Precursors,
before they're patterns.
Five agents read your operation in real time: classifier picks the ICAO category, risk drafts the score, bowtie populates the threats and barriers, precursor surfaces the cluster building underneath, analyst answers questions in plain English. Every run is captured, citable, and replayable years later.
Locked at run.
Cited at audit.
Every agent run captures its inputs, output, cited source records and confidence score on a locked row. Years later, the auditor reads exactly what the safety officer saw — not a regenerated approximation.
Five moments.
No black-box appeals.
- When an occurrence is classified.The classification is captured the moment it runs.
The classifier reads the narrative against your enumerated taxonomy and writes one answer. That answer is captured along with what the AI looked at and which records it referenced — and never silently re-runs. If the analysis changes, that is a deliberate, signed event with a reason.
- When you ask the analyst agent a question.A cited answer with the underlying records attached.
Natural-language questions about the operation. The answer cites the records it was derived from — not "according to recent data" prose. Every assertion is reviewable, every citation links to the source occurrence, finding or audit.
- When the regulator asks how the AI made a call.Open the run. The full audit trail is right there.
No black-box appeals. Every AI run is a record: what the AI saw, what it concluded, the confidence, the cited source records, the time, the validating user. The audit lane opens the run. The auditor sees what the analyst saw on the day.
- When a pattern emerges across the operation.The precursor agent surfaces what you have not asked yet.
Latent conditions lining up across reports, hazards, SPIs and barriers. The agent surfaces the cluster as a typed finding, not a chart in a tab nobody opens. The post-holder gets a record, not an inkling.
- When the AI is wrong.Override is a signed event, with the reason captured.
AI gets things wrong. When it does, the override is a record — who overrode, when, why, against what new evidence. The audit trail is built into the workflow, not bolted on as an export.
Agents don't live
in a separate dashboard.
Every analysis attaches to the record it analysed; every alert routes to the post-holder who owns it. Four examples that matter the day a regulator asks for the trail.
Open the platform.
Replay a run with the founder.
See a classification run, an override, and the audit replay — on your data shape, with the model, the prompt and the cost open on the screen, in 30 minutes.