Skip to article frontmatterSkip to article content
Site not loading correctly?

This may be due to an incorrect BASE_URL configuration. See the MyST Documentation for reference.

Evidence-first investigation rubric

A correct scalar with unsupported reasoning is not complete. This rubric can be used analytically by dimension or holistically for a research memo. Adapt point weights to the course; preserve the distinctions among evidence classes.

Task-specific applicability

Before assigning a unit, publish which dimensions apply and their weights. A dimension may be marked not applicable (N/A) without penalty when the task does not ask for that reasoning—for example, derivative interpretation in a value-only activity or failure analysis in a tightly bounded identity check. Students should briefly justify an N/A claim; it must not be used to avoid a declared learning objective.

DimensionStrong evidenceDeveloping evidenceInsufficient evidence
Prediction qualityCommits to the task-relevant units, sign, limits, invariants, failure states, conditioning, and derivative meaning before executionPredicts a value trend but omits one or more applicable structural expectationsReconstructs a prediction after seeing the result or gives none
Model, units, and scaleSeparates model assumptions, parameters, state, observables, units, and numerical scalingNames the model and most units but conflates a role or scaleTreats arrays or formulas as self-interpreting
Method inspectionIdentifies the executed algorithm, tolerances, branches, status, shapes, and cost-relevant telemetryNames the method but retains only partial stateReports only the final scalar
Independent auditUses an actually independent identity, limit, FD check, convergence study, conservation law, or source recordPerforms a relevant check that shares important assumptionsRepeats the same computation under a new label
Derivative interpretationNames the differentiated map, held-fixed quantities, units, smooth domain, and finite-map versus implicit meaningComputes a derivative with partial interpretationTreats any finite AD output as proof of the intended derivative
ProvenanceRecords public API, configuration, precision, source/evidence ID, and reproducible commandRecords code and some configurationCannot identify which implementation or evidence produced the claim
Failure analysisUses a failed or boundary case to diagnose model, numerical, program, or evidence assumptionsNotices failure but offers an untested explanationHides, retries, or discards failure without analysis
Warranted claimStates the strongest supported claim plus explicit limitations and non-claimsGives a mostly calibrated conclusion with vague scopeGeneralizes beyond the tested method, domain, or evidence class

Feedback language

Prefer feedback that names the contract:

Minimal completion standard

Every submission should contain a pre-execution prediction, one units-explicit metric table, one independent audit, and a warranted claim. Include failure or boundary analysis when the task declares it applicable. Extensions may deepen the mathematics or science but should not replace the selected dimensions.