F-0014 · first recorded 2026-08-22 · updated 2026-08-24
Both venues' ±1% BTC book held roughly a quarter of a billion dollars; measured retail-size sweeps cost about 1 bp and typical clusters were small relative to standing depth
PRELIMINARY
| Observation | standing notional within ±1% of mid is $247.6M median on Binance futures and $225.6M median on Hyperliquid — within 9% of each other. At burst level, a same-side sweep worth 0.5–1% of that book costs the taker 0.8–1.1 bps of slippage. The median published liquidation cluster is 0.44% of the Hyperliquid book; the p90 cluster is 5.1%. |
|---|---|
| Sample | 221,778 aggressive bursts ≥ $50,000 over 2 Binance futures days (2025-03-10/11), joined to bookDepth ±1%; 5,112 Hyperliquid l2Book snapshots 2026-08-20/21; 433 published clusters |
| Method | contiguous same-side aggTrades within 200 ms grouped into one burst; cost = executed VWAP against the price at which the burst began, signed against the taker; conditioned on the PREVIOUS minute's range so that volatility already in progress cannot be credited to the burst |
| What is not established | two Binance days and 34 hours of Hyperliquid book. The order of magnitude is not in doubt; the day-to-day variation is not established, and neither is behaviour in a stressed regime, where F-0002 measured near-book depth falling to 0.657 |
| Who else publishes this | liquidation heatmaps publish cluster notional without ever stating it as a fraction of the book standing in front of it, which is the only form in which the number means anything |
| Evidence | market/impact2.py, market/impact.py, market/CONTRACT-impact.md §4 P1/P2, ~/genesis-evidence/hl2/btc-l2book.jsonl |
This was meant to be pre-condition P1 of IMPACT-1 — a check that minute-level aggregation had not distorted the cost estimate. It found a distortion in the opposite direction to the one the contract predicted, and the correction is larger than the experiment.
What was measured
Burst-level cost, median bps paid by the taker, by the previous minute’s range (the control) and the burst’s size as a fraction of the standing ±1% book:
| prior range | 0+ | 0.005+ | 0.01+ | 0.02+ |
|---|---|---|---|---|
| 5–10 bps | 0.0 (25k) | 0.8 | 1.4 | 2.3 |
| 10–20 bps | 0.0 (74k) | 1.0 | 1.6 | 2.7 |
| 20–40 bps | 0.1 (85k) | 1.1 | 1.8 | 3.0 |
| 40+ bps | 0.1 (25k) | 1.1 | 2.2 | 3.7 |
The rows are not flat, so the relationship IMPACT-1 set out to measure is real at burst resolution too. But the magnitudes are an order of magnitude below the minute-level table — 0.1 to 3.7 bps against 16 to 44 — for a reason that is obvious in hindsight and was not stated in the contract: the minute-level outcome was the whole minute’s high-low excursion, which is everyone’s trading, while this is what one taker actually paid.
The number that reframes the product
| ±1% standing notional | |
|---|---|
| Binance futures BTCUSDT | $247.6M median |
| Hyperliquid BTC | $225.6M median |
Two things follow, and they point in opposite directions.
1. F-0006’s transfer assumption looks far safer than assumed. The two venues’ BTC books are within 9% of each other in depth. Binance-derived microstructure is not obviously the wrong model for Hyperliquid at this asset. This does not settle F-0006 — depth parity is not the same as identical response under stress — but it removes the largest reason to doubt it.
2. The measured BTC retail-size execution cost is small. A $2.1M position is 0.9% of that book. In the sampled conditions, closing that size in one sweep cost about 1 bp — roughly $200. Whether anyone pays for that information is a product question, not a result established by this measurement.
And it explains F-0010 rather than merely repeating it
F-0010 measured +40.07 bps at 15 minutes after liquidation episodes, which lost to a volatility-matched control at 44.52 bps. That rejected the tested cascade signal without identifying a mechanism.
Here is the mechanism, in the same units:
| published cluster | notional | as % of HL ±1% book | implied one-sweep displacement |
|---|---|---|---|
| median | $1.0M | 0.44% | under 2 bps |
| p90 | $11.4M | 5.1% | roughly 4 bps |
| p99 | $85.8M | 38% | outside the measured range |
The measured one-sweep displacement for a p90-sized cluster is about a tenth of the move F-0010 observed. This scale comparison is consistent with the cluster being largely a passenger on the move, but does not by itself prove that causal interpretation. Liquidation heatmaps are drawn at a scale that invites the reader to compare clusters to each other, never to the book — and against the measured book, almost all of them are small.
Where the value survives
Nothing here says impact modelling is worthless. It says it is worthless where the book is deep, and both BTC books are deep. The measurement retains value in exactly three places, none of which has been measured yet:
- Thin markets — Hyperliquid altcoins, where a $2M position is a meaningful share of the book
- Stressed regimes — F-0002 measured near-book depth at 0.657 in the worst quarter; the cost that matters is the cost when depth is gone, and every estimate above is a median across calm
- The tail cluster — the p99 at 38% of the book is a different object from the median at 0.44%, and IMPACT-1 P2 forbids quoting it because the cells backing it are empty
The methodological note
The contract’s P1 asserted that minute aggregation understates single-sweep cost. It
overstates it, by an order of magnitude. The assertion was written before the measurement and was
wrong, which is the correct order in which to be wrong; it is recorded in
market/CONTRACT-impact.md as an amendment rather than a silent edit, because the contract was
still unfrozen when the error was found.