Hypothesis Mining

You have done the research and now you are drowning in it: transcripts, support tickets, keyword volumes, a stack of papers, competitor one-star reviews. The temptation is to test the assumption that is easiest to check or most pleasant to confirm. That is how founders spend three months proving the technology runs while the real killer, that nobody will pay, goes untested. Mine the input into a ranked stack of falsifiable bets and test the one that is most uncertain and most fatal first.

The method

Mine, classify by risk, score by uncertainty times fatality. Full framework: Handbook Chapter 02. Deep method, scoring grid, and a worked example: references/hypothesis-mining-method.md.

Step 1: Mine every source for raw beliefs

Sweep all the input and pull out the implied beliefs. Different sources surface different risks, so mine widely before you rank.

Step 2: Write each as a falsifiable hypothesis

State every belief as: "We believe [X]; we'll know we're wrong if [bar]." Set the pass/fail number before any test. If no result could fail it, it is not a hypothesis, it is a wish. Forbid solution-shaped claims ("users want our dashboard"); restate as a problem belief.

Step 3: Classify the risk

Sort each hypothesis into one bucket, because you test them differently.

Step 4: Score uncertainty times fatality, pick the top

Score each on how uncertain you are (1-5) and how fatal it is if wrong (1-5). Multiply. The highest product is your first test, not the easiest or the most flattering one.

Output

• A hypothesis backlog, each written "We believe X; we'll know we're wrong if [bar]", risk-typed.
• An uncertainty-times-fatality score per hypothesis and a ranked order to test.
• The single riskiest assumption ([RISKIEST_ASSUMPTION]) and its first, cheapest test named.
• Next skill: frame-the-hypothesis to sharpen that top bet and run the devil's advocate on it.

Constraints

• Solution-shaped hypotheses are the failure mode: a belief about your feature, not the customer's world, is untestable theatre. Restate every one as a problem or demand belief before scoring.
• Every hypothesis carries a pass/fail bar set before the test, or it is unfalsifiable. No bar, no bet.
• Do not let the model rank by what is easy to confirm (sycophancy); rank by fatality if wrong.
• Stay theme-agnostic; the founder supplies the domain, you supply the rigour.

Dictionary

sycophancy · the Remove-the-AI test

Copy-paste version

For non-coders: paste into any chatbot (Claude.ai, ChatGPT, Gemini). Tag: chat.

Act as a sharp product strategist who turns research into testable bets. My idea: [ONE_LINER]. For:
[ICP]. Domain: [DOMAIN]. Here is my raw research: [PASTE NOTES, SUPPORT LOGS, REVIEWS, DATA].
1. Mine every source for the beliefs my idea quietly rests on. List them all, including the ones I'd
   rather not examine.
2. Rewrite each as a falsifiable hypothesis: "We believe X; we'll know we're wrong if [bar]." Set the
   bar as a number BEFORE any test. Reject any claim about my feature; restate it as a belief about
   the customer's problem or demand.
3. Tag each with its risk type: problem, demand, solution, or feasibility.
4. Score each on uncertainty (1-5) and fatality-if-wrong (1-5), multiply, and rank by the product.
CONSTRAINTS: do not rank by what is easy to confirm; rank by what kills the company if wrong. No
solution-shaped or unfalsifiable claims. Do not invent market numbers; label any figure an assumption.
OUTPUT: the ranked hypothesis backlog as a table. End by naming the single riskiest assumption and
the cheapest experiment that would expose it.