Evidence is the product: verified, correlated, risk

Ask the console a question and you get an answer, not a data dump. But the answer is only useful if you know how much weight each sentence can bear. That is why every claim on the platform — in a Rapid Report, on the twin, in the API — carries one of three reserved evidence levels.

The three levels

• Verified (sage). Cross-confirmed on two or more independent sensors. A fence-line detection matched to drone video and a SAR change pair is verified. Safe to act on, and safe to defend to a board, an insurer or a court.
• Correlated (slate). Consistent with the pattern, but single-source or anecdotal. Two nearby yards reporting the same after-hours clustering is correlated — valuable context, not yet ground truth.
• Risk (terracotta). A forecast. Intrusion risk of 0.71 for the next 30 days is a probability, and it is labeled as one, with its confidence attached.

The colors are reserved. They never decorate a heading or brighten a button. When one appears, it means exactly one thing — which is the point.

Why alerts end in action

An alert with no evidence level is a demand for attention. An alert tagged verified, with the sensors that confirmed it one click away, is a decision already half-made.

This is also why a licensed human operator sits between AI detection and your phone. The AI watches every frame; the operator confirms what the mesh believes before anyone is woken up. False alarms end at our desk — and every dismissal feeds the baseline, so the same false alarm gets rarer.

A forecast is a claim with a number attached

The discipline that matters most is negative: a forecast must never masquerade as an observation. The risk color exists so that "we predict elevated intrusion risk" can never be misread as "we observed an intrusion." Persistent observation earns the right to predict; honest labeling keeps that right.