A human example
A service is unreachable even though its processes look healthy.
What the caller supplies
The agent proposes candidate causes and read-only tests, with a current state snapshot.
What happens next
Jev recommends a test or review. The agent runs the test, updates its hypothesis and verifies repair invariants.
Illustrative example, not a recorded result.
Potential value: high
Repeated incident and debugging work contains many choices about which evidence to collect next.
Evidence confidence: moderate
The SRE study measured more passing attempts overall, with regressions. Trace attribution also showed low joint diagnosis accuracy.
The rating describes support for this claim. It is separate from Jev's returned probability. How we assign ratings.
Evidence, including disagreement
- Find the likely failure in an agent trace. reported. Trace classification is promising, but most full diagnoses were wrong.
- Choose diagnostic tests and review repair evidence. reported. More attempts passed in one small study; some incidents regressed.
The next test
30 reproducible local failures with known causes, misleading logs and cases where the correct hypothesis is omitted.
Compare against
- Direct agent investigation
- Fixed diagnostic checklist
Measure
- Successful repairs
- Diagnostic calls
- Time to a supported cause
- Durable repair tests
Decision after the test
Require better task outcomes or fewer diagnostic calls without more incorrect repairs. Report omitted-hypothesis failures separately.
The report will retain inputs, question versions, every attempt and failure examples. We will update the confidence rating after reviewing the result.
Use a related Engine recipe
Recipes are implementation starting points. Their presence does not mean the protocol above has passed.