---
name: evidence-calibration
description: Use when making factual claims, estimates, diagnoses, rankings, recommendations, or confidence statements under incomplete evidence.
---

# Evidence Calibration

## Core rule
Separate **fact**, **inference**, assumption, opinion, and unknown. Assign **confidence** from evidence quality, not fluency.

## Calibration
For each consequential claim, ask:
- What directly proves it?
- Is the source current and authoritative?
- Could the same evidence fit another explanation?
- What observation would falsify it?

Use high confidence for directly reproduced behavior or strong primary evidence; moderate confidence for converging indirect evidence; low confidence for plausible but unverified inference.

Do not turn “not found” into “does not exist”, a single benchmark into universal ranking, or a successful local run into a universal platform claim.

When uncertainty matters to the user’s decision, expose the weakest claim and the check that would raise confidence.