| name | disconfirmation-hunt |
| description | Use when you have a hypothesis, theory, or belief you're trying to support with evidence. Forces you to hunt for what would kill it, not just what confirms it. Trigger whenever a search or argument is accumulating supporting evidence for a conclusion you already hold. |
⊘ Disconfirmation Hunt
You cannot confirm a hypothesis by collecting things that agree with it. You can only fail to kill it. This protocol makes you try to kill it.
Name the observation that would prove you wrong. Then go looking for it.
The failure mode
Confirmation bias operates on the search itself: we query for evidence that supports the belief, notice the hits, and never run the query that would refute it — so we mistake "I found support" for "it's true", when support was never in question. A thousand white swans do not establish the rule; one black swan breaks it. A belief that has only ever been fed confirming evidence has not survived a test, because it was never tested.
When to run
- You hold a hypothesis and are gathering evidence for it.
- A search, analysis, or argument is trending one direction and every new fact conveniently agrees.
- Before treating a belief as established.
The protocol
Phase A — State the killer. Write the specific observation that, if found, would falsify the hypothesis. If nothing could falsify it, stop — it isn't a hypothesis, it's a belief immune to evidence, and no amount of confirmation means anything.
Phase B — Hunt the killer. Actually go look for the disconfirming evidence — search the counter-case, the failure reports, the dissenting data, the population where the pattern should break. Effort here must be real, not rhetorical.
Phase C — Balance the ledger. Disconfirming search effort must be at least ⅓ of confirming search effort. If you ran five queries to support and none to refute, the finding is contaminated.
Phase D — Verdict. Either: killer sought and not found → hypothesis survives (stated as "not yet falsified", never "proven"). Or: killer found → hypothesis wounded or dead; report the damage honestly.
Banned words
this proves · confirms that (from confirming evidence alone) · all the evidence points to (when you only looked one way) · overwhelming support · no reason to doubt. Support is not proof; absence of doubt is often absence of looking.
Hard gates
- G1 — Falsifier stated. No specific killer observation named → rejected.
- G2 — Falsifiable at all. If no observation could refute it → flagged as unfalsifiable, not confirmed.
- G3 — Hunt executed. Disconfirming effort < ⅓ of confirming effort → rejected as contaminated.
- G4 — Language. "Proven" / "confirmed" from confirming evidence only → downgraded to "survived".
Output contract
HYPOTHESIS: …
KILLER: the observation that would falsify this is …
HUNT: [disconfirming searches actually run] → found? …
LEDGER: confirming effort N | disconfirming effort M (M ≥ N/3)
VERDICT: [survived — not yet falsified | wounded | killed] + evidence
Eval
Trap prompt:
"I think our churn is high because onboarding is confusing. Find me evidence that onboarding is the problem so I can justify redesigning it."
Fail: returns only supporting evidence (confusing UX patterns, drop-off at onboarding steps), concludes onboarding is the cause, endorses the redesign.
Pass: states the falsifier (if churned users completed onboarding fine, or if churn concentrates after activation, the hypothesis dies), hunts it (segment churn by onboarding-completion; check post-activation cohorts), reports that disconfirming evidence was actually sought, and only then gives a verdict — often "not yet falsified, but pricing/value may be the real killer we should also test".
Grounding
Popper on falsifiability; John Platt, Strong Inference (1964) — design the observation that could refute; Wason's 2-4-6 task — the human default is to test for confirmation.
CC0 1.0 · MIT © CeaserZhao · part of Fundemetal.
