Disconfirmation Adversary

The one idea

Argue the case against, on purpose, before anyone is allowed to argue for.

A project arriving at a gate has already been shaped by the people who like it. Every retelling has smoothed an edge; every objection that was easy to answer has been answered and is now part of the pitch. What has not happened is a serious, resourced attempt to kill it — because nobody is assigned to, and because by the time the idea reaches a gate, being the person who kills it is socially expensive.

So assign it. One pass, at one moment, whose entire job is to build the strongest available case that this project should not proceed. Not devil’s advocacy as a courtesy, not a risks section appended to a proposal — a genuinely adversarial reading, produced before the upside is put on the table, so it cannot be written around the pitch it was meant to test.

The output is a kill-case, not a verdict. It argues; the human decides.

The ordering is load-bearing. An adversary that reads the upside first writes objections the upside already survives. An adversary that goes first has to find the real load-bearing assumption on its own, which is the only way it finds the one nobody has looked at.

When to reach for this

• At a consequential gate — the moment resources get committed: a build starts, a window gets booked, a dependency gets taken on, something becomes expensive to reverse.
• When the decision is hard to unwind. Cheap and reversible does not earn this ceremony; it earns a try.
• Above all: exactly when you feel most certain the project is good. That certainty is the bias this exists to defeat. A gate where everyone already agrees is the gate that most needs an adversary, and is the one where an adversary is least likely to be invited.

When NOT to fire it — and this is half the skill

Never at the intake mouth. The mouth stays permissive on purpose. An adversary at intake teaches people not to write ideas down, and the cost of capturing a bad idea is a line in a list — approximately nothing. The cost of not capturing a good one is unbounded and invisible. Filtering belongs at the gate where resources are committed, not at the door where ideas arrive.

If you find yourself wanting to run this on an idea nobody has committed anything to yet, you are not doing quality control; you are doing gatekeeping at the wrong seam.

The anatomy — the moves that make it work

1. Kill-case first, upside second. The adversary reads the proposal and writes its case before any benefit is scored, ranked, or argued. Sequence it in the artifact too: the kill-case document is written and timestamped before the upside section exists. Why: an objection written next to its own rebuttal is a rebuttal wearing an objection’s clothes.

2. Attack ONE assumption — the riskiest — not a list. A ten-item risk register is a way of saying nothing at length; every item gets a sentence, nothing gets thought about, and the reader skims. Name the single assumption on which the whole project rests and go at that. Why: projects do not usually fail from ten small risks. They fail because one thing everyone took as given was not.

3. Name the cheapest test that would settle it. The kill-case is not finished when the objection is stated. It is finished when it says what would resolve the objection and what that costs — an afternoon, one measurement, one phone call. Why: this converts an argument into a decision procedure. An objection with no test attached can only be won by whoever talks longest.

4. Argue from a skeptic’s vantage, not the owner’s. Write as someone with no stake in the project’s success and a professional interest in its failure modes — the reviewer who has seen this shape fail before. Why: the owner’s vantage is already fully represented in the room. Adding a second copy of it produces objections the owner has already answered.

5. Produce a KILL-CASE, never a verdict. The output does not say “do not proceed.” It says: here is the case for not proceeding, here is the assumption it turns on, here is what would settle it. Why: an adversary that renders verdicts becomes an approver, and then it accumulates the authority that made the gate necessary in the first place. Its power must stay rhetorical.

6. The rebuttal is written, and only a GO requires it. If the human decides to proceed, they answer the kill-case in writing, and that rebuttal ships with the decision. A park or a kill resolves without one. Why: the asymmetry is the entire control. Requiring a written answer to proceed makes proceeding the expensive option — which is correct at a resource-committing gate — while never taxing the two outcomes that spend nothing. Symmetric ceremony would just make the adversary an obstacle to route around.

7. The rebuttal is a record, not a formality. It is the artifact you read later when the project is in trouble, and it is the only place the assumption that failed was named in advance. A one-line “noted, proceeding” is not a rebuttal; it is the ceremony without the control.

The line it must never cross

It argues the case; the human decides.

The adversary has no veto, no score, no threshold, no automatic outcome. It cannot park a project, cannot demand evidence, cannot block on its own objection being answered to its satisfaction. The instant it can do any of those, the decision has quietly moved from the person accountable for it to a process nobody is accountable for — and the failure mode is not that bad projects get through, it is that the good project everyone was excited about dies to a technicality nobody owns.

Concretely: the adversary writes, the human answers, the human decides, and the decision is the human’s alone.

Template (skeleton)

KILL-CASE — <project> · gate: <what is being committed> · <date>
  (written BEFORE any upside is scored)

THE RISKIEST ASSUMPTION
  <the one thing this project rests on that has not been tested>
  Everyone has been treating this as given because <why it went unexamined>.

THE CASE AGAINST
  <the strongest argument that this project should not proceed, written from
   a skeptic's vantage. One argument, developed — not a risk register.>

WHAT IT WOULD COST TO BE WRONG
  <the failure this produces, and whether it is recoverable>

THE CHEAPEST TEST THAT SETTLES IT
  <the smallest observation that would confirm or refute the assumption>
  Cost: <hours / one measurement / one conversation>

WHAT WOULD CHANGE MY MIND
  <state it explicitly — an adversary that cannot be answered was not arguing>

─── the human answers below; a GO requires it, a PARK or KILL does not ───

DECISION: GO | PARK | KILL          decided by: <the human> · <date>
REBUTTAL (required on GO only)
  <the written answer to the case above — this is the record read later>

Anti-patterns — the adversary that does nothing

• Firing at intake → people stop writing ideas down, and you have moved the filter to the one place where filtering has no upside.
• Reading the upside first → the objections are shaped to what the pitch already survives.
• A ten-item risk register → breadth is how an adversary avoids saying the one thing that would hurt. One assumption, seriously attacked.
• An objection with no test attached → unfalsifiable, so it resolves by seniority rather than by evidence.
• Rendering a verdict → the adversary becomes an approver, and the gate now has two decision-makers and one accountable party.
• Requiring a rebuttal to park or kill → symmetric ceremony makes the adversary an obstacle; the asymmetry is what makes it a control.
• “Noted, proceeding” → the ceremony survived and the control did not.
• Skipping it because everyone agrees → that is the exact condition it was built for.

Provenance

Earned at a gate where a project everyone liked was funded before anyone had been asked to argue against it, and the assumption that later sank it had never once been written down. The asymmetric rebuttal rule — required on GO, waived on park and kill — came later, after a symmetric version was observed being routed around: people stopped bringing decisions to the gate at all.