme kinetics and coupled carrier-buffered thermodynamics	Fail
BIO2	Molecular biophysics and nonequilibrium thermodynamics	Single-molecule work fluctuations with Donnan partitioning	Pass
BIO3	Cellular electrophysiology and ion transport	Eyring ion-channel permeation with thermal and electrolyte linkage	Fail
BIO4	Population genetics and evolutionary dynamics	Underdominant-inversion traveling waves with viscoelastic and homeostatic coupling	Fail
BIO5	Ecological biogeochemistry and nutrient cycling	Soil nitrification–denitrification with crowding and mass-transfer coupling	Pass
CHEM1	Electrochemistry / impedance spectroscopy	Electrochemical impedance with coupled conformational equilibrium	Pass, Fail
CHEM2	Polymer / soft-matter thermodynamics	Polyelectrolyte hydrogel swelling and nonlinear mechanics	Fail, Pass
CHEM3	Theoretical / quantum chemistry	Radical-cation Hückel molecular-orbital analysis	Fail, Fail
CHEM4	Organometallic catalysis / reaction kinetics	Wilkinson-catalyzed hydrogenation microkinetics	Fail, Fail
CHEM5	Thermodynamics / reaction equilibrium	Non-ideal ammonia synthesis using a cubic equation of state	Fail, Fail
EARTH1	Earthquake physics / rate-and-state fault dynamics	Inverse Dieterich–Ruina Nonlinear Phase-Plane Excursion, Hopf Bifurcation, and Rebound Peak	Pass
EARTH2	Geothermal convection / porous-media stability	Inverse Variable-Conductivity Anisotropic Porous Thermal Convection Sturm–Liouville BVP	Fail
EARTH3	Glaciology / Antarctic ice-shelf hydrofracture	Antarctic Ice-Shelf Firn-Aquifer Meltwater Percolation & Hydrofracture BVP	Fail
EARTH4	Chemical oceanography / marine iron cycling	Hydrothermal-to-Abyssal Dissolved Iron Organic-Siderophore Stabilization BVP	Fail
EARTH5	Atmospheric chemistry / polar ozone depletion	Polar Stratospheric Cloud (PSC) Type Ia NAT Denitrification & ClONO2-HCl Activation BVP	Pass
ENG1	Magnetohydrodynamics and liquid-metal cooling	Hartmann-channel ohmic dissipation and microwave-transducer heat flow	Fail
ENG2	Geotechnical mechanics and optical sensing	Strip-foundation stress and causal optical-sensor pulse dispersion	Pass
ENG3	Information theory and signal processing	Gaussian source reverse water-filling and coupled quantum ground-state energy	Pass
ENG4	Control systems and electrochemical transduction	Control-loop stability and coupled electrostatic–osmotic free energy	Fail
ENG5	Chemical process engineering and electro-optics	Fixed-bed adsorption breakthrough and Pockels-waveguide pulse propagation	Fail
MATH1	Algebraic geometry / arithmetic curves	Frobenius Slopes and Hasse-Witt Invariants of a Genus 2 Curve over $\mathbb{F}_7$	Fail, Pass, Fail
MATH2	Partial differential equations / nonlinear elasticity	Radial Equilibria and Non-Homogeneous Annular Deformations for Polyconvex Hyperelasticity	Fail, Pass, Fail
MATH3	Probability / random matrix theory	Normalized Expected Sum of Singular Values of Gaussian Matrices	Fail, Pass, Fail
MATH4	Analytic number theory / modular forms	Analytic Invariants of the Normalized Ramanujan Delta L-Function	Pass, Fail, Fail
MATH5	Topology / knot invariants	Determinant, Signature, and Cyclotomic Tristram-Levine Invariants of Pretzel Knots	Pass, Fail, Fail
MED1	Cardiovascular hemodynamics	Womersley blood flow coupled to a three-element Windkessel	Fail
MED2	Renal physiology	Glomerular ultrafiltration with tubuloglomerular feedback and protein transport	Pass
MED3	Cancer genetics and Bayesian risk modeling	Two-hit retinoblastoma, mosaic selection, and sequencing-biased genetic risk	Pass
MED4	Medical imaging physics	SPGR MRI T1 mapping with field scaling, protein binding, and T2* decay	Fail
MED5	Clinical-trial biostatistics	Non-inferiority trials with constancy drift and crossover correction	Fail
PHYS1	Condensed-matter physics	Quantum oscillations and polariton wavepacket dispersion	Pass
PHYS2	Quantum optics and open quantum systems	Quantum-jump photon statistics and driven-probe heat transport	Pass
PHYS3	Kinetic plasma physics	Vlasov–Landau plasma damping and driven-probe heat transport	Fail
PHYS4	Fluid mechanics and hydrodynamic stability	Görtler boundary-layer instability and dispersive waveguide pulses	Fail
PHYS5	General relativity and spin dynamics	Relativistic gyroscope precession and dispersive waveguide pulses	Fail

Validation and integrity

All 40 reference answers pass the unchanged packaged verifiers. All 55 saved Argon outcomes reproduce, including pass/fail, reward, and applicable recorded field diagnostics. The 370 retained task and trajectory files are byte-identical to the eight normalized input samples; merging has not changed numerical answers, reference targets, tolerances, or historical content. Task IDs are unique and metadata file links are checked against the merged directory layout.

docs/SHA256SUMS.json lists every delivered file except itself. Its paths are relative to the package root, not to docs/. It is regenerated for this final package, not copied from any individual domain archive. Runtime Python bytecode caches are ignored by the integrity checker. No private images, duplicate run trees, nested ZIPs, or temporary build reports are shipped.

Passing a bundled verifier establishes compatibility with that benchmark's reference answers and tolerances. It does not constitute independent scientific peer review of the problem statements or reference calculations. The Medicine entries are benchmark problems, not guidance for patient care.