Automated findings

Production failures, filed as issues.

Provon diagnoses complete conversations, groups the ones that share a cause, and opens a tracked issue with the evidence already attached.

Read the findings guide Connect a tracker
Claimed success after tool error high · 7 occurrences

0.92

confidence

7

occurrences

3

evidence spans

run_tests exit 1
Final answer "all tests pass"

Create work item

GitHub
GitLab
Linear
Jira

From trace to ticket

Evidence first. Routing second.

  1. Observe

    01

    A settled conversation becomes one evidence snapshot.

    $ provon findings discover

  2. Diagnose

    02

    Five built-in rules compare tool results against claims.

    $ provon findings rules list

  3. Cluster

    03

    Repeat symptoms group under one failure mechanism.

    $ provon findings list --min-score 0.7

  4. Hand off

    04

    A confirmed finding becomes a tracked issue.

    $ provon findings handoff <id> --provider linear

Built-in rules

Five lenses on every completed run.

Runtime reliability

False success after a failed step

Tool correctness

Failed tool loops and ignored errors

Task fulfillment

Skipped verification, early completion

Answer grounding

Final claims without support

Conversation health

Runaway context, no convergence

Signal quality

Actionable, not just alarming.

Evidence attached

Every issue carries cause, trace refs, and quoted spans.

Deduplicated

Replaying the same activity never inflates the count.

Verified closure

A merged GitHub PR with passing checks can resolve it.

Cost controlled

Deterministic-only mode screens without a model call.

Make production failures actionable.

Review one Finding and send it where work happens.

Read the findings guide