Platform · Findings

Failures become tests. Tests become PRs.

Repeated production failures cluster into a finding with a suspected cause, a repro test, and a remediation PR. Engineering work, not another trace table.

Four steps, one workflow.

Cluster

Detectors mine ClickHouse traces every 6h: silent failure (promised action without tool call, tool error swallowed, no-progress loop, claimed completion without tool), behavioral drift (output length z-score, refusal-rate spike, tool-call drop), and red-team failures.

Explain

Each finding ships with a suspected cause derived from the cluster's shared evidence. Not 'something looks wrong' — 'tool schema allows money movement before identity confidence is verified, seen in 7.3% of refund conversations.'

Repro + remediate

Auto-generated repro test ships to your repo at vouch-tests/<finding-id>.spec.ts. Vouch-bot opens a remediation PR with the policy gate snippet. Tests stay in CI.

Triage

Findings inbox with severity, source (REDTEAM / FIREWALL / INTENT / CERT / SBOM / NETWORK / vouch-ai), supporting traces, and triage actions. RBAC-gated for security review.

Findings ship in your inbox, not a slide deck.

Severity, source, supporting evidence, and a PR draft — all in one place.