What this is and is not
Steelmanning builds the strongest version of your opponent’s case. Falsification is a different operation: you specify in advance what would show your position to be false, then you go looking for it in earnest. Steelmanning can be done entirely inside your own head and still leave every one of your beliefs untouched. Falsification requires you to go outside and risk losing something.
Two separate targets, and it is worth keeping them apart:
- The conclusion. Is the thing I am claiming true?
- The argument. Do my premises actually get me there?
You can be right for bad reasons. On stage, a bad argument for a true conclusion loses just as badly as a false conclusion, and it loses worse in front of an informed audience, because the flaw is demonstrable in real time.
Step 1: Decompose the position into falsifiable units
Most debate positions are not one claim. They are a bundle, and the bundle usually contains at least four different kinds of claim that fail in different ways. Write out your position, then sort every load-bearing sentence into one of these:
| Claim type | Example form | What would falsify it |
|---|---|---|
| Empirical | ”X happened,” “the rate is N” | Contrary data, better-sourced counts, primary documents |
| Causal | ”X caused Y,” “X made Y more likely” | Confounders, absent counterfactual, reverse causation, cases with X but no Y |
| Conceptual / definitional | ”X counts as genocide/aggression/censorship” | Counterexample that satisfies your definition but that you would refuse to call by that name |
| Normative | ”X was wrong,” “we ought to Z” | The principle you invoke, applied consistently, generates a conclusion you reject |
| Interpretive | ”the evidence best supports X” | A rival reading that explains the same evidence at least as well, more simply |
If a sentence cannot be sorted, it is probably rhetoric, not a claim. That is fine — but mark it, because rhetoric cannot be defended when attacked, and audiences do not distinguish.
The most common self-deception at this stage
Falsifying a weakened proxy of your own claim. You believe the strong version, you test the weak version, it survives, and you walk away feeling audited. Write out the strongest version you actually intend to assert on stage — including the adverbs — and test that one.
Step 2: Pre-commit to disconfirmers, in writing, before you search
This is the highest-value single move in the whole protocol, and it takes twenty minutes.
For each load-bearing claim, complete this sentence and date it:
If I found __, I would abandon this claim entirely. If I found __, I would narrow it to __. If I found __, I would keep the claim but drop this supporting argument.
Written before you go looking, this prevents the standard failure: you find something inconvenient, and the standard for what counts as disconfirming quietly rises to just above whatever you found. Post-hoc, you will always be able to explain why the thing you found does not really count. Pre-registration makes that move visible to you.
A claim for which you cannot complete these sentences is not a claim you are entitled to assert as factual. Either it is a value commitment — say so openly and defend it as such — or it is vague, and you need to sharpen it before it can be defended at all.
Step 3: Rank by load, and spend effort proportionally
Not every claim deserves the same scrutiny. Score each one on two axes:
- Structural load. If this falls, how much of my case falls with it? Is it a foundation, a supporting wall, or decoration?
- Attack surface. How likely is a competent opponent to go at it, and how cheaply can they do damage?
The highest priority is the quadrant nobody audits: high load, low apparent attack surface. These are the claims you have held so long that you no longer remember where you got them. They are the ones that get destroyed on stage.
The second priority is low load, high attack surface — the decorative flourish that is slightly wrong. It costs you nothing to cut it and a great deal to defend it. A single demonstrable error on a trivial point licenses the audience to discount everything else. Cut aggressively here; you lose nothing structural.
Step 4: Attack from five directions, one at a time
Sequential hostile personas work better than a general instruction to “look for problems,” because different attacks come from different directions and adopting one role at a time stops you from softening all of them at once.
The empiricist. Where does each number come from? Not “which source cited it” — where does it originate? Trace every figure to its root. Check for citation laundering: A cites B, B cites C, and C says something narrower, or is a press release, or is an estimate later revised. Check whether your key study has been corrected, retracted, superseded, or failed replication. Check the denominator: many of the worst errors are category mismatches between numerator and denominator, or an aggregate compared against a subset.
The definitional pedant. Take every term doing moral or legal work — proportionate, civilian, deliberate, systematic, unprecedented — and ask whether you are using it consistently across the whole argument. The characteristic failure is drifting between a technical definition when you need rigor and a colloquial one when you need force. Opponents who know the field will catch this, and it reads as bad faith even when it is only carelessness.
The moral philosopher. State the normative principle you are relying on in its general form, stripped of the case at hand. Then apply it to three other cases, at least one of which you have a prior commitment about in the opposite direction. If the principle condemns your side elsewhere and you would not accept that, either the principle is wrong or you are applying it selectively. Better to find that yourself.
The historian / contextualist. Why does your account begin where it begins? Periodization is an argument in disguise. Whoever chooses the starting point has usually chosen the conclusion. Ask what the account looks like if you begin ten years earlier, and whether you can justify your start date on grounds other than that it favors you.
The bad-faith debater. What is the strongest cheap shot? The out-of-context quote, the guilt by association, the reframe that makes your position sound monstrous in one sentence. You are not obliged to respect this attack, but you are obliged to have a prepared answer, because it will land with the audience whether or not it is fair.
Step 5: Find the differential prediction
This is the sharpest test available and the most underused.
Your position and the best rival position probably both explain the evidence already on the table — that is why the debate exists. So stop asking which explains the past better and ask: where do the two theories predict different things? Identify a case, a document, a dataset, or a future event where your account says one thing and theirs says another. Then go check it.
If you cannot find a single point of divergence, that is diagnostic. It means either the two positions are notational variants of each other, or yours has been formulated so flexibly that it accommodates any evidence. A theory that survives everything explains nothing, and a sufficiently skilled opponent will demonstrate exactly that.
Step 6: Run the symmetry checks
Three tests, all cheap, all uncomfortable.
The label swap. Take your best piece of evidence. Imagine it had come out the other way, or that the identical methodology had been used to support the opposing conclusion. Would you accept it? If you would be reaching for methodological objections in that world, those objections apply now.
The isomorphic standard. Compare the evidence you required to adopt the belief with the evidence you are now demanding to abandon it. If the second is heavier than the first, you are not reasoning, you are defending. This asymmetry is nearly invisible from the inside, which is why it has to be checked explicitly.
The Ideological Turing Test. Write your opponent’s opening statement — full length, best version, no tells. Then judge honestly whether a neutral reader could identify it as written by an opponent. Failure here is not a rhetorical weakness; it means you do not know what you are arguing against, and everything upstream in this protocol was run against a phantom.
Step 7: Convert results into three artifacts
The protocol only pays off if it changes what you walk on stage with.
The falsification register. One row per load-bearing claim: claim as stated → pre-registered disconfirmer → what search was actually run → what was found → verdict (survives / narrowed / dropped) → revised wording. This is your defensive map. When challenged, you will already know exactly how deep the floor goes under each claim.
The concession list. Everything you now know to be wrong, overstated, or unsupported, prepared with framing in advance. Conceding a point you have already planned to concede is a position of strength: it is fast, it is calm, it costs you nothing you needed, and it buys enormous credibility for the claims you then refuse to concede. Conceding a point you had not anticipated is a rout.
Distinguish three moves that look similar and are not:
- Concession. “That’s right, and I was wrong about it.” Use it cleanly, without hedging. Hedged concessions read as evasion and cost more than the original error.
- Scope narrowing. “That’s true of the 2019 data; my claim is about the period after the policy change.” Legitimate — but only if the narrowed claim still does the work your case needs. If it does not, this is a concession wearing a disguise, and the audience can tell.
- Irrelevance. “That’s true and it doesn’t bear on the question, because __.” The most demanding of the three: you must supply the _because immediately, or it sounds like dodging.
The revised claim set. Your actual position after the audit, with scope conditions stated. It will be narrower than what you started with. That is the protocol working, not failing.
Stopping rule
You are done when you can, unprompted and from memory, state the single strongest piece of evidence against your own position — in its strongest form, without editorializing — and your claim still stands with its scope conditions attached.
If you cannot name that evidence, you have not searched. If naming it makes you want to add qualifiers you had not planned, the position is not yet ready and the qualifiers belong in the claim.
Failure modes to watch for in yourself
- The shallow search. Declaring “there is no evidence for that” after twenty minutes on the open web. Absence of evidence found by you is weak evidence of absence, and the claim will be checked by someone who searched properly.
- Vindication drift. One test survived, so the claim is now treated as established. Surviving a test raises confidence a little. It does not convert a contested claim into a settled one.
- Weak-opponent selection. Auditing against the worst public advocate for the other side rather than the best scholar. Comfortable, and worthless.
- Source loyalty. Failing to read hostile specialist reviews of the books and papers you rely on most. Your key sources have critics; those critics are doing your falsification work for free, and your opponent has read them.
- Falsifying the conclusion but not the argument. The conclusion survives, so the reasoning is left unexamined. On stage, the reasoning is what gets attacked, because it is what is visible.