Where the late price comes from

The mechanism kills late pool entry on purpose. A stake arriving after the last opposing dollar is paid exactly 1×, so a rational trader with any belief short of certainty will not make it, and the measured unconditional EV of the buzzer strategy is −1.05% here against +25.27% under classic rules.

That is the design working. It also creates a hole, and the hole is the most serious open problem in the paper.

If informed money stops entering the pool late, the pool's ratio stops updating late. And the last hour of a prediction market is when its forecast matters most.

The hole, measured

Brier score of the pool ratio against the realized outcome, by phase of market life, agents responding to whichever rule they face (lower is better):

And the harsher metric, mean absolute gap to the generator's true probability:

Three readings the paper insists on, all of them uncomfortable:

The volume migration buys no measurable early forecast gain. First-third Brier is identical across arms — 0.239 everywhere — and |q − p| is weakly worse under the mechanism in every phase. Within this crowd model, paying for time-priority does not improve the price. It protects the payout. The paper's own conclusion is that "the instrument case for the mechanism is dilution protection, the deleted lock, and the seeding economics, not forecast quality, and this paper stops implying otherwise."

The fair headline is against the deployed baseline, not the ideal. 0.174 against the lock arm's 0.150. The unlocked classic's 0.148 is a bound, not a practice — no venue runs it, because +25% buzzer EV is not survivable.

It is not a sampling artifact. A fixed-clock control — one end-of-phase snapshot per market at identical instants across arms — agrees: 0.167 for the mechanism against 0.145 unlocked and 0.152 locked.

The separation principle

The paper's answer is a division of labour, stated as two claims.

The primary layer settles. No payout depends on the final pool ratio, so end-of-life manipulation of that ratio has nothing to grab: you cannot expropriate settled claims by trading against them. And manipulating the displayed ratio late is strictly more expensive here than in a classic pool or a book, because the manipulator's stake earns nothing even when it is correct.

That is a real security property and it is the reason the hole exists in the first place. The pool is a bad late forecaster precisely because it is an unprofitable place to express a late view — which is the same fact as being an expensive place to lie.

The secondary layer prices. A position is four numbers (outcome, principal, entry accumulator, vintage), so it trades cleanly. An informed trader late in a market does not inject into the pool — P4 makes that pointless — but buys positions from holders. That reroutes late information through prices paid to risk-bearers rather than taken from them.

What is established, and what is not

The paper used to concede this section as a weakness. It now has a number, and the number is better than expected — with a very specific scope.

Adding a stylized dealer to the behavioural study, where every late-arriving informed agent prints a transaction at their belief shaded by half the spread:

Better than every pool arm, by a lot, and flat across a ten-fold range of spreads — which is the surprising part. The layer does not need to be tight to work; it needs to exist.

Now the scope, which the paper states before anyone else can:

The model is deliberately generous (no inventory risk, no quote withdrawal, every informed arrival prints), so read it as an upper bound on what the layer can deliver and a lower bound on nothing.

What it establishes is narrow and useful: the late information exists, and a thin transaction layer suffices to surface it. So the open question is not whether a quote would help. It is the adverse-selection economics of who quotes.

The structural worry that stands

There is no guaranteed source of uninformed selling late in a market.

A standing cash-out vault is a designated adverse-selection sink: it will widen or decline exactly when information is most valuable, because that is when it is most likely to be picked off. That is not a bug to be engineered around; it is the job description.

This is also where λ bites in an unexpected direction. Hedging demand is a principal source of the uninformed flow the layer needs, and λ = 1 removes hedging entirely. Setting the dial to maximize the mechanism's own properties makes the layer that repairs its biggest weakness harder to run. The paper flags this as one of the reasons λ exists at all.

The cheaper repair does not exist

Before accepting that the late price must come from a position market, the paper tested whether it can be recovered from primary-layer state alone — which would be free, since every position's vintage and stake are stored anyway.

Three candidate estimators, against the classic arm as a control under identical treatment (vpm-estimators.mjs, 20 seeds × 2,000 markets):

Recency-weight the pool by vintage. It works — final-phase Brier improves from 0.1736 to 0.1395 at weight (t/T)⁴. But it improves the classic arm more (0.1478 → 0.0942), so the vpm-minus-classic gap widens from 0.0258 ± 0.0007 to 0.0453 ± 0.0014. It is a generic recency result owned by neither mechanism.

Weight toward early vintages. Halves the gap to 0.0126 ± 0.0006, but degrades both arms absolutely (vpm to 0.1948). That is convergence, not repair.

Invert the entry rule into a revealed-belief bound. Destroyed by noise entries, which reveal no bound and are indistinguishable from those that do: final-phase Brier 0.4700, far worse than doing nothing.

Against the dealer arm's 0.054 on the same instants, no primary-layer construction the paper found is competitive.

The negative result is the point: §8 is not a convenience.

That is the strongest statement in the paper about its own weakest section, and it is worth reading as an instruction to implementers. You do not get to skip the layer and post-process your way to a late price.

What to build

Two shapes, by duration.

Markets long enough for a position market to form. Ship transferable positions and a cash-out affordance before real money — §15 makes this a precondition, not a follow-up. The position's payout splits usefully: a floor tranche (principal plus accrued claims — a digital option with a known payout) and a flow tranche (whatever vests later), so a holder can keep conviction and sell activity, or the reverse.

Short recurring rounds. No position market can form inside a fifteen-minute round, so the paper specifies the degenerate form: a venue-side RFQ cash-out at a published spread, and its quotes are the composite's late price.

One structural requirement on that vault, which is easy to miss and expensive to get wrong:

The vault's quote must never be a mechanical function of the current pool ratio.

If it is, the manipulation the separation principle disarms comes straight back through the vault — at P5's near-zero printing cost for a book-dominant party. The quote must price from the venue's own model over wash-excluded flow and off-market inputs. Manipulation-robustness is the second half of the vault's pricing problem, and the paper lists both halves as open.

A venue unwilling to run the vault should run those markets at λ < 1 and accept the other trade instead.

The most valuable missing experiment

The paper names it explicitly in §13.3:

Backtesting the composite estimator against recorded live tapes remains the most valuable missing experiment.

Everything above is simulation. The dealer arm shows a thin layer would carry the late forecast; it does not show who quotes it, at what cost, or how the composite behaves on real order flow with real withdrawal behaviour. Until that exists, the live late price is simply worse here, and any venue adopting the mechanism is accepting that on the paper's own evidence.

Which is the correct place for a whitepaper to leave its weakest claim: with a bound, a negative result ruling out the cheap alternative, a named experiment, and no pretence that the design is finished.

Next: Why there are only a few thousand prediction markets.