Skip to content
NOT PUBLISHED

The engineering detail, shared in a conversation.

Enterprise teams reasonably want to know how this works before they trust it during a release. We will go through all of it with you: architecture, modeling, ranking, verification, failure modes, and where the method is weakest. We just won't put it on a page that anyone can read.

01
How the population is modeled

Why a thousand modeled users is not a thousand language-model agents, and where each technique, reasoning, behavioural, deterministic, conventional load, is actually used.

UNDER NDA
02
How failures are recognised

What counts as a failure when there is no expected screenshot, how invariants are inferred and confirmed, and how we avoid drowning you in false positives.

UNDER NDA
03
How incidents are ranked

The inputs behind users-at-risk and money-exposed, and why we do not publish the weighting.

UNDER NDA
04
How verification is proven

What we rerun beyond the original failure, and why a passing replay alone is not sufficient evidence that a fix held.

UNDER NDA
05
Architecture and isolation

Tenant boundaries, execution isolation, evidence storage, region pinning, and the model-data boundary in full detail.

ON REQUEST
06
Where the method is weakest

The failure classes we are honestly bad at, the ones we are improving, and what we will not claim to catch. We would rather you hear this before you buy.

ON REQUEST
Why we don't publish the method

A detailed public description of how we model populations, rank scenarios and generate neighbouring failures is a build specification for a competitor. It is also the part of the product that took longest to get right. You get the conclusions and the evidence, which is what a release decision actually needs, and under NDA you get the rest.

  • Scenario ranking weights and the confidence formula
  • How neighbouring failure scenarios are generated
  • Model routing and selection strategy
  • Anything learned from other customers' applications, which never leaves their tenant regardless
  • Internal prompts and behavioural model internals

Request access

A 45-minute technical review with an engineer, not a demo.

Mutual NDA before the call if you want one. We sign yours or send ours.

Request the deep dive

Already public

You don't need the deep dive to evaluate us. These are open to everyone:

Or just run it.

One free simulation on one application tells you more about whether this works than any architecture diagram. No credit card.

Start with a Reality Check