Iterate the design. Test the autonomy. Ship the data. Package the evidence.

Freeze a sheet per variant, run every variant against one scenario matrix, rank them by envelope. Forty variants overnight is forty frozen sheets against one matrix.

Every report names the limiting factor on every run, so the next iteration is a decision, not a guess. Thirty-six scenarios simulate in seconds; results are unvalidated design-iteration estimates, and the report says so.

Plug your guidance into the policy API — observations in, setpoints out — and fly it through detection sites and GNSS or datalink denial zones. Two baseline policies ship so every report has an A/B: a naive direct route against a cost-map replanner that trades endurance for terrain masking.

Firmware-in-loop via PX4 SITL is the higher-fidelity tier: the actual autopilot code flies the simulation. Threat models are generic capability classes with fictional presets; your real threat data stays inside your enclave.

Every rendered run is a labelled dataset: pose-stamped RGB and depth with a manifest tying each frame to its run and scenario. Batches stay headless; you pay for rendering only on the runs you select.

Manual flights with a controller log schema-identical trajectories, so operator demonstrations arrive in the same format as autonomous runs — imitation-learning starter sets from the same session. Every dataset ships with written sim-to-real caveats.

For a programme milestone you need something you can hand over. The evidence bundle packages the report, chart PNGs, results.csv, selected trajectories and photoreal replays, with a provenance appendix that chains every number to a frozen sheet and the hashes of its source documents.

What it is not: certified evidence. Results are unvalidated design-iteration estimates, the disclaimer footer cannot be removed from any export, and formal verification, validation and accreditation is deliberately deferred until a validation programme has published its deltas.