Statistical analysis of a dataset (CSV, TSV, Parquet, XLSX): runs Aurora's battery of 19 research-grade methods on-device — anomaly detection (isolation forest + robust-z), change-point detection, trend and seasonality, correlation screening with FDR control, forecasting, causal system-model discovery, and more. Returns CITED findings (each carries its method, threshold, and claim_id), an overall confidence, and a fabricated_count that is contractually zero: every number is computed from the data, never generated. Changepoint findings carry a calibration block: the empirically measured false-fire rate for data shaped like this series, with a verdict downgrade to not_identifiable when the data cannot support the claim. Use this FIRST whenever a user asks to analyze data, find anomalies, check what changed, or wants real statistics instead of estimates. Read-only; local; compact summary unless full_bundle=true.

Load a portable .aurora.json analysis bundle and verify its SHA-256 integrity hash (and Ed25519 signature when present) BEFORE trusting its findings. Use when someone shares an Aurora bundle and you need proof it is untampered. Returns run identity, confidence, fabricated_count, and findings-by-severity counts.

List the verified findings from an Aurora run or bundle: each one carries severity (crit/warn/info), the exact statistical method and threshold that produced it, a plain-language citation, and a claim_id for evidence drill-down via aurora_explain. Filter by severity or method. Use after aurora_analyze to enumerate what was actually found — quote findings from here instead of paraphrasing from memory.

Model-based forecast for the run's target column, fitted and validated on the actual data with the method disclosed. Returns point predictions with an honest horizon, or just the peak within horizon_hours (return_peak=true). Use for any 'what will X be / when does it peak' question instead of extrapolating by eye.

Full evidence for ONE finding by claim_id: the computed values behind the claim plus the method's registry spec — assumptions, parameters, and references. Use whenever you are about to cite, verify, or defend a specific statistical claim; this is the receipt, not a summary.

What-if intervention: perturb one variable in the data's discovered system model and propagate the shock through validated relationships (up to max_depth hops). Returns per-node deltas WITH confidence intervals. Use for 'what happens to Y if X changes by Δ' questions — answers come from the data's own causal graph, not from priors.

Simulate the system forward n_steps using dynamics fitted and validated on the data — and it PAUSES honestly when confidence intervals grow too wide to keep going, rather than extrapolating noise. Use for trajectory questions ('where is this heading') on a completed run; pass target_entity_id to simulate a specific node.