Angel Eyes is being built to tell you what made a radio signal - and to say so plainly when it can't.
Angel Eyes listens to the radio link between a drone and its controller and identifies the drone when it recognises it. Flagging a drone it has never seen, instead of guessing, is what it is designed for, but that is not yet reliable. It detects and hands off. It does not jam, spoof, or engage.
The aim: for a signal it has never seen, state what kind of object could have made it - a control link that hops between channels, a wide steady video stream, or something that is not a consumer drone at all - with a confidence level. That is the goal, not a finished capability. What exists today is below: identifying known drones, an unreliable first attempt at refusing unknown ones, and an early measurement of signal width. The step from those to naming the cause has not been built or tested.
Why refusing to guess matters
Most classifiers are built to always answer. Shown a drone they have never seen, they return the closest thing they know, with high confidence. In counter-drone work that is the dangerous failure: a confident wrong name is worse than an honest "unknown", because it ends the investigation instead of starting one.
Angel Eyes is built and measured around that second case - the drone that is not in the catalogue. Measured honestly, it is also the hardest part: see Limits.
Where it stands today
Mean of four different known/unknown splits (seeds 42 to 45), public RFUAV dataset. Each split is on the Evidence page.
Scope, stated plainly
- Does: detect a drone control link, identify the model when known, try to refuse when unknown (not yet reliable, see Limits), hand the result to partner hardware.
- Does not: jam, spoof, take over, cue a weapon, or intercept. Not now, not planned.