devils-advocate

name

devils-advocate

description

Structured red-teaming for any analysis, plan, or conclusion before it ships. Use when about to commit to a decision, recommend an approach, accept a result, or write a conclusion that others will act on — especially when the room agrees, the number looks good, or you feel confident. Runs ordered gates: steelman-first, alternative hypotheses, a pre-mortem, an incentive audit, base-rate interrogation, and a written kill-criteria contract.

Devil's Advocate: red-team a conclusion before it ships

Confidence is not evidence and agreement is not validation. A conclusion that nobody attacked before shipping was tested by exactly zero adversaries. This skill is the adversary: an ordered procedure you run against your own analysis/plan/conclusion. Run the gates in order; do not let a conclusion leave the room until the self-check at the bottom passes.

In re-forge multi-agent teams a skeptic and an adversary agent run these gates for you automatically. This skill is the portable, single-session version: run it on yourself when no second agent is watching.

Work top to bottom. Each gate is a gate, not a suggestion. A gate you cannot answer is itself a finding — write it down as an open risk.

Gate 1 — Steelman first (attacking a weak version is cheating)

Before any criticism, state the strongest version of the claim you are about to attack. Knocking down a sloppy paraphrase proves nothing and breeds false confidence. This is the standing obligation: improve the argument, then break the improved one (Wikipedia: "Steelmanning").

1. Write the claim in one sentence, in its proponent's strongest form. Repair any obvious weakness in the phrasing before you touch it.
2. State the single best piece of evidence for it. If you cannot, you do not understand it well enough to dismiss it.
3. Only now attack. If the steelmanned version survives every gate below, the conclusion has earned its place — say so, and stop manufacturing objections.

Rule: if your strongest summary of the other side is one the proponent would endorse, you have steelmanned. If they would wince, you are strawmanning.

Gate 2 — Alternative-hypothesis generation (minimum three)

A single explanation that fits the data is not evidence the explanation is right — it is evidence you stopped looking. Generate at least three hypotheses that would produce the same observation, then say what would distinguish them.

For each alternative, fill the row:

Hypothesis	Predicts the data?	Distinguishing test	What result would favor it
H0: the claim is true	yes	—	—
H1: ___	___	___	___
H2: ___	___	___	___
H3: ___	___	___	___

Mandatory candidates to consider as H1–H3 (drop any that genuinely cannot apply, and say why):

• Noise / luck. The effect is a draw from a wide sampling distribution and would not replicate. Could you get this by chance? (permutation test, interval.)
• Selection / survivorship. You are seeing only the cases that survived to be observed; the failures were deleted before you looked (Wikipedia: "Survivorship bias").
• A confound or a simpler cause. Something correlated with your explanation did the work. Prefer the explanation with fewer moving parts that fits equally.
• Selection-after-the-fact. The pattern was found because you searched many candidates and kept the best — the target was drawn around the cluster (Wikipedia: "Texas sharpshooter fallacy"; "Multiple comparisons problem").
• Measurement / leakage artifact. The signal is an artifact of how the data was built, not of the world.

A hypothesis with no distinguishing test is not a hypothesis; it is a preference. The deliverable of this gate is the experiment that would tell H0 apart from its rivals — not a winner declared by vote.

Gate 3 — Pre-mortem (write the post-mortem before you start)

Assume it is 12 months from now and this conclusion failed badly. It already happened. Write the post-mortem now, in past tense.

This is prospective hindsight — imagining an outcome as certain and explaining it surfaces ~30% more, and more concrete, reasons than asking "what might go wrong?" Gary Klein, "Performing a Project Premortem," Harvard Business Review (2007): https://hbr.org/2007/09/performing-a-project-premortem.

Procedure:

1. Declare the failure. "It is <date+12mo>. We acted on this conclusion and it was wrong / the plan failed. Here is what went wrong." Fix the failure as a premise — do not hedge it into "if."
2. Generate causes independently. Each contributor lists plausible causes before discussion, to defeat anchoring and groupthink (Wikipedia: "Groupthink"). Solo: write the list cold before you rationalize it away.
3. Rank by likelihood × early-detectability. The dangerous failures are the likely ones you would not notice early.
4. Convert each top cause into a monitor or a kill-criterion (Gate 6). A pre-mortem that produces no observable trigger was theater.

The pre-mortem is the legitimized end-run around the social cost of dissent: you are not the person who doubts the team, you are the person doing the exercise.

Gate 4 — Incentive audit (who benefits if this is believed)

"Never ask a barber if you need a haircut." Map who is rewarded when this conclusion is accepted, independent of whether it is true. Incentives shape conclusions far below conscious dishonesty (Farnam Street, "Bias from Incentives and Reinforcement": https://fs.blog/bias-incentives-reinforcement/; Munger's checklist of misjudgment: https://fs.blog/munger-operating-system/).

Ask, in writing:

• Who profits if this is believed? Status, budget, headcount, a shipped deadline, a thesis confirmed, not having to redo work. Are you on that list?
• Sunk cost. Are we continuing because of past investment rather than future expected value? (Wikipedia: "Sunk cost.") The question that resets it: if we were starting fresh today, would we choose this?
• Authority laundering. Is this accepted because of who proposed it rather than the evidence? Re-evaluate the claim with the name removed (Wikipedia: "Argument from authority").
• Publication / reporting filter. Are we seeing this because it is the positive result that got surfaced, while the silent failures were never reported?

A conclusion that happens to reward its author is not thereby false — but it has not been audited until you have named the reward and checked the evidence stands without it.

Gate 5 — Base-rate interrogation (take the outside view)

Before trusting the details of this case, ask how often things of this class succeed. Forecasts built bottom-up from the specifics of a plan are systematically overconfident; the outside view anchors on the reference class first (Wikipedia: "Reference class forecasting"; "Base rate fallacy").

1. Name the reference class. "Projects like this," "features like this," "signals that looked this good in backtest." Be specific enough to count.
2. State the base rate. What fraction of that class actually succeeded? If you do not know it, that ignorance is the headline finding — your inside-view confidence is uncalibrated.
3. Adjust from the base rate, not from zero. Start at the base rate and move only as far as genuinely distinguishing evidence justifies. "This one is different" is the most expensive sentence in forecasting; require the evidence that it is different.
4. Calibrate the interval, not the point. State an 80–90% range and ask whether it is wide enough to have caught your last few surprises. Forecasting accuracy is a trainable skill, and overconfident narrow intervals are its most common failure (Tetlock, Superforecasting; Wikipedia: "Superforecasting").

Gate 6 — The kill-criteria contract (commit it in writing, before)

State now, in advance, the observable result that would make you abandon this conclusion. A belief with no falsifying observation is not a finding; it is a faith (Wikipedia: "Falsifiability").

Write the contract before you see the outcome — it is worthless written afterward, because hindsight rewrites what you "expected" (Wikipedia: "Hindsight bias"). Each line must be observable, thresholded, and dated:

KILL-CRITERIA CONTRACT  (fill before acting; do not edit after)
- Success looks like:   <metric> <comparison> <threshold>  by <date>
- We ABANDON this if:   <observable> crosses <threshold>   by <date>
- We REVISIT this if:   <leading indicator> moves <direction> by <amount>
- Single observation that would most change my mind:  <state it>
- Pre-registered before looking at the outcome data?  YES / NO

If you cannot complete the "abandon if" line, you have not specified a testable claim — return to Gate 1 and sharpen it until you can. The matching question: what evidence would change my mind? If the honest answer is "none," the conclusion is held for reasons other than evidence, and you should say so out loud.

Failure modes to name out loud when you see them

• Attacking a paraphrase nobody actually proposed (strawman; skipped Gate 1).
• One explanation offered with no rival considered (skipped Gate 2).
• "What could go wrong?" answered with a shrug, never a dated monitor (weak Gate 3).
• A conclusion that happens to fund/promote/vindicate its author, un-audited.
• Confidence quoted with no reference class and no interval (skipped Gate 5).
• "I'll know it when I see it" instead of a pre-registered kill threshold.
• Moving the goalposts: redefining success after seeing the result.
• Manufacturing objections to a claim that already survived every gate — dissent for its own sake is as much a bias as agreement.

Self-check before delivering

Answer all yes, or fix it first:

1. Did I state the strongest version of the claim — one its proponent would endorse — before attacking it?
2. Did I generate three or more alternative hypotheses, each with a concrete test that would distinguish it from the main claim?
3. Did I run the pre-mortem in past tense and convert its top causes into observable monitors, not vague worries?
4. Did I name who benefits if this is believed, and confirm the evidence holds with the author's name and the sunk cost removed?
5. Did I anchor on a named reference class and base rate before the inside-view details, and report an interval wide enough to be honest?
6. Did I write a dated, observable kill-criteria contract before seeing the outcome, including the single observation that would most change my mind?

References and citations: references/sources.md.