- A parametric flood policy pays millions the instant a satellite calls a site underwater. But cloud shadow, radar shadow, and a permanent reservoir darken the same pixels a flood does. A single-frame trigger can't tell them apart, so a $2M payout can fire on a shadow. 🧵
- So we built TriggerProof, the adjudication layer between a fired trigger and the payout. It takes the trigger as given and proves whether it should pay. A deterministic physics verifier in plain Python makes the call. The LLM only advises, and it can be overruled.
- The signature catch is a cloud shadow. Step through the acquisition frames and the dark patch is there only in the trigger frame, gone by the next pass, while SAR backscatter stayed normal throughout. The shadow moved at cloud speed. Real floodwater would have stayed.
- Five inspectable rules do the separating, no model vote: temporal persistence, SAR-optical agreement, DEM slope (water can't pool on a cliff), a permanent-water mask, and hydrological linkage to the drainage network. Each verdict names the exact rule that fired.
- When the physics is genuinely ambiguous, it doesn't guess. On a weak-signal site whose river gauge never crossed bankfull, confidence lands at 0.151, under the 0.65 automation bar, so the case escalates to a human with the evidence instead of a coin-flip $0.8M payout.
- On an 8-AOI storm book, the legacy trigger queued $8.0M in automatic payouts across 6 fires. TriggerProof confirms 2 real floods to PAY ($4.0M), denies 3 look-alikes, escalates 1. $4.0M of $8.0M is stopped or held for proof. (The AOIs are synthetic, physics-faithful.)
- On a fixed 60-case labeled benchmark (synthetic, physically faithful), it makes 0 unsafe automated decisions vs 48 for the single-frame baseline, auto-resolves 80%, and among those suppresses 100% of false positives and pays 100% of real floods. Not a field guarantee.
- Every decision emits a forensic dossier: data lineage, per-rule evidence, the false-positive elimination log, the ground-signal cross-reference, and a SHA-256 hash of the record, ready to hand a reinsurer or auditor. (Live retrieval and PKI signing are stubbed in this demo.)
- For the parametric desk: when confidence is genuinely ambiguous, should the payout auto-fire to keep the disputeless promise, or escalate to a human and eat the delay? We put the veto in code, outside the model. #ParametricInsurance #FloodRisk #InsurTech #RemoteSensing
- Watch the cloud-shadow filmstrip flip a $1.2M auto-payout to DENY, and the escalate case refuse to guess, in the live demo: veriprajna.com/demos/satellite-flood-intelligence
Published on X · August 19, 2026
On social media
See this post on its original platform
In our archive