What the discipline inversion means if you actually trade
If you trade prediction markets and you grew up on Manifold, your sizing instinct is wired for the wrong venue. Not in the “calibration still takes practice” way. In the measurable, data-visible, statistically-significant-at-p-equals-0.0001 way.
The Convexly Research paper that went up yesterday at /research/edge-score-methodology-v1m is the long version. This post is the short version for anyone who is about to size a real-money position and wants the operator summary first.
One number
1,647 Manifold users cleared the twenty-resolved-bet floor in both the pre-sweepcash bridge window and the six months of sweepcash (Manifold's real-money program). Same users, same platform, same UI, same question catalog. The only thing that changed was whether a subset of contracts paid out in dollars.
Median within-user concentration on those 1,647 traders fell by 8.9 percentage points when real money showed up.
Not “some users got more careful.” The median user got 8.9 percentage points less concentrated. When sweepcash ended, concentration rose back up by 3.5 points on the 1,839 users active across that transition.
Calibration did not move. The users were just as accurate when their mana was redeemable as when it was not. What moved was how much capital they put behind any one forecast.
Why this is the opposite of the naive read
The naive read of “real money changes behavior” is “real money makes people take bigger swings, because the payoff is real.” The data say the exact opposite. When real money arrives, the same trader takes less concentrated swings. When it leaves, they concentrate again.
The mechanism is not a mystery. Play money on Manifold has no ruin barrier. If you bet your entire 1,000 mana stack on a single thesis and you are wrong, you get 250 mana at the start of next month. Real money on the same platform introduces a barrier that play-money trading never had. The absorbing barrier does not reward maximum conviction. It rewards surviving long enough to compound.
This is the direction Taleb has been arguing from first principles for twenty years. The Convexly research is the empirical test, done on the same users across an incentive regime shift that Manifold accidentally ran for us.
The reading for operators
Three practical implications for anyone trading real-money prediction markets (Polymarket, Kalshi, Prophet, whatever comes next):
Your play-money sizing instinct is wrong by design. If you cut your teeth on Manifold, your default position size is calibrated to a venue that has no absorbing barrier. On a venue that does have one, the same default is systematically too concentrated. The within-user 8.9 point shift is the visible correction those 1,647 users made when the barrier showed up. You will have to make the same correction or the market will make it for you.
Copy-trading the Polymarket leaderboard without accounting for venue is a misread. The frozen V3b coefficients from the 8,656-wallet Polymarket cohort put conviction at +2.72 and discipline at -1.15. The refit on the 15,106-user Manifold cohort puts the same pillars at +0.16 and +0.36. Same features, same z-score procedure, same OLS fit, different venue, an order-of-magnitude difference on conviction and a sign flip on discipline. If your copy-trade strategy is “hit rate” and you are applying it across venues, you are applying three different measurements under one label.
Hill alpha below one is not a typo. The per-user PnL tail on the full 15,106-user Manifold cohort is alpha = 0.86, 95% CI (0.80, 0.93). Below the mathematical boundary where the mean is even defined. On Polymarket the same estimator returns 1.28 (1.20, 1.36). If you are sizing Kelly off “average historical edge” on a play-money venue, you are sizing off a statistic that has no well-defined value. The honest read is to use a venue-specific fractional-Kelly multiplier, anchored to each venue's measured tail, not to import a finite-variance equity default.
What this is for, at Convexly
Three things this research directly changed in the product:
Convexly's Polymarket wallet analyzer uses V3b coefficients fitted on the Polymarket cohort. Any wallet score you see there is venue-specific and explicitly so. When the Kalshi integration lands (on V3b-M, the multi-outcome extension introduced in this paper), the scores will use a Kalshi-specific refit. There will not be a single set of coefficients rescored across venues. That was a decision made on the basis of this data.
Convexly's Kelly Replay calibrates fractional-Kelly sizing against the measured Hill alpha of the venue you are actually trading. The default is not half-Kelly. The default is what the tail on your venue says the fraction should be. For Polymarket that is roughly quarter-Kelly; for Manifold the honest answer is that naive Kelly is not well-defined below alpha = 1, and the tool says so.
Convexly's Trade Journal compares your sizing to the cohort of traders on the venue you logged the position on. If you are entering a Polymarket position using a Manifold-calibrated gut, the journal will flag it as oversized relative to the Polymarket cohort's distribution. That is the pre-mortem the 1,647 sweepcash traders' data argues for in general form.
The honest caveat
These findings are structural, not personal. The data cannot tell you that any specific Manifold user is careless or that any specific Polymarket user is careful. What the data can tell you is that a cohort of 1,647 traders, the same individuals measured before and after the incentive regime changed, adjusted their sizing in a particular direction. Your job is to decide whether you are the exception or not. The cohort-level prior is that you are not.
The full paper lives at /research/edge-score-methodology-v1m. The underlying data bundle (15,106 hashed users, sweepcash deltas, coefficient tables, reproduction script) is at /research/v1m/v1m-data-bundle.tar.gz. Paste any Polymarket wallet into the analyzer to see the three pillars split out for a specific trader. Pro tier is $49 a month.
Comments and corrections to research@convexly.app.