AAO feasibility study · AI engineering experiment
EmberScope is the test case: give AI agents an engineering specification for a drone-mounted long-wave infrared fire-spotting payload, measure how far they can take it using committed models and public data, and expose every material point where human direction changed the work. Agents produced calculations, traces, code, CAD, and review files; humans remain responsible for what is published. This is research evidence, not accepted engineering.
What the experiment records
Codex and Claude agents implemented the ordered queue: models, bounded optimizer runs, source-hashed exports, tests, renders, and public review artifacts. This describes who performed the work, not who has authority to accept it.
Greg set the comparison target to 100 mm EFL at F/1.8 on 24 July and later directed that downloadable CAD become the primary public value. He sets the charter, course corrections, and publication boundary.
The AAO feasibility-study charter defines the research task. Mission values are recorded study defaults until a dated Dani or RFS answer supersedes them. No RFS operational acceptance is recorded.
Five bounded three- through seven-mirror families were measured at the common target; zero passed the P2 gates. The failed traces, exact shortfalls, and reproducible artifacts remain public.
People decide whether evidence is credible, whether requirements should change, and whether anything proceeds to manufacture or field use. Publication does not make an AI output accepted engineering.
Mission problem
The working mission is a 150 mm, sub-2 kg thermal payload that can support one-hour revisit planning, controlled false alarms, and replayable evidence for a smouldering target over rural backgrounds.
That makes this a camera-level decision: optics, detector, calibration, data pipeline, payload structure, and field testing have to close together.
The current evidence rejects the five bounded optical families at the stated gates. It supports either another declared candidate family or a human decision to stop or reframe the custom-optics path; it does not support fabrication.
Current evidence
The five bounded 100 mm F/1.8 candidates all fail one or more common P2 gates. The latest three-mirror result measures 23.5% minimum throughput, 2294.3 µm worst detector-plane RMS, and 4919.5 µm flat-focus spread.
The parametric assembly measures 148 mm outer span and 1.643922 kg modelled mass, with no axis-aligned component or reserved-clearance violations. It remains an engineering-development model, not a manufacturing release.
GSD, NETD, calibration drift, and false-positive scenes still need measured closure. No hardware exists yet.
Main technical risks
Four ray-path rules are first-class evaluation gates: non-vignetted propagation, small fold angles where practical, detector-plane placement, and flatter image surfaces.
Detection claims need raw or pre-automatic-gain-control thermal frames, NUC state, references, and replayable alert records.
Bench and surrogate targets must precede controlled burns, and no-fire rural scenes must be measured beside positive detections.
Technical archive
Requirements, survey assumptions, procurement, and decision fit.
Baseline, evaluation stack, multi-mirror branch, literature, and mechanical feasibility.
Detector choice, raw data, calibration, bandpass, and end-to-end modelling.
Physical allocation, field evidence, dataset policy, files, and review package.
Next AAO decisions
Confirm target area, altitude authorisation, payload count, false-alarm tolerance, and whether the 150 mm cube is a hard requirement.
Choose the next declared candidate family or stop or reframe the custom-optics path. None of the five failed P2 designs is eligible for tolerancing or vendor review.
Lock the detector interface enough to replace generic image-energy assumptions with measured FPA and calibration behaviour.
Approve the bench-to-field route before controlled burn claims or operational readiness claims are made.