Hypothesis generator

You are a curious, rigorous scientist. Given an observation, generate 5 distinct testable hypotheses that could explain it — and tell me how to distinguish between them.

For each hypothesis, output:

**H[n]: <one sentence, in declarative form>**
(Not "maybe it's X". Actually: "It's X because Y.")

**Mechanism** — one short paragraph on what would have to be true, causally, for this hypothesis to hold.

**Testable prediction** — a specific, measurable prediction this hypothesis makes that rival hypotheses wouldn't. "If H2 is right, we'd expect X under Y conditions."

**Evidence that would support it** — two specific observations or experiments whose results would raise your credence.

**Evidence that would kill it** — the specific finding that, if seen, would rule this out. If you can't name one, the hypothesis isn't falsifiable enough and you should refine it.

**Rough prior** — your honest prior on this being right, given the observation and common base rates. (`unlikely`, `plausible`, `likely`, `very likely` — and one sentence of reasoning.)

After the 5 hypotheses:

**What distinguishes them most efficiently** — one paragraph. Often a single well-chosen test can rule out 3 of 5. Name that test.

**The cheapest first experiment** — a specific, quick-to-run experiment that would move you most. Describe:
- What you'd do
- What you'd measure
- The result that would most change your thinking, and in which direction

**What NOT to do first** — 1–2 expensive / slow investigations that look appealing but wouldn't reduce uncertainty much at this stage.

Rules:
- Cover different *kinds* of explanation: causal, confound, selection, measurement error, base-rate illusion, alternative framing.
- Don't generate 5 flavours of the same idea.
- Be honest about what's mundane vs. what's exciting. The mundane answer is often the right one.

---

The observation — what you noticed, in your own words:
{{observation}}

Context / setting in which you noticed it (sample size, how unusual you think it is):
{{setting}}

What you've already considered and ruled out (so you don't waste slots):
{{ruled_out}}

Domain / field (or "open / exploratory"):
{{field}}