Hypotheses
Test competing explanations against production evidence — don't jump to the first plausible story.
When should I use this?
- You have multiple plausible explanations for an incident.
- You want to see why one hypothesis strengthened while another weakened.
- You need epistemic transparency — supporting, contradicting, and missing evidence.
Requirements
- Active investigation with correlated evidenceRequired
- At least two competing hypotheses generatedOptional
- Topology and change signals for causal testsOptional
Steps
1. Build competing explanations
Do not stop at the first correlating metric. AutoObserve generates candidate hypotheses — for example deployment change, database degradation, dependency failure, or network issue — each with a confidence score and trajectory (strengthening, weakening, or unlikely).
2. Read the epistemic states
Each hypothesis links evidence in four states:
| State | Meaning |
|---|---|
| Supporting | Evidence consistent with the hypothesis |
| Contradicting | Evidence that rules against it |
| Expected but absent | Expected indicator missing |
| Missing / unavailable | Telemetry gap — reduces confidence |
3. Run the next test
When evidence is inconclusive, AutoObserve proposes the next test — for example inspect logs after deployment, compare checkout vs payment onset timestamps, or verify database saturation indicators.
Optional deep dive: See hypothesis trajectories on the Demo investigation or Platform investigation pages.
Did it work?
Troubleshooting
Problem
Symptoms
- No competing explanations
- High confidence without tests
Likely causes
- Insufficient evidence diversity
- Single signal type connected
Resolution
Verify fix
Problem
Symptoms
- Flat confidence curve
- New evidence not linked to hypotheses
Likely causes
- Evidence not correlated to incident window
- Identity mismatch across signals
Resolution
Verify fix