Skip to content

Pipeline findings

Pipeline findings is the per-session aggregation of every detection the Pipeline made during that Session.

"PII detector fired 4 times, redacted 2 emails and 1 phone number." "Prompt-injection step flagged 1 attempt with confidence 0.92." "Exfiltration heuristic matched 0 times."

Findings are derived, not stored separately — they're a query over the audit trail's pipeline-step events.

Derived fromAudit trail events tagged pipeline.step.fire
Scoped toone Session (typical) or aggregated across sessions
Grouped bystep id → severity → reason
SurfacesSessions UI per-session detail · per-agent rollup · compliance reports
Statusnow (basic), next (richer aggregation surfaces)

For a single session:

StepPhaseFiredOutcome
detect.pii.contactbeforeredacted 2 email, 1 phone (1 false-positive flagged)
detect.secrets.api-keybefore
defense.prompt-injectionbeforeflagged confidence 0.92, request blocked
redact.findings.responseafterreplaced markers in 2 chunks
cost.budget.sessionduringcontinuouswithin budget (8.4k / 50k tokens)

Per-agent and per-project rollups answer: "which steps fire most often for this agent? which detections recur? where are we most exposed?"

The Audit trail records every event — including detection details. Findings is the operator-facing projection of that data: noise-suppressed, grouped, classified, with a UI shape that maps onto pipeline-step ids rather than raw event types.

Findings is also the input to:

  • Compliance evidence. "Show me PII redactions in production over the last quarter" — a findings query.
  • Tuning. Operators iterate on step config based on findings (false-positive rate, missed detections).
  • Feedback to detection authors. Step authors see which detections fire most across customers.
PersonaSurfaceWhat they see
OperatorSessions detail page → Findings tabPer-session breakdown, expandable per step
OperatorProject dashboard → Findings rollupStep-level frequency, severity distribution
Compliance teamCompliance reportFindings filtered by compliance_tag (e.g., gdpr.art-32)
Step authorStep telemetry view (planned)Cross-customer fire rate, FP rate trends