Hypothesis engine
Live hypotheses#
| Hypothesis | Verdict | Confidence | Basis |
|---|---|---|---|
| General cleanliness | Rejected | 18% | Morning cleaning completed |
| Missed cleaning rounds | Rejected | 8% | Morning cleaning completed |
| Supplies depleted by Thursday | Validated | 92% | Soap and bags near zero by Thursday |
| Cadence mismatch | Validated | 88% | Supplier arrives Monday only |
Why rejected hypotheses stay on screen#
A validated finding means more when you can see its rivals. Rejection at 8 percent confidence, with the operational log that caused it, is a stronger argument than a confident assertion with nothing behind it.
It also protects against the failure mode of analysis by conviction, where the first plausible story wins because nobody tested the second.
Each hypothesis generates its own questions#
A hypothesis that cannot be settled with data on hand produces the follow-up question that would settle it. One question tests one explanation:
| Tests | Question | Outcome |
|---|---|---|
| General cleanliness | When you say dirty, what specifically did you notice? | Answered |
| Missed cleaning rounds | What time did you visit on Thursday? | Answered |
| Supplies depleted by Thursday | What specifically was empty or overflowing? | Validated |
| Cadence mismatch | Have you noticed this issue before, or was this one-time? | Validated |
This is the loop that closes: hypothesis to question, question to evidence, evidence to verdict.
Continuous, not on request#
Agents keep forming and testing hypotheses as new signal arrives, so a confidence score can move before anyone thinks to re-run an analysis.
Next#
- Adaptive surveys covers how the generated questions get asked.
- Read the evidence shows how to audit a verdict.