Detector Marketplace
Five detectors for the failures you can't see.
Most AI failures don't throw. Agents loop instead of stopping. PII leaks into prompts. Bias creeps in across a rolling window. Models hallucinate sources that look right. Prova ships five detectors that catch each of these, all running inline in TypeScript, no separate analyzer service to deploy.
What ships today
- Coordination loop. Tarjan SCC over the agent trace. Finds the cycle the moment it forms.
- Prompt injection. Regex and heuristic patterns against the inbound prompt, run sync.
- PII leak. Regex over the completion text for the standard categories, run sync.
- Bias drift. Rolling-window approval-rate divergence across protected attributes, sampled at 10%.
- Hallucination. Claude Haiku entailment check over RAG context, sampled at 20%. Requires ANTHROPIC_API_KEY.
What it does for you
Catch the failure mode that doesn't throw.
A LangGraph that loops forever doesn't crash. A model that emits PII doesn't return a 4xx. A bias drift is invisible to any single response. The detectors are the layer that turns these into signals you can alert on.
Inline, not an external service.
Every detector runs as TypeScript inside the ingest path. No separate Python service, no inter-service latency, no extra deployment. Async detectors get wrapped in a 1.5s timeout so a slow one can't stall ingest.
Sample what's expensive, run what's cheap synchronously.
Coordination loop, prompt injection, and PII are cheap and synchronous. Bias drift is sampled at 10%, hallucination at 20%. Sampling rates are tunable; production cost stays bounded.
Author your own detector.
The detector interface is a single function: receipt in, findings out. Drop a TypeScript module into the registry and it runs alongside the built-ins, with the same audit-trail guarantee.
Every detection writes a receipt
A detector catching a coordination loop emits the same signed AIDecisionEvent receipt as a policy violation or a model call. The Audit Vault is one stream, and detectors are one source feeding it.
That means a quarter-end report on "every AI failure we caught" is one SQL query against the receipt store, not five disparate dashboards.