---
name: tangxiangru/answer-the-why-not-only-the-what
source: https://app.decimal.ai/s/tangxiangru-answer-the-why-not-only-the-what@1/SKILL.md
source_sha256: c0fdcb7cfda7
---

# A measured effect is half the result

Plenty of the questions a study is judged on are not "what is the number" but
"why is it that number": why two methods that should differ give the same gain,
why an effect is larger in one regime, why the improvement vanishes past a
threshold, what the failure cases have in common.

A report can measure such an effect precisely and score nothing on it, because
it never says what the effect is *evidence of*. Numbers are the input to that
argument, not the argument.

## What a mechanism claim needs

- **The candidate explanations, named.** Usually two or three are plausible. Say
  which they are before you argue for one; a single explanation asserted looks
  like the only one you thought of.
- **Something that distinguishes them.** An ablation, a stratification, a regime
  where they predict different things. This is often cheap once the main result
  exists — you already have the pipeline.
- **The one that survives, and what would overturn it.** A mechanism nothing
  could refute is a story.

## Where it goes

In the Discussion, and again in one sentence in the Results next to the effect
it explains. A reader who sees the number and no account of it forms their own,
and it is rarely yours.

## When you cannot explain it

Say that, and say what you would run. "We do not know why X and Y agree here;
distinguishing the two accounts would need Z" is a real contribution and reads
as honesty. Silence next to a surprising number reads as not having noticed it.