| name | modeling-problem-execution |
| description | Drive a concrete scientific modeling problem from supplied inputs through model selection, falsifiable plan, one-time approval, environment setup, smoke test, full execution, downstream analysis, provenance, and review. Use when the user provides sequences, structures, molecules, omics data, trajectories, or a specific modeling objective and expects an actual result rather than instructions. |
| license | MIT |
Modeling Problem Execution
Do not stop at a tutorial when a concrete problem and usable inputs are present.
Continue through execution after the required gate.
Intake
- Inspect supplied files/data and state the concrete problem, prediction unit,
deliverable, inference-time information, and non-goals. Ask only for missing
facts that materially change execution; choose and state safe local defaults.
- Search the audited catalog and load the narrow specialized modeling skill plus
required preparation/analysis skills. Prefer a current open implementation
whose modalities, license, hardware, and evidence boundary fit the problem.
- For a scientific claim, preregister the question, falsifiable hypothesis,
prediction, baseline, primary metric, and success threshold. Define the
smallest falsifying experiment before the expensive run.
One-time gate
Ask once with the exact package/repository and pin, weight/database downloads,
network hosts, private-data transfer, expected disk/GPU/time, and output path.
After approval, do not pause between reversible steps inside that scope.
Execute
- Create an isolated environment or digest-pinned container. Capture the code
revision, dependency lock, weight checksums, input hashes, hardware, seeds,
and configuration under
artifacts/<run-id>/.
- Validate schemas, chemistry/sequence alphabets, coordinate/feature mappings,
and privacy. Run the official smoke example or smallest real input first.
- If preflight succeeds, run the full input. On failure, diagnose and retry
bounded alternatives within the approved plan; preserve every failed attempt.
- Compare against the preregistered baseline or an orthogonal model where the
claim requires it. Never select a favorable model after seeing results without
labeling that comparison exploratory.
- Chain downstream work automatically: confidence/geometry and visualization
for structures; sequence design and independent refolding for generated
proteins; pose/interactions for complexes; donor/split-aware diagnostics for
omics; convergence/uncertainty for simulations.
- Record commands, environment, inputs, outputs, metrics, failures, and claims
with
$science-provenance. Run $science-review, resolve findings when
possible, then report results, limitations, and exact artifact paths.
Docking route
For receptor/ligand inputs, load $cx-molecular-input-preparation, the selected
docking engine, $cx-docking-validation, and $cx-plip-interaction-analysis.
Resolve the pocket in this order: an intended bound ligand whose use is valid for
the evaluation, user-provided residues or box, then a separately validated pocket
prediction. A different pocket is an interpretation-changing fork.
Prespecify ligand microstate and seed handling, pose-to-compound aggregation,
score ties, and sensitivity to receptor/pocket preparation. Redock a reference
when available. Without a valid positive control, label ranking exploratory.
Report the top result as the "highest-priority predicted docking candidate", not
the best binder; docking score and interaction counts are not affinity.
Stop conditions
Stop only for a new permission boundary, missing essential input, incompatible
hardware/license, unsafe data transfer, or an interpretation-changing fork.
Do not present installation success, a completed process, or model confidence as
a scientific result.
