The brief
Six beats. One number each. Everything else on this site is evidence for these.
If you read nothing else here, read this page. It takes about two minutes. Each beat links to the surface that demonstrates it, so any number below can be checked rather than taken on trust.
A maneuvering threat does not have an impact point.
Classical impact-point prediction assumes one ballistic arc. A glide vehicle with a cross-range budget breaks that assumption on its first turn. A solver that answers anyway returns a precise, confident, wrong point — and precision is exactly what makes it dangerous, because a wrong point gets acted on.
The correct product is a footprint with a stated containment, not a pin.
See the footprint →The footprint collapses, and that collapse is the product.
Over twenty minutes of track custody the 50% containment area falls from roughly 656,000 km² to about 1,150 km². Three honest mechanisms drive it: the vehicle spends energy and has physically less ground left to reach, more observations tighten the state estimate, and the archetype classifier narrows the vehicle parameter band.
That third mechanism is the only place machine learning earns its keep here. It does not predict the future. It recognises which physical archetype is flying and collapses the prior.
Watch it collapse →Confidence arrives after the window has closed.
Scrub the console forward and the rule finally fires — and turns red, not green. By the time containment is high enough to justify committing an asset, the remaining time is shorter than the action takes to execute. This is not a bug in the demonstration. It is the finding.
So stop optimising the model.
Sensor detection through to a unit beginning to move takes about 336 seconds in the notional chain. Model inference is 1.4 of those seconds. Humans are 74% of it. Making inference ten times faster moves the total by a rounding error; taking ten seconds out of the operator handoff moves it more.
A capability pitch that leads with model latency is optimising a term that does not control the outcome.
See the timeline budget →The real metric is not accuracy. It is the exchange rate.
At a 0.30 alert threshold the watch floor absorbs 11.8 false alarms a month and modelled operator trust falls to 5% within six months. At 0.80 trust holds at 88% but 14 of roughly 15 real events a month go unwarned. There is no setting that avoids both.
Choosing the operating point is a command decision about which error is more survivable. It is not a vendor's decision, and a product that hides the choice has made it for you.
Move the threshold →And this is a demonstrator, not a capability.
It runs live, at speed, on realistic geometry, which makes it present like TRL 5–6. System readiness is the minimum of its load-bearing components, and the archetype classifier sits at 3 — it is trained on synthetic tracks generated by the same physics it is then used to constrain, which is circular and flatters the model.
The validation gate is closed, not merely incomplete. Verification can be finished with care and open references. Validation needs measured flight data, and no quantity of software substitutes for it.
See the readiness assessment →If you only do one thing
Open the console and press Play.
It opens on the problem state — many tracks, footprints in the hundreds of thousands of square kilometres, no rule firing. Playing forward walks through the collapse to the moment the rule fires and is already infeasible. That single pass is the argument.
What this is not
Stated up front, not in a footnote.
Not a targeting system — no aimpoints, weapon pairing, or engagement recommendations, and Target and Engage are struck from the kill chain on the console itself. Not validated — every corpus is synthetic and seeded. Not connected — no live sensor, no classified input, no backend. Not 6-DOF — point mass only. Not a program of record.