Trust center
This page is maintained by the Aegis team and describes the behaviour of this build. It is documentation, not an independent certification.
Aegis cannot emit a recommendation without a traceable path from telemetry to driver to judgment. The workspace shows the proof graph, the ranked drivers, the supporting evidence lines and the validator rules that fired.
There is no generated prose standing in for reasoning. Every sentence in a report is assembled from artefacts the engine produced.
The engine is rule-based and structural. It contains no language model, no sampling and no randomness. The same dataset and question always produce the same recommendation, the same confidence and the same evidence chain.
Engine behaviour changes only when a new version is released, and versions are certified before release.
Analyses run in your browser. Your dataset is parsed, profiled and judged locally; it is not transmitted to an Aegis server as part of producing a recommendation.
Decision records, feedback and pilot context are stored in your browser's local storage under keys owned by this application.
You own your telemetry, your decision records and your outcome annotations. Aegis claims no licence over them.
Because storage is local in this build, exporting or clearing your data is entirely under your control.
The workspace keeps the most recent analyses in local storage and lets you delete any of them individually. Clearing site data removes everything Aegis has stored.
Retention terms for hosted or private deployments are agreed in writing during onboarding, not assumed here.
This build has no server-side account system and no customer database, which removes the corresponding attack surface. Data stays on the device that produced it.
Enterprise deployments — private hosting, SSO and audit logging — are scoped with your security team before any data leaves your environment. Aegis makes no certification claim it has not been audited for.
Customer data is never used to tune, train or calibrate the engine. There is no model to train: reasoning is deterministic rules over inferred structure.
Engine calibration is performed exclusively against the internal benchmark's synthetic datasets.
Aegis is designed to refuse. Insufficient coverage blocks judgment when the dataset lacks enough operational structure. The oscillation gate blocks judgment when telemetry reverses repeatedly with no durable trend. Premise conflict rejects a stated assumption that the evidence contradicts.
In each case the engine returns HOLD_JUDGMENT with the reasons, rather than a low-quality recommendation dressed up with a confidence number.
Aegis Engine v2.0 is scored against a fixed benchmark of adversarial, cross-domain scenarios with hidden oracles: missing upstream telemetry, contradictory metrics, correlation traps, false causation, competing root causes, unstable trends, pure noise, deceptive executive framing, and manufacturing, healthcare, retail and abstract causal datasets.
Scenarios are scored on five dimensions including root-cause precision, premise handling and refusal correctness. The release candidate certification run was executed three times with identical results, confirming determinism.
Questions we have not answered here
Security reviews, DPAs, residency and audit requirements are handled case by case. Nothing on this page substitutes for a written agreement.