Findings: the full arc, fail included.

Every number on this page is computed from committed, hash-receipted artifacts in the repository — the pre-registration, the published failure, the rebuild, and the holdout verdict. Nothing here was chosen after the fact. Try the fee engine →

1 · The problem: LPs pay every single day

Native ETH/USDC pool, Unichain mainnet. Adverse-selection cost per sampled day across the calendar-2025 validation window.

Pooled cohort: 126 of 126 active pool-days negative for LPs · this pool: $4.5M net adverse selection ($8.1M gross losses, sampled days)

2 · The pain is concentrated

3 · The honest failure

The v1 signal could not predict individual toxic trades. Published exactly as pre-registered — which is why the later pass is credible.

4 · The diagnosis → the rebuild

Toxic flow clusters in time. A fee that rises with realized volatility collects while it lasts. The deployable signal — EMA of the pool's own 60-second realized move, no oracle — recovered the volatility component at ρ = 0.36, and the calibrated policy beat static fees in 18 of 18 out-of-sample configurations.

5 · It runs on-chain

SentinelHookV1, live on Unichain Sepolia — the demo's 72 fee updates, from committed broadcast receipts (each verifiable on Uniscan):

14k gas overhead per swap (budget: 40k) · 100,000-step stateful fuzz campaign, zero invariant violations · manipulation suite: split-trade dodge not cheaper, poisoning costs the attacker 3.29 units per 3 minutes with zero revenue.

6 · The holdout verdict — PASS

Jan–Jul 2026, evaluated exactly once under a protocol frozen before ingestion. 1,243,673 labeled trades, 161 pool-day clusters, zero parameter changes.

ΔNet vs static 5bps by pool · M2 precision 0.65, coverage 2.76× · M1 precision 0.56, coverage 4.41×

7 · Artifacts

Final performance claims are bounded by the replay methodology: trade sizes are held fixed, so volume elasticity is measured only in the pilot. The failure of Gate 1 is part of the record on purpose.