Section 16Experimentation and Evidence

Experiment Orchestrator requirements

Defines immutable runs, runtime-action plans, sweeps, random sampling, comparisons, parallel execution, calibration, and optimization hooks.

Section relationships
IDNamePriorityRequirementVerification
EXP-001Immutable run specificationMVPA launched run SHALL reference an immutable compiled scenario, controller identity, seed, initial state, metric set, and stop conditions.Inspection
EXP-002Run identityMVPEvery run SHALL receive a stable unique identity before execution begins.Test
EXP-003Stop conditionsMVPAn experiment SHALL support virtual-time, mission, safety, failure, convergence, and explicit user stop conditions where applicable.Test
EXP-004Early terminationMVPA metric or guard SHALL be able to terminate a run with a named reason at a safe scheduling boundary.Test
EXP-005Named metricsMVPAn experiment SHALL select metrics by stable identity and version.Inspection
EXP-006Reference rover metricsMVPThe reference experiment SHALL report mission success, elapsed time, trajectory tracking or declared reference-comparison measures, stability, slip indicators, torque, current, energy, and termination reason.Demonstration
EXP-007Finite parameter sweepMVPThe orchestrator SHALL generate and execute the Cartesian product or explicit list of selected parameter values.Test
EXP-008Random samplingBetaThe orchestrator SHALL generate repeatable samples from declared parameter distributions.Test
EXP-009Correlated samplingBetaThe orchestrator SHALL support documented correlation among selected parameters.Analysis
EXP-010RepetitionsMVPAn experiment SHALL support repeated runs with explicit seed policy.Test
EXP-011Variant comparisonMVPA result set SHALL compare metrics across robot, controller, and scenario variants without requiring identical internal backend state.Demonstration
EXP-012Independent-world parallelismBetaThe orchestrator SHALL execute independent worlds concurrently subject to deterministic mode and resource policies.Analysis
EXP-013Progress and cancellationMVPA batch SHALL expose progress and support graceful cancellation without marking incomplete runs as complete.Test
EXP-014Resume batchBetaA batch SHALL resume pending runs without repeating verified completed runs unless explicitly requested.Test
EXP-015Resource budgetBetaAn experiment SHALL support configured CPU, memory, GPU, storage, virtual-time, and wall-time budgets where measurable.Test
EXP-016Monte Carlo summaryBetaA stochastic experiment SHALL report sample count, seed policy, distributions, confidence summaries, and failed-run handling.Analysis
EXP-017Sensitivity analysisBetaBeforeMetal SHOULD support local or global sensitivity analysis for selected parameters and metrics.Analysis
EXP-018Parameter calibrationBetaBeforeMetal SHALL support calibration of selected parameters against a declared calibration dataset and objective.Analysis
EXP-019Held-out separationBetaCalibration workflows SHALL prevent designated validation observations from being used by the fitter.Test
EXP-020Optimization hookFutureBeforeMetal MAY expose experiment evaluation to an external design-optimization algorithm.Demonstration
EXP-021Pareto comparisonFutureBeforeMetal MAY report Pareto-efficient designs for multiple selected objectives and constraints.Analysis
EXP-022Runtime-action planBetaAn experiment SHALL support a versioned plan of stable-identity runtime-action requests whose type and version, origin, target or compiled definition, requested virtual time or trigger, typed payload, bounds, and invalid, late, or failure policy are fixed before the run.Test
EXP-023Calibration result recordBetaEvery completed calibration SHALL record parameter identities, bounds, initial and final values, residual definitions and weights, time alignment or delays, dataset identities, optimizer identity and settings, seed where applicable, termination and convergence status, final residuals, and uncertainty method and result.Inspection
EXP-024Practical identifiability diagnosticsBetaA multi-parameter calibration SHALL provide a local practical-identifiability analysis using a declared rank-revealing method and SHALL report parameter scaling, sensitivity source, observation weights, sample points, rank criterion, spectrum or conditioning, correlated or near-null directions, and an explicit unidentifiable result where applicable.Analysis

Change rationale (EXP-006): The previous unconditional “trajectory error” implied a physical truth dataset. The reference workflow always reports its simulated trajectory and may report error only against an identified analytical, published, manufacturer, community, independent, or optional empirical reference.

Change rationale (EXP-022): Weather timelines and fault injection already imply scheduled changes to live state, while topology and controller changes lacked one native, immutable authoring contract. The plan describes requests against a future world instance; it does not mutate the project or compiled scenario and does not grant omnipotent runtime access.

Change rationale (EXP-023): EXP-018 requires parameter fitting but not enough evidence to reproduce or judge the fit. A complete result record is necessary before identifiability diagnostics or a calibrated model can be interpreted honestly.

Change rationale (EXP-024): Parameter sensitivity alone does not establish whether selected observations distinguish the fitted parameters. The new diagnostic is explicitly local, method-specific, and permitted to conclude that parameters are unidentifiable; it is not a claim of global or structural uniqueness.

Generated from the canonical specification. Edit section metadata or prose in docs/requirements.md; the website rebuilds this page and its relationships automatically.