Proof, batch by batch.
Every batch clears at one price, inside a band. In session the band is centred on a reference published after the batch closed; while the market is closed, on the last close. This page is the record as it happens, and the research behind the rule.
Who pays the latency sniper?
Five venues, one price path: 1,000,000 Monad blocks (83 hours) at 45% volatility, with 24 news jumps a day. A sniper who sees the true price mid-block trades whenever it has an edge. Unison's rows run the real clearing engine.
| Venue | The sniper takes, a day | Liquidity keeps | Uninformed traders pay | Sniper fills |
|---|---|---|---|---|
| Constant-product AMMfee 30 bp | $6,171 | $13,459 | 69.1 bp | 12,563 |
| Oracle AMM±10 bp, 50 bp push trigger | $473 | $2,175 | 9.3 bp | 63 |
| Order bookmakers ±2 bp, half the races lost | $487 | $84 | 2.0 bp | 5 |
| Unisonvault ±10 bp, fee 3 bp | $0 | $2,858 | 13.0 bp | 0 |
| Unisonvault ±2 bp, fee 1 bp | $0 | $646 | 3.3 bp | 0 |
| Unison, rule brokenreference published before the batch closed | $1,000 | $1,740 | 13.0 bp | 8 |
Sniping doesn't pay.
In session, the auction prices against a reference published after the batch closed. A mid-batch edge only fills when the price has already moved against it, so the sniper stops trading.
Liquidity keeps what snipers took.
At the vault's shipped setting (±10 bp, 3 bp fee) it keeps $2,858 a day; order-book makers keep $84. Takers pay for the width: 13.0 bp against 2.0 bp. At an identical ±2 bp quote, a setting no live vault runs yet, it would keep 7.7× as much ($646) and takers would pay 3.3 bp. All of it in market hours.
The rule is the mechanism.
Break it on purpose, pricing against a reference published before the close, and the sniper's edge returns at once. So the venue enforces it on-chain: each reference is bound to its batch, and one published too early is rejected.
Assumptions and limits, stated ›
- Stylized actors: one sniper with a perfect mid-block feed; order-book makers refresh with one block of latency and races are coin flips.
- The reference relay is modelled with no lag at the batch close. A relay lagging by δ gives back an edge of order σ·√δ, far below a 2 bp spread over tens of milliseconds except during news, where discovery and halt bands cap it.
- Synthetic, independent uninformed flow; no strategic market makers or inventory aversion.
- Live data supersedes this: the tape above is the venue's real record, and replaces the simulation as it accumulates.