| name | computational-investigation |
| description | Methodology for running a rigorous, tiered computational-science investigation — DFT/MD/ab-initio/simulation campaigns that test competing hypotheses about a material, molecule, structure, or mechanism and end in a testable prediction. Load when planning or executing a multi-stage compute investigation (materials/chemistry/physics): enumerating hypotheses, choosing methods + compute tiers, gating input-model validity before production, separating mechanism from root cause, designing the confirming experiment, and writing claims without overclaiming. Distilled from a proven real-world investigation plan; domain-agnostic. |
Computational investigation — the tiered-rigor playbook
The shape of a good investigation: enumerate hypotheses → discriminate at the cheapest tier that can → escalate rigor only on the survivors → separate mechanism from root cause → end with a designed, testable prediction → anchor every claim. Most "AI science" stops at a conclusion; this playbook ends at a falsifiable experiment an experimentalist can run.
The one governing principle
Use the lowest-cost method that can answer the current question — but never sacrifice model validity. A valid cheap tier that can't yet decide is fine; that's what escalation is for. A cheap invalid model is worse than no result — it makes every downstream number wrong.
The loop
- Frame the question + enumerate hypotheses — the human's and your own. For each, state the observable that would confirm or kill it. A hypothesis with no discriminating observable isn't yet a hypothesis. Add candidates the human didn't raise.
- Validate the input model BEFORE any production compute (Gate G1). Source data traceable + preserved unchanged; reconstructed models physically valid (stoichiometry, charge, geometry, no clashes); every construction step logged (source hash, recipe, seed). No production run starts before this gate.
- Tier the compute — and each tier declares what it CANNOT conclude.
- Screening (cheap, consistent): remove invalid / obviously-wrong candidates, scope cost & convergence. Do-not-infer: final values, fine orderings.
- Production (converged, common settings): reportable values for the survivors only.
- Add higher tiers (tighter functional/basis, free energy, dynamics) only when a lower tier genuinely cannot separate the live hypotheses. Don't burn HPC hours proving what a screening run already settled.
- Cheap pilot first. Before the ensemble: one single-point/short run to measure real cost, memory, convergence — then a sensitivity test (cutoff / k-mesh / timestep / box) on one representative case. Size the campaign from measured cost, not a guess.
- Escalate on survivors; separate mechanism from root cause. When two hypotheses survive, ask whether one is a mechanism through which the other operates vs a root cause — and pivot the higher-rigor calculation onto the root-cause candidate.
- Honour the resolution limit (Gate G3). If candidates can't be separated within method + experimental resolution, do not force a ranking — state the ambiguity and recommend the characterization that would resolve it. A well-supported "we can't tell yet, and here's how to find out" is a real result.
- Design the confirming experiment. Close by proposing a distinct, testable prediction — e.g. a novel composition / substitution / molecule whose computed signature would confirm the surviving hypothesis beyond reasonable doubt. This is the hinge from discovery to qualification; it is the deliverable, not an afterthought.
Decision gates (question · pass · if-failed)
- G1 Input — model physically/structurally valid & traceable? → exact stoichiometry/charge, integer occupancy, provenance. Fail → stop and rebuild; no production.
- G2 Method — settings converged enough for the intended comparison? → energy/force/stress stable, baseline sane. Fail → raise cutoff/sampling, change solver, test the correction (dispersion, etc.).
- G3 Interpretation — candidates separable within method + experimental resolution? → a robust difference or a well-supported ambiguity statement. Fail → don't force a ranking; recommend new characterization.
- G4 Confirm — does the prediction/perturbation survive arrangement + method checks? → trend consistent across cases/settings. Fail → report as a single illustrative microstate only.
- Final — can a new person reproduce it and understand the limits? → archive + report + figures + README + citations complete. Fail → hold the handoff.
Disciplines that keep it honest
- Artifact-first. The foundation is primary data, commits/diffs, logs, changed files — transcripts/scrollback are enrichment. Anchor every claim to a reference, a file, or a gate.
- No overclaiming. Validate against experiment with absolute error + a physical explanation of the residual (zero-point motion, finite-T, disorder, functional error) — never a bare pass/fail. A computed result is not material qualification.
- Common settings before any relative comparison. Never compare energies/observables computed with different settings.
- Package around decisions and observable predictions, not a list of outputs.
Compute routing
Route heavy jobs to whatever HPC/cloud resources are available; keep quick checks local. Escalate up a tier only when a gate says you must. For the live hypothesis stack + collaborator stance, pair with the scientific-collaborator rule.
