---
description: "Test whether the evidence behind a claim actually supports it, or just fails to contradict it."
model: sonnet
---

# Falsify

Take a claim and apply Popperian falsification: is the claim falsifiable, what evidence would
disprove it, and does the existing evidence actually support it — or does it merely fail to
contradict it? The distinction matters. A claim that cannot be proven wrong is not a strong
claim; it is an unfalsifiable one.

Use this when a decision is being justified by evidence that has not been scrutinized, or when
something "seems supported" but the reasoning has never been stress-tested.

## Behavior Constraints

<output_contract>

Present the falsification analysis as:

- **Claim** — the claim restated precisely and completely
- **Falsifiability** — is the claim falsifiable? If yes, state what a disconfirming observation
  would look like. If no, explain why and flag it as unfalsifiable.
- **Disconfirming evidence** — what specific evidence, if observed, would weaken or disprove
  the claim. Be concrete: not "evidence against X" but "if Y were measured and found to be Z."
- **Evidence audit** — assess the evidence currently cited in support of the claim:
  - Does it actually support the claim, or does it only fail to contradict it?
  - Is it subject to selection bias, confounding, or base rate neglect?
  - What is the quality and source of the evidence?
- **Verdict** — one of: well-supported, weakly supported, unsupported, or unfalsifiable —
  with a one-sentence justification

</output_contract>

<completeness_contract>

The task is complete when the evidence audit goes beyond surface-level review — it must
distinguish between "this evidence supports the claim" and "no evidence contradicts the claim."
If the claim is genuinely well-supported and the evidence holds up to scrutiny, say so.

</completeness_contract>

<reasoning_rules>

- Absence of evidence is not evidence of absence, but it is also not confirmation. Flag when
  a claim is being treated as supported simply because it has not been challenged.
- Do not conflate this with think-devils-advocate. Devils-advocate challenges the position.
  Falsify challenges the evidence supporting it.
- Unfalsifiable claims are not necessarily wrong — but they are not scientific claims and
  should not be treated as empirical justification for decisions.
- Be precise about what the evidence actually shows. "Studies suggest X" is not the same as
  "X has been demonstrated under conditions Y and Z."

</reasoning_rules>

## Workflow

1. Ask the user for the claim and any supporting evidence if not already provided. Optionally ask:
   - What decision this claim is being used to justify
   - Where the evidence comes from
2. Assess falsifiability.
3. Audit the evidence and present the output per the output contract above.