Estrix Risk Horizon

Risk teams aren’t short of information. They’re short of a defensible answer. Most tools tell you something is happening and then stop, which leaves the hard part exactly where it was.

A platform that takes a scenario, say a tanker route through a strait that has just become contested, and war-games it. Five agents work the problem. One reads the live signal. One plays out how events unfold. One proposes moves, three at a time, each with its arguments for and against. One adjudicates, weighing those arguments, throwing out the implausible ones and setting a probability in five percent steps. And one exists purely to argue the other four have got it wrong (this one is my favourite).

The disagreement between them is treated as the signal, not as something to smooth away.

The tradecraft underneath. Every likelihood has to start from a reference class and a base rate before the agents are allowed an opinion, and confidence is stated separately from probability, in the same seven bands the UK intelligence community uses. The report ends with a pre-mortem that actively tries to disprove its own conclusion.

There’s an end-to-end test harness that runs a full simulation after every deploy and checks the assertions inside the function, so a change to one agent can’t quietly break another.

The first version put everything on one screen at equal weight: feed, map, indicators, sentiment, agent activity, all competing for attention from the moment a session opened. It looked impressive in a screenshot and nobody could use it. Watching a real user get lost in it was the useful bit. The August rebuild turned it into a journey with four stages, brief, threat picture, war-game, report, each with one view, one question and one action. Everything else moved one deliberate step away.

Selling it needed its own system. The CRM lives in the same database, 1,200 contacts scored against an ideal customer profile, with a daily action sheet that tells me exactly who to reply to, who to follow up and who to invite, in order, with the text already drafted. The articles that feed the campaign are produced by running the platform itself on a live scenario and writing up what it found. Nothing sends without me pressing the button. That was a decision, not a limitation.