Technology

Search finds the past. Risk lives in the next step.

We first built scenario prediction the industry-standard way: an embedding model fine-tuned on 60,000 industrial events, clustering by similarity. It worked until it mattered: similarity can't see a scenario that has never happened. So we built an encoder that learns how risk propagates instead.

Principles

Four commitments the engine is built around.

Risk logic, not lookup

Retrieval systems answer 'what past event resembles this?' Useless for the event that has no precedent. The Aegis risk encoder is trained to answer a different question: given this source, this flux, this target, this context, what can physically happen next? That's a property of the world, and it generalizes.

Generate wide, assess hard

The encoder proposes the full scenario space without regard to probability; nothing is unthinkable at generation time. Relevance is then earned through physics-based assessment. Creativity and rigor are separate stages, so neither corrupts the other.

Physics as the referee

Each scenario is scored by the best-fitting micromodel from a catalogue of 40+ (inundation, blast, dispersion, network reflow), selected against the data actually available for your system. Model choice is recorded per scenario, including what would improve it.

Auditable to the ground

Every output traces to its risk pair, model, inputs and thresholds. Regulators, review boards and insurers get an evidence chain, not a confidence score asking to be trusted.

The loop

Every target is the next source.

One pass of the engine, drawn honestly: a hazard source reaches its targets through defined fluxes; assessed-relevant scenarios re-enter the loop as new sources until the tree closes.

Disclosure policy

What we publish. What we don't.

You're evaluating a vendor whose engine you can't fully inspect, so we're explicit about where the boundary sits and why.

We publish

The method lineage: MADS-MOSAR's source → flux → target reading of risk
The training corpus scale: 60,000+ structured industrial event records
The pipeline: generation, physics-based assessment, propagation to closure
The evaluation bar: scenario recall and granularity versus expert-built trees

We keep

The encoder's architecture and training objectives
The data formatting and representation work that makes it perform
The heuristic core's model-selection internals
Benchmark internals ahead of publication

Design partners see more under NDA, including evaluation runs on their own system, before committing.

Skeptical? Good. Bring a scenario our engine shouldn't be able to see.

The strongest evaluation is adversarial: give us a configuration outside our training distribution and judge the scenario tree yourself. That test is part of every design partner onboarding.