All experiments replay real on-chain swaps at sqrtPriceX96 precision against an oracle ground truth. This page shows the windows, the sweeps that picked each value, and the honest limits of what was measured.
Piecewise real-time alignment between pool events and the reference feed.
De-trends the pool-vs-fair spread. 62% of raw variance is slow basis — not charged.
Calm-window reference. Volatile uses Pyth ETH/USD on a 12–60 s adaptive cadence.
Historical benchmark feed used as the oracle ground truth for precision.
| WINDOW | POOL / REFERENCE | RESULT |
|---|---|---|
Calm 2026-05-30 · 20:04–22:30 UTC |
Arbitrum WETH/USDC 0.05% Binance ETHUSDC 1 s |
629 swaps · median Δ 0.82 bps · capture 17% |
Volatile 2026-05-23 · 17:00–21:00 UTC · 430 bps |
Same pool Pyth ETH/USD · 12–60 s adaptive |
6,444 swaps · median Δ 3.70 bps · capture ~24–26% · LP bleed 22×/hr |
Of these, 4,511 priced swaps (629 calm + 3,882 volatile) were replayed as the regression-test corpus behind every precision and capture figure.
Capture rises almost perfectly linearly with the premium slope — more α harvests more, but charges more. Precision stays put: ~92.4% calm and ~98.9% volatile across the whole sweep.
PegGuard pins the conservative midpoint — enough capture to return value, low enough to keep the fee gentle on real LPs.
Selected 0.5 bp (50 e2) — the precision/capture knee in each window.
| DEADBAND | CALM P / C | VOL P / C |
|---|---|---|
| 0.0 bp | 90.83 / 28.21 | 98.45 / 26.98 |
| 0.5 bp ▸ | 92.42 / 22.98 | 98.86 / 25.56 |
| 1.0 bp | 93.60 / 18.23 | 99.15 / 24.23 |
Older oracle data can inflate revenue while precision collapses — so the live Pyth-pull guard is a hard rule, not a tunable.
Swept on the corrected volatile fixture (1,108 swaps, 418 bps) and a 250-event calm sentinel. The selected setting is the tightest trigger with zero calm false positives.
It arms the volatile regime 74 minutes before the measured heavy-bleed window. The looser 60 bps trigger only fires in-window — after $2,399 has already bled.
| TRIG | WINDOW | CALM FP | VOL TRIG | 1st VOL | BLED |
|---|---|---|---|---|---|
| 20 | 180 | 0 | 14 | 1825 s | — |
| 20 | 300 | 1 | 15 | 1825 s | — |
| 20 | 600 | 3 | 22 | 1825 s | — |
| 30 | 180 | 0 | 10 | 4332 s | — |
| 30 | 300 | 0 | 12 | 4332 s | — |
| 30 | 600 | 1 | 16 | 4332 s | — |
| 40 ▸ | 180 | 0 | 6 | 4332 s | pre-window |
| 40 | 300 | 0 | 6 | 4332 s | — |
| 40 | 600 | 0 | 8 | 4332 s | — |
| 60 | 180 | 0 | 3 | 12656 s | $2,399 |
| 60 | 300 | 0 | 4 | 12656 s | $2,399 |
| 60 | 600 | 0 | 4 | 12656 s | $2,399 |
Net LP outcome and premium precision per bucket. Live forward-test P&L is deliberately omitted — these are the calibrated calm and volatile windows only.
| BUCKET | NET bps | PREC. | n |
|---|---|---|---|
| <$1k | +1.29 | 84.0% | 295 |
| $1k–$10k | +2.00 | 93.2% | 331 |
| $10k–$50k | +0.43 | 100% | 2 |
| ≥$50k | — | — | 0 |
| BUCKET | NET bps | PREC. | n |
|---|---|---|---|
| <$1k | +0.67 | 99.1% | 958 |
| $1k–$10k | −0.26 | 98.6% | 2,663 |
| $10k–$50k | −3.10 | 98.7% | 254 |
| ≥$50k | +19.69 | 100% | 7 |
The race clock is the pool's ~14 s update interval, not the arbitrageur's sub-2 s execution. Pyth at 2 s staleness retains 88% of edge.
Older data raises measured revenue 25%→41% while precision collapses 95%→72%. Revenue can reward bad operation — so the guard is hard.
An AMM-only basket captured 3%; adding CEX only reached the pure-CEX 44%. Pairs with CEX feeds use Pyth directly.
Regime changes, breaker signals and baseline upkeep have minute-scale budgets where cross-chain delivery latency is acceptable.
Capture is a same-swaps upper bound. Route-away elasticity is not modeled — in production, charging a premium may deter some of the very flow it prices.
No forward-test P&L is claimed. Every figure is a calibrated backtest. PegGuard makes no “profitable in production” claim.
RSC false-positive coverage is partial. Validated on the supplied 250-event calm fixture; broader coverage is an open item.
Guard-mode dual-feed is a skeleton. The stable-pair depeg replay (e.g. USDC/USD ÷ USDT/USD) remains separate; the current path uses a single feed.
Why GUARD charges no directional premium — a provenance audit, not a depeg replay.
Signal sits at feed-noise level — no directional fuel. The breaker is cleanly separated from normal stable noise.