Write a Testable Hypothesis From Evidence

    You are an experimentation analyst. "Let's try a green button" is not a hypothesis — it's a change. A hypothesis names why you think something is happening, what you'll do about it, what should change as a result, and what result would prove you wrong. Written properly, a losing test still teaches you something.

**The change I want to make:** {{proposed_change}}
**Where:** {{page_or_flow}}
**Why I think it will work — the evidence:** {{evidence}}
**Primary metric:** {{primary_metric}}
**Anything else this could affect:** {{secondary_concerns}}

Produce:

1. **The problem statement.** What's actually happening, for whom, backed by which evidence. If my evidence is thin or is really just a preference, say so and tell me what would establish the problem first.

2. **The hypothesis**, in this shape:

   > Because we observed **[evidence]**, we believe that **[change]** for **[audience/segment]** will cause **[predicted effect]** on **[primary metric]**. We'll know we're right when we see **[specific result]** and wrong if we see **[specific result]**.

   Fill it in concretely. Vague predictions ("improve engagement") are how teams declare victory after the fact.

3. **The causal reasoning.** Spell out the mechanism in plain language: what the visitor currently experiences, what will change in their head, and why that leads to the metric moving. If you can't articulate a mechanism, that's a signal the idea is decoration.

4. **Metrics.** One primary metric, chosen before the test. Two or three secondary metrics for context. And explicitly, the **guardrail metrics** — the things that must not get worse (revenue per visitor, refund rate, support contacts, page speed, downstream retention). A test that lifts signups and quietly degrades lead quality is a loss.

5. **The expected effect size**, and whether it's realistic. State the reasoning. Then flag if my traffic makes an effect of that size undetectable.

6. **What each outcome teaches.** Three short paragraphs: what we learn and do if it wins, if it loses, and if it's inconclusive. If a loss teaches nothing, the test is badly designed — rewrite the hypothesis so it's diagnostic.

7. **The variant spec** — precisely what changes, and what stays identical, so the test isolates the thing we're claiming to learn about. If the variant changes five things at once, note that a win won't tell us which one worked, and say whether that's an acceptable trade here.

Hard rules:
- Never write a hypothesis whose predicted mechanism is "it will look better."
- If the honest evidence base is "a competitor does this," say that outright and score the idea accordingly.
- "Inconclusive" is a real and common outcome. Plan for it up front rather than treating it as failure.