Persistent Overwatch
Without Cameras.

In 2013 DARPA disclosed a sensor called ARGUS-IS: a 1.8-gigapixel airborne camera array — reportedly on the order of 368 imagers — that could hold a medium-sized city in a single frame and let an analyst pull a “soda straw” window anywhere inside it, after the fact. It was the visible face of a broader idea the defense world calls wide-area motion imagery (WAMI): stop staring through one narrow lens at one moving thing, and instead record everything in an area, continuously, so you can rewind and follow any target you did not know was important until later.

WAMI works. Constant Hawk, Gorgon Stare, and their successors turned persistent overwatch from a fantasy into a program of record. It also collided, hard, with the law. When a WAMI system was flown over Baltimore, a federal appeals court ultimately held that persistently recording the movements of a whole city was the kind of dragnet the Fourth Amendment was written to constrain. The strategic capability was never really the problem. The cameras were the problem. Faces, license plates, and high-resolution video are what make persistent observation feel — and legally look like — mass surveillance.

So here is the question we build around at Argus Dynamics: what does persistent overwatch look like if you delete the cameras entirely?

The inversion: track the device, not the face

Almost everything that moves through a modern space is broadcasting. Phones and laptops probe for Wi-Fi networks. Wearables, earbuds, tags, and beacons chirp over Bluetooth Low Energy. Smart-home and industrial gear talks Zigbee. Aircraft squawk ADS-B; increasingly, drones broadcast Remote ID. None of this is video. All of it is signal — and signal can be observed passively, from the ground, with inexpensive radios, without pointing a single lens at a single human being.

That is the whole thesis. You can reconstruct a surprising amount of a space’s activity — how many distinct devices are present, when they arrive and leave, how they move relative to fixed sensors, which ones are new and which are part of the furniture — purely from the RF exhaust of the electronics people carry. It is the strategic shape of WAMI (persistent, area-wide, rewindable, target-agnostic) with the biometric payload stripped out. You are tracking an emitter’s signature, not a person’s face. In RF there is no face to search.

// the shape of the claim

Same strategic capability — persistent area awareness, target tracking, pattern-of-life. Different physics — devices, not faces.

Architecture: dumb edge, smart core

The temptation with any sensing system is to make the sensors clever. We do the opposite. The design principle is dumb edge, smart core: the radios at the edge do exactly one job — receive, timestamp, apply a coarse filter, and push. They make no decisions. All of the fusion, correlation, and identity logic lives in one place, on hardware the operator physically controls.

That gives a clean pipeline: many cheap, disposable sensors generate observations; observations become evidence; evidence is fused into one entity per real-world emitter. The unit of value is never the sensor. It is the entity — the external device you have detected and are trying to understand. A system that measures its own success by “number of sensors online” has confused the map for the territory. The metric that matters is how confidently, and how honestly, you can describe the things those sensors see.

The confidence triad, and why it is the whole product

RF is noisy. Signal strength wobbles with a passing body or an open door. Devices randomize their identifiers. A single sensor’s view is a probability, not a fact — and a system that renders probabilities as facts is not a situational-awareness tool, it is a confident liar.

So every entity in our model carries a confidence triad: an identity confidence (how sure are we this is one distinct device and not two, or half of one), a spatial confidence (how sure are we where it is), and a temporal confidence (how sure are we about when). These are numbers, timestamped, shown to the operator — not hidden behind a single reassuring dot on a map. The entire discipline of the platform is to raise those numbers over time through corroboration across sensors and modalities, and to refuse to pretend they are higher than they are. Our internal shorthand for this is blunt: observability precedes intelligence. You earn the right to reason about data only after you have measured how much you can trust it.

The honest limits (because elite means honest)

It would be easy, and dishonest, to end there and let you imagine a Gorgon Stare for RF. So here is the calibration, stated plainly to any technical reader.

• Scale. A ground-up RF mesh built from commodity radios is a property, building, or campus instrument today — not a thirty-square-mile aerial one. The WAMI comparison is a description of capability shape and a roadmap north star. It is not a claim of present-tense equivalence, and we say so to every buyer.
• Position is inference, not truth. Localizing a device from received signal strength is a heuristic. Real metric positioning — floor, elevation, a defensible XYZ with an uncertainty ellipsoid — requires time-of-flight ranging hardware (ultra-wideband anchors), and that is a build step, not a checkbox. Until a range fix is logged, spatial confidence stays honest and low.
• Identifiers fight back. Modern operating systems rotate MAC addresses precisely to defeat this kind of tracking. High-confidence identity is a layered, multi-modal, over-time result — never a guaranteed read of a true hardware address on first contact. Protected-management-frame networks can defeat active identification entirely. “One hundred percent identification” is an asymptote you approach with corroborating evidence and operator confirmation, not a feature you ship.

None of these limits are embarrassing. They are the difference between a research-grade instrument and marketing.

The privacy question we refuse to hand-wave

“Devices, not faces” genuinely removes the biometric-video liability that sank the airborne programs. It does not buy a free pass, and any vendor who tells you it does is selling you a lawsuit.

The reasoning that constrained persistent aerial imagery — that continuous, retrospective tracking of everyone’s movements is qualitatively different from a single observation — was not fundamentally about the medium being optical. In Carpenter v. United States (2018), the Supreme Court held that 127 days of cell-site location records were a search requiring a warrant, precisely because persistent location history offers “near perfect surveillance” and reveals the privacies of life. Point a persistent RF sensor at an occupied floor for a month and you could, in principle, infer shift patterns and co-location from device identifiers alone. That is the same totality concern, in a different band.

So we are precise about the defensible use, and we design for it:

// the defensible frame

We track emitter signatures, not individuals — for anomaly detection and asset inventory on a perimeter the operator owns, with the same minimization, notice, and retention discipline a responsible organization already applies to a badge-reader log. Consented, owned-space deployments — a substation, a data-center campus, a forward operating base — not covert public-area mass tracking.

Leading with that frame is not a compliance tax. It is the only version of persistent RF awareness that is worth building, because it is the only version that survives contact with a serious customer’s counsel.

Why it has to be sovereign

The last architectural decision follows from the first two. A system whose entire value is a high-fidelity map of who and what is present in a sensitive space cannot have that map leave the operator’s control. So the correlation core runs on hardware the operator owns; inference runs locally, air-gap-capable, with no dependency on a vendor cloud that could be subpoenaed, breached, or simply switched off. The same instinct runs through everything we build: produce evidence the operator can hold, verify, and defend — not answers they have to take on faith.

That is the wager behind persistent overwatch without cameras. The strategic capability the defense world spent a decade and a fortune proving out is real and it is durable. The liability that keeps cancelling it is optical. Delete the optics, keep the discipline — passive observation, a confidence triad instead of false certainty, an owned perimeter instead of a public dragnet, and a sovereign core — and you are left with something rare: situational awareness you can actually stand behind.

— Meta Strawman, Argus Dynamics