↩ ExploRational · A Catalogue of Divergent Rationality
explorational · a divergent move
You have not heard a view until its holder would certify your statement of it. Rapoport's rules put the repair before the rebuttal: restate, agree, learn — only then criticize. What does criticism earn — and what does it cost — when it is only allowed after the repair?
Rapoport's rules, as Dennett relays them: first, re-express your opponent's position so clearly, vividly and fairly that they say, "Thanks — I wish I'd thought of putting it that way." Second, list every point of agreement, especially if it is not general or widespread. Third, mention anything you learned from them. Only then are you permitted a word of rebuttal. Anatol Rapoport was a game theorist — his four-line tit-for-tat won Axelrod's cooperation tournament — and the rules have a game theorist's shape: a payment schedule, not a courtesy.
The sequence is epistemic, not polite, and the order is load-bearing. The restatement forces you to hold the actual position rather than your memory's paraphrase of it, against an external standard — not "sounds fair to me" but "they would sign it." The agreements fix the ground both parties stand on, giving the disagreement a precise address. The learning step certifies that information flowed toward you; if nothing did, you were reading to reload. And the rebuttal, arriving last, lands on the position as held instead of the position as convenient. Run the same steps in reverse and you have an ordinary argument.
the signature instrument
The bench enforces the order; it does not grade the work. The rebuttal pane opens only after a restatement, a claimed certification, at least one recorded agreement, and at least one recorded learning. Two readouts run live on the restatement — key-term overlap and a framing tripwire — genuinely computed, and labeled what they are: rough lexical proxies for structure, blind to quality. Load the two worked exemplars — same original, two treatments — and watch the readouts pull apart.
Steelman a position
hard sequence lock · readouts are rough proxies, labeled · nothing here is a grade
Waiting for a position and a restatement.
Waiting for a restatement.
The bench records this claim; it cannot verify it — only they can. Editing the certified text voids the signature.
Text changed after certification — signature voided. Re-certify with the holder.
—
—
certification: not claimed
—
earn the right to criticize
Criticism is cheap to produce and expensive to deliver. Anyone can compose an objection; getting it heard by the one person it targets is the rare event, and Rapoport's sequence is the purchase price — paid upfront, in the only currency a holder reliably values: proof of having been understood. The restatement converts your opponent into a proofreader of your understanding; while checking your statement of their view for signature, they are briefly working with you rather than against you.
What the wait earns is uptake. A rebuttal that arrives after the repair cannot be waved off as misunderstanding — the signature is on file. What it costs is just as real: some rebuttals do not survive contact with the strongest form of the position. Expect to feel that as loss. It is the move working; the demolition you were saving had been priced for the weak version.
what to try
Steelman a position you actually oppose — live, with a named holder. Draft the restatement fast, then read the tripwire before fixing anything: the distancing words — claims, apparently, so-called — appear on their own; you never decide to write them. Rewrite until the wire is quiet, and notice the restatement getting stronger, not softer.
Watch the not-reused chips. The key terms of theirs you failed to carry are, suspiciously often, the load-bearing ones — the term you dropped is the matter you were avoiding. Then let the forced learning step run: if nothing surfaces, you have relabeled your memory of the position, not restated it.
Certify, then edit one sentence of the restatement. The bench voids the signature — the one verification it can actually perform. Then load the two exemplars and diff them: same original, two restatements, and the overlap readout and the framing flag pull apart on cue.
the certification is theirs
The sign-test is the only real grade a steelman can receive, and you cannot issue it. A steelman is a claim about the inside of someone else's view; the displayed values are readings taken from outside. The proxies do see something: if the original's key terms are absent from your restatement, you were engaging some other matter, and the overlap says so. If verdicts rode in dressed as description, the tripwire often catches the costume.
But no lexical check can see a paraphrase, weigh an emphasis, or notice that you handed them a reason they do not hold — the signature failure of confident steelmen. The holder catches all three in one read. So the checkbox is framed as what it is: a recorded claim, never a measurement. The bench performs exactly one verification of its own — edit the certified text and the signature is void — because certification attaches to words, not to intentions.
the honest bench
The overlap is a content-token intersection: stopwords dropped, plurals and a few suffixes folded, everything else literal. A fine paraphrase scores low; a lazy quotation scores high; the number is a floor on engagement, not a ceiling on quality. The tripwire is a fixed word-list plus a scare-quote pattern — the full list is disclosed under the readout — and it reads words, not intent: it false-positives on honest quotation and stays silent through a motive-swap.
Neither number is ever averaged into a score, because there is no honest score for a steelman that its holder has not issued. The apparatus is the lock order itself: fifteen words of position, a signed restatement, one agreement, one learning, and only then the rebuttal. The thresholds are chosen, not measured — disclosed so you know which discipline is real.
the move ↔ the bench
| The move | The bench |
|---|---|
| the original pane | the position as held, not as convenient — their words, unedited, the fixed reference everything else is checked against. |
| the restatement | the repair to strongest form: the same matter, re-expressed until its holder would wish they had put it that way. |
| key-term overlap | whether you engaged their actual matter — a token intersection with the original; a floor, never a grade. |
| the framing tripwire | smuggled editorializing — claims, so-called, scare quotes: verdicts dressed as description, caught lexically and roughly. |
| the certification | the sign-test — the only real grade, issued by the only qualified grader; the bench merely records the claim. |
| the locked rebuttal | criticism that waits until it can hit the real thing — and is answerable for it, signature on file. |
how this opening fails
risk · token opposition
A beautiful restatement built to be demolished. The sequence can be performed: agreement entered as throat-clearing ("I agree commutes exist"), a learning that is about the holder rather than the position, certification ticked with no holder in the room. The bench resists this — minimums enforced, wire watching, claims recorded — but cannot prevent it, and says so; the strawman exemplar obeys every lock and still fails. The tell is downstream: a performed steelman leaves your rebuttal exactly as it was before you started. A real one costs you at least a sentence of it.
risk · a holder who does not exist
Repairing to strongest form a position nobody holds flatters everyone and tests nothing: you collect the credit for charity, the opposition is upgraded free of charge, and the sign-test can never run, because there is no one to sign. The certification line is the check — name the holder. If your strongest form cannot be attached to a person who would certify it, you have discovered something about the debate; report that, instead of rebutting a ghost.
Restate until they would sign it; only then have you earned the right to disagree.
can you use it?
RECOGNITION — Which is steelmanning? A: listing the flaws in the opposing view. B: restating the opponent's position so well they'd say 'yes, better than I put it' — before you rebut a word. C: finding the opponent's best-sounding quote.
B. Rapoport's rule: re-express the other side to their own satisfaction first. You have not earned a rebuttal until they would certify your statement of their view. A attacks; C cherry-picks.
THE NEAREST NEIGHBOR — The devil's advocate also engages an opposing case. What separates steelmanning from it?
Direction of effort. Steelmanning builds the opponent's view up to understand it; the devil's advocate argues against the house view to test it. One repairs a rival, the other breaks a consensus.
PRODUCTION — Pick a view you reject. Write the one sentence its strongest holder would most want you to understand — the version they'd sign.
A version (on a policy you oppose): 'The point was never the first-order effect; it was removing a precedent we'd regret handing the other side.' Yours works if a real holder of the view would say 'yes, that's it' — not 'that's a caricature.'