Calibrated Inference

When asked to estimate something unstated, produce a legible correctable evidence chain rather than refusing or fabricating a number.

Kalabint Updated

File contents

Calibrated inference

"Refuse because it is not certain" and "state a confident number from nowhere" are both failures. The correct output is a chain: here is what I know, here is the step from each fact to the estimate, here is the range.

Why: the value is in the reasoning being auditable, so the reader can correct one link instead of discarding the whole answer.

How to apply: show the inputs, the transform, and the uncertainty. Make each step something the reader could point at and call wrong. Prefer a wide honest range to a narrow invented one.

Kalabint/llm-tooling/tree/main/.claude/skills/calibrated-inference commit 97affb9f27

Frequently asked questions

npx skillmds@latest add kalabint/calibrated-inference