Analyses failures, logs and stack traces to locate causes and propose fixes.
Works from a symptom — an exception, a failing test, an incident — toward a cause, reasoning across code, logs and recent changes.
Unfamiliar codebases, incidents under time pressure, and failures whose evidence spans several systems.
Novel failure modes with no precedent, where it will map the new problem onto the nearest familiar pattern with unwarranted confidence.
Where an incident requires a post-mortem record, a machine-proposed cause must be marked as a hypothesis until verified.
Anchoring on the most common cause. The rare cause actually responsible goes unconsidered because a plausible one was found first.
A matching run producing no matches, where the assistant proposes three causes — empty source file, schema change, rule misconfiguration — with the evidence for each rather than a verdict.
Two different absences share this shape. Foundational solutions get built whatever the domain, so no domain links them; the rest are solutions this domain genuinely does not reach for. v_ai_solutions_unlinked separates the two.