Liquidation autopsy: the method

91% of 95,095 liquidated positions broke Guard’s leverage or liquidation-buffer default at that morning’s snapshot, against 38% of positions not liquidated; about 16 times as likely to be liquidated that day. All the studies.

How it was done

• Data, archived only. Our bucket’s copy of Hydromancer’s public archive:
  • the daily position snapshot (snapshots/perp, about 00:08 UTC, every open main-dex perp position with entry price, liquidation price, notional, account value and leverage setting);
  • all perp fills of the previous day (fills/perp/all, for “opened in the 24 h before”).
  • No request was sent to Hyperliquid.
• Unit: every snapshot position with notional ≥ $10 and account value > 0, outside the system addresses. Liquidated: a liquidation fill of the same account and market, on the snapshot’s side, between the snapshot and the end of the day.
• Rules: Guard’s defaults (rules schema v1), each judged from the snapshot as the protocol’s table says. Not one of them is run through the engine: the engine needs a trade and a stop, and the snapshot holds neither.
• Why not the recorder. The recorder’s trades carry no liquidation flag, and its userFills cover the HLP liquidator vaults only (backstop liquidations; on 3 Oct every liquidation fill was method market). Its clearinghouseState sweeps cover about 15,000 large accounts. The Hydromancer snapshot is the cheapest honest source of every position’s leverage and liquidation price.

Limits

• Not at entry. The rules are read at the daily snapshot. Leverage and liquidation price drift with the price and, in cross margin, with the account’s other positions. A position that was within the rules when opened may break them by the next morning because it lost money. The 24-hour subset narrows the gap, and its result is the same.
• Mechanical link. A liquidation price close to the entry makes a liquidation likelier by definition, and R2 alone separates the groups most (76.8% against 12.8%).
  • The relative risk says how strongly the rules separate positions that were liquidated from those that were not.
  • It says nothing about what Guard would have prevented. A refused trade might have been replaced by another, and the replay of real accounts (survival-study.md, compliance-study.md) is the place for what-ifs.
• Coverage: 62.5% of liquidations, by count, are seen (above).
• Equity. The snapshot’s account value can include assets outside the perp account (unified and portfolio-margin accounts), which lowers measured leverage. A third of positions report no liquidation price and pass R2 by construction. Both bias the shares down.
• Stops are not in the data. R3 assumes Guard’s attached 2% stop. A trader’s own tighter stop would allow a larger position, so R3 overstates breaches for traders with stops; the compliance study found that few use them.
• Daily intervals. The per-day files use Wilson intervals on positions. These ignore that one account holds several positions, so they are too narrow. The pooled intervals are clustered by account.
• One month, a volatile one. The share of open positions breaking the rules (the base) moved only between 36% and 40%.

Trial count

Nothing else was tried. No definition, threshold or subset was changed after the first share was computed.

Deviations

None in method.

• The pooled per-rule table sums the day files’ counts. It is descriptive, as the protocol lists it.