---
name: falsifiable-hypothesis
description: "Turn a hunch you are about to test into a falsifiable hypothesis before you collect data — state the null (no-effect) version and the alternative, define how each variable is actually measured, commit to a predicted direction, and name the specific result that would prove the hypothesis wrong — instead of shipping a vague directional guess like \"I think X drives Y\". Use when someone has a hunch, prediction, or \"I bet that…\" and wants to set it up as a test, experiment, or study. Do NOT use to disconfirm a conclusion already reached (that is self-refutation-pass), to critique a finished study's methods (that is methodology-critique), or to compute how many subjects a test needs (that is power-and-sample-size)."
metadata:
  decimalai-skill-id: "e163bffe-b056-44eb-ad28-d4811c67106f"
  decimalai-version-id: "e6a1527d-d220-4bbe-aa10-6803da5195b0"
  decimalai-version: "1"
  decimalai-body-sha256: "f07efbef2d92833cc967adeb654557278170e0b8cd6a678d312349d5fc0becfd"
  decimalai-name: "falsifiable-hypothesis"
  decimalai-format: "1"
  decimalai-invocation: "model"
  decimalai-category: "research"
  decimalai-source: "https://app.decimal.ai/s/falsifiable-hypothesis@1/SKILL.md"
---
# State the Hypothesis So It Could Be Proven Wrong

The default response to "I think X drives Y" is to restate it as a confident directional sentence — "the price increase raises churn" — and move on to testing it. That guess is not yet a hypothesis you can test, because it never said what result would count as *wrong*. A hypothesis earns the name only when a specific observation could disconfirm it. The discipline here is to pin a hunch into a form that is exposed to being wrong **before** any data is collected: an explicit no-effect baseline, variables defined by how they are measured, a committed direction, and the single result that would sink it.

## When to Activate

Activate when there is a hunch, prediction, or belief that is *about to be tested* and has not yet been pinned down: "I think X causes Y", "my bet is that…", "I want to check whether…", "let's test that…", "set this up as an experiment." The target is a claim heading into data collection.

Do NOT activate for a conclusion that has already been reached and is being defended or shipped — running one disconfirmation pass on a committed claim is self-refutation-pass. Do not use it to critique someone else's finished study (methodology-critique), to compute required sample size or power (power-and-sample-size), or to analyze results that have already come in (that is significance testing on collected data). This is the setup step, done once, before the test runs.

## The Form

Produce all five. Skip one only by saying why it does not apply.

1. **Null and alternative, both explicit.** State the null (H0) as the boring no-effect / no-difference / no-relationship baseline, and the alternative (H1) as the effect you actually expect. The null is not a formality — it is the thing the test is set up to reject. "H0: churn is the same in the raised-price and control cohorts. H1: churn is higher in the raised-price cohort."
2. **Operationalize every variable.** For each variable in the claim, say *how it is measured* — the metric, unit, instrument, window, and threshold — not the abstract concept. "Churn" becomes "percent of active paid accounts that cancel within 30 days of a renewal, from billing records." "Price increase" becomes "list price raised by ≥10%." A variable you cannot state a measurement for cannot be tested.
3. **Commit to a predicted direction (and magnitude if you can).** Say which way the effect should go, and how big would matter — one-sided when you have a real prior, with a smallest effect worth caring about if you can name one. "Churn at least 3 percentage points higher in the raised cohort." A hypothesis that would be "confirmed" by an effect in either direction predicts nothing.
4. **Name the disconfirming observation — the specific result that proves it wrong.** State the concrete finding that would falsify H1 and leave H0 standing. "Churn equal or lower in the raised cohort than in control over 60 days falsifies it." If you cannot name any result that would prove the hypothesis wrong, that is the finding: the claim is unfalsifiable and cannot be tested as written — fix it or flag it.
5. **State the scope and the comparison.** Name the population the claim covers and what it is compared *against* — the control, baseline, or counterfactual. "Among paid accounts on monthly plans, versus a matched cohort whose price did not change." A claim with no comparison group has nothing to be measured against.

## Failure Modes This Guards Against

- **Directional guess with no null.** "X raises Y" with no stated no-effect baseline is a prediction, not a testable hypothesis — there is nothing set up to be rejected.
- **Unmeasured variables.** "Engagement improves" with no definition of engagement can be declared true or false after the fact to taste; operationalize it up front.
- **Two-tailed dressed as a prediction.** "There will be a difference" is confirmed by any result and predicts nothing — commit to a direction.
- **Unfalsifiable claim.** If no possible result would count as disproof, the hypothesis explains nothing; flag it rather than testing it.
- **HARKing setup.** Leaving the prediction and the disconfirmer vague so the hypothesis can be quietly reshaped to match whatever the data shows. Fixing the disconfirmer *before* the data is the guard.
- **No comparison.** Measuring the treated group alone, with no control or baseline, leaves the effect uninterpretable.

## Discipline (what the base skips by default)

1. **Write the explicit null** — the no-effect baseline the test is built to reject, not just the effect you hope for.
2. **Operationalize every variable** — metric, unit, window, threshold; no untestable abstractions.
3. **Commit to a direction before the data** — one prediction, fixed in advance, not "some difference."
4. **Name the result that would prove it wrong** — a concrete disconfirming observation, decided up front.
5. **State the comparison and the scope** — against what, and for whom, or the effect is uninterpretable.