RESEARCH · DATA TO 6 OCTOBER 2026

What the data says about Guard’s rules

Four studies on real Hyperliquid data, with registered protocols, disclosed corrections after outcomes, methods, limits and trial counts. Exploratory subgroups are identified separately. Aggregates only: no account is named or shown.

Three findings to start with

  1. Defaults held up out of sample

    No supported default switch in the corrected replay: the best challenger scored +0.092 (95% interval −0.154 to +0.289). Recommendation: keep the current defaults.

    Sample and trials

    262 Hyperliquid accounts (train 136, validation 63, held-out test 63); each account’s last 180 days to 6 October 2026.

    5,946 rule-set variants (5,940 grid, 6 ablation); repeated analyses and deviations are disclosed in the method. The original test was already seen before the correction.

  2. Liquidations and rule violations

    91% of liquidated positions against 38% of the others; relative risk 15.7 (95% interval 15.0–16.4). An association, not a claim that Guard prevents liquidations.

    Sample and trials

    All of Hyperliquid over 30 days, 4 Sep – 3 October 2026: 95,095 liquidated positions and 7,942,674 position-days that were not; 62% of liquidations covered.

    18 comparisons reported, listed in the method.

  3. Guard on likely bots: corrected replay

    26.4% → 6.8% lost half their stake; paired difference −19.6 percentage points. Sample median return rose; the paired return estimate is inconclusive. Fixed historical replay engine, without current live or flows-v4 parity.

    Sample and trials

    148 likely bots in the original-cache primary; 152 in a separate restored-input sensitivity, never pooled. 84 likely humans provide descriptive context with different selection and follow-up.

    Conservative minimum 71 measurements/deviations, including post-outcome source recovery, input restoration, metric correction and an exploratory return breakdown. Intervals are unadjusted; repeated cached calculations are not independent confirmations.

The principal trade-off

It costs the best traders part of their gains: in the replay the median long-term winner’s return fell from 54% to 48%. Opening the halt alone does not recover that median return. With all five engine limits opened, the median is 44%; attached stops, the position cap and costs still apply. This does not isolate their separate effects. 9 winners, training and validation accounts.

In the original 63-account test split, the median account’s full-size entry pass-through was 19% (95% interval 2–32%). Separately, across 76,071 pooled entries from 199 training and validation accounts, 39% were held by a halt (35% by the drawdown halt, once an account was already 25% down), 21% refused and 11% resized.

See the cost alongside the protection →

These are historical measurements. The full findings below retain their intervals, methods and limits; they do not promise future results.

The studies
StudyDataTrialsDesign
Guard defaults study 6 October 2026 5,946 Pre-registered; corrected after unblinding Frozen train selection and original validation/test splits
How traders compare with Guard’s defaults 6 October 2026 42 Pre-registered Two cohorts, replayed
Liquidation autopsy 4 Sep – 3 October 2026 18 Pre-registered A count over the whole market
Guard on likely bots 7 October 2026 71 Pre-registered; corrected after outcomes Dedicated cohort; separate cache sensitivity

Trials: every variant or measurement the study took, as its method lists them. Each study’s protocol summary is at the top of its method page.

Guard defaults study · 6 October 2026

Defaults against alternatives

We tested 5,940 registered rule sets on the real trades of 262 Hyperliquid accounts. The corrected replay did not support changing our defaults.

A replay of each account’s last 180 days to 6 October 2026 through Guard’s engine. “Beat” means our pre-registered score, with its constraints, on accounts the rule set was not chosen on; the best challenger scored +0.092 (95% interval −0.15 to +0.29). The original test split was rerun after unblinding to correct historical-data defects; this is not a fresh independent confirmation.

How to read this

The dot is the best challenger’s score against the defaults, the bar its 95% interval. The defaults sit at 0. An interval that crosses 0 gives no evidence that the challenger is better.

Best of 5,940 challengerson accounts it was not chosen on
+0.092−0.15 to +0.29

Pre-registered score, best challenger minus the defaults. Right of 0 favours the challenger.

Guard defaults study · held-out test, 63 accounts

The held-out test

Original test split: 63 accounts, corrected after the original results had been seen. Rules and membership stayed fixed; this is not a fresh blind test. Medians; 95% bootstrap intervals. A mixed sample: ordinary active traders, long-term winners and recently liquidated accounts. Post-hoc filter sensitivity: excluding 2 now-ineligible test accounts leaves 61 accounts and 61 of 171 events flat (35.7%; 95% interval 15.5–54.1%). Those excluded accounts supply 62 of the primary split’s 123 flat events.

How to read this

Left: each measure with its 95% bootstrap interval; the drawdown cut is in percentage points. Pass-through is the median of each account’s entry share; liquidation flatness is a pooled event share. Right: the median account’s worst day as traded and behind the defaults; switch a series off with its key. Hover, focus or tap a row for the exact values.

  • −43 ptsmax drawdown, median accountcut by 43 points (95% interval 23–58)63 accounts held out
  • −67% → −17%worst day, median accountas traded, then behind the defaults; no interval computed for this difference63 accounts held out
  • 60%return per unit of drawdown improvedshare of accounts (95% interval 48–73%)63 accounts held out
  • 47%liquidation events with the guarded account flatin 123 of 260 liquidation events, the guarded account held no position in that coin (95% interval 21–63%). Not “prevented”: positions Guard kept smaller are not shown to survive, and part of the sample was picked for having been liquidated.63 accounts held out
Max drawdown cut, median accountpercentage points
−43 pts23–58
Return per unit of drawdown improvedshare of accounts
60%48–73%
Liquidation events, guarded account flat123 of 260 events
47%21–63%
Entries through at full sizemedian account’s share
19%2–32%

Medians and shares, 63 accounts held out; 95% bootstrap intervals.

As traded
−67%
Behind the defaults
−17%

Return on the worst day, median account, 63 accounts held out.

Compliance study · data as of 6 October 2026

Winners and liquidated accounts

Long-term winners mostly sized their trades within Guard’s limits; liquidated accounts mostly did not.

75%
long-term winners · median share of entries within Guard’s limits (95% interval 7–100%) · 18 accounts
0%
liquidated accounts · the same share (95% interval 0–4%) · 57 accounts
How to read this

Each bar counts the accounts whose share of entries within Guard’s limits falls in that 10-point band. The cohorts differ in size (18 and 57): switch to “Share of cohort” to compare their shapes. Under each cohort, the dot is its median and the bar the median’s 95% bootstrap interval.

Limits of this result

“Within” counts Guard’s own 2% stop: most winners used no resting stop. Winners are survivors of today’s leaderboard, with much larger accounts. Last 180 days to 6 October 2026; descriptive, not a promise.

Long-term winners n = 18

6
1
0
0
1
0
1
0
1
8

median 75% · 95% interval 7%–100%

Liquidated accounts n = 57

42
4
5
1
2
1
1
0
0
1

median 0% · 95% interval 0%–4%

Share of each account’s entries that fit Guard’s limits with Guard’s 2% stop, in 10-point bands; the median and its 95% interval below each cohort.

Liquidation autopsy · 4 Sep – 3 October 2026

Before liquidation

91% of liquidated positions broke Guard’s leverage or liquidation-buffer default that morning, against 38% of the others.

91%
liquidated positions over the limits · leverage above 5× or liquidation within 10% of entry (95% interval 90–91%) · 95,095 positions
38%
positions not liquidated that day · the same rules, the base rate (95% interval 37–38%) · 7,942,674 position-days
How to read this

The share of positions over Guard’s leverage or liquidation-buffer default at that morning’s snapshot, for positions liquidated later that day and for all others. The samples are large, so the 95% intervals are narrower than the dots.

Limits of this result

Over the 30 days 4 Sep – 3 October 2026, all of Hyperliquid. Read from that morning’s snapshot of open positions, not at entry; covers the 62% of liquidations whose position was open at the snapshot. An association, not a claim that Guard prevents liquidations.

Liquidated that day95,095 positions
91%90–91%
Not liquidated7,942,674 position-days
38%37–38%

Share of positions over the limits at the morning snapshot, with 95% intervals.

Relative risk 15.7 (95% interval 15.0–16.4): the association between breaking the limits and being liquidated that day.

Guard defaults study · the cost

The cost of protection

It costs the best traders part of their gains: in the replay the median long-term winner’s return fell from 54% to 48%. Opening the halt alone does not recover that median return. With all five engine limits opened, the median is 44%; attached stops, the position cap and costs still apply. This does not isolate their separate effects. 9 winners, training and validation accounts.

In the original 63-account test split, the median account’s full-size entry pass-through was 19% (95% interval 2–32%). Separately, across 76,071 pooled entries from 199 training and validation accounts, 39% were held by a halt (35% by the drawdown halt, once an account was already 25% down), 21% refused and 11% resized.

The “active” preset lets more trades through (paper and testnet only for now).

How to read this

Left: the median long-term winner’s return as traded, behind the defaults, and behind the defaults with all five engine limits opened to their schema bounds. Right: pooled outcomes for the 76,071 entries it judged across 199 training and validation accounts; “by the drawdown halt” is part of “held by a halt”.

As traded
+54%
Behind the defaults
+48%
Limits opened
+44%

Median return over the replay window, 9 long-term winners (training and validation accounts).

Held by a halt
39%
by the drawdown halt
35%
Refused
21%
Resized
11%

Share of 76,071 judged entries from 199 accounts.

Dedicated bot cohort · 148 likely bots · corrected replay

Guard on likely bots

In the corrected original-cache replay of 148 likely bots, the share losing half their flow-adjusted stake fell from 26.4% to 6.8% behind the defaults. The sample median return rose; the median within-account return change remains inconclusive.

Paired loss-of-half difference: −19.6 percentage points (95% interval −26.4 to −13.5 pp; n=148). Original-cache primary; fixed historical engine and defaults.

The cost: Guard shrinks or blocks entries, and limits can reduce upside too. Full-size pass-through, halts and return uncertainty are shown alongside the loss results below.

The sample median return rose from −1.13% as traded to +0.36% guarded: a +1.49 pp difference of medians, with no interval calculated for that difference. The median within-account change was +0.019 pp (95% interval −0.3 to +3.0 pp): the paired estimate is inconclusive. These are different summaries of the same 148 accounts.

Likely bots are identified from public order and fill patterns, without ground truth about automation or AI. Median follow-up was 84.9 days (IQR 21.2–180.0); these are modeled historical paths, not live outcomes. The fixed study engine predates flows-v4 and does not establish parity with current live Guard.

How to read this

The left chart counts accounts whose flow-adjusted stake index fell below half its start. Each account is replayed as traded and behind the defaults. The paired difference uses a paired bootstrap, rather than subtracting the two share intervals. The right chart shows median paired relative reductions: worst-day loss among accounts with actual losing days, and drawdown among accounts with positive actual drawdown. Relative reductions are not percentage-point cuts.

Limits of this result

Fees, funding and ledger flows enter a fixed historical replay with assumed execution and stops. Later bot decisions remain fixed even when Guard changes an earlier trade. Main-dex perpetuals only. The engine retains prior deposit/withdrawal halt behavior: this study establishes no parity with current live Guard or flows-v4.

The metric correction and source recovery occurred after outcomes were seen. Original-cache and restored-cache profiles use different portfolios, metadata and candles; existing accounts change too. They are not independent confirmations. All 95% intervals are unadjusted across a conservative minimum 71 measurements/deviations. Leaderboard survivorship, activity filters and variable follow-up limit generalization.

As traded39 of 148 accounts
26.4%19.9–34.0%
Behind the defaults10 of 148 accounts
6.8%3.7–12.0%

Share of accounts; bars are unadjusted 95% Wilson intervals.

Original-cache primary: 148 likely bots; same accounts in both paths.
Worst-day loss137 accounts · relative reduction
22.3%0.04–41.9%
Max drawdown148 accounts · relative reduction
26.4%13.3–38.4%

Relative reduction, not percentage points; unadjusted 95% bootstrap intervals.

The worst-day lower interval is 0.04%; its small positive value should be read with the study’s multiplicity and modeling limits.
  • 53.6%entries through at full sizeMedian account share; 95% interval 30.1–81.0%n=148 accounts
  • 43/148accounts hit the drawdown halt29.1%; 95% interval 22.3–36.8%n=148 accounts
  • 60/148accounts hit a daily halt40.5%; 95% interval 33.0–48.6%n=148 accounts
  • +0.019 ppmedian paired return change95% interval −0.3 to +3.0 pp; includes zeron=148 accounts
Separate restored-input sensitivity · 152 likely bots

Loss of half: 40/152 (26.3%) → 10/152 (6.6%); paired difference −19.7 pp (95% interval −26.3 to −13.8 pp). Relative worst-day reduction 24.6% (95% interval 0.04–41.9%; n=141); drawdown reduction 26.4% (95% interval 14.2–39.2%; n=152).

A separate input snapshot, not just extra accounts. Never pooled with the primary. Median full-size pass-through 54.0% (95% interval 30.3–82.7%; n=152); drawdown halt 44/152; median paired return change +0.019 pp (95% interval −0.3 to +2.7 pp).

Human context · 84 likely humans · descriptive only

The human sample has different membership, selection and windows: median follow-up 180 days, ending at a shared historical date; bot windows end at each account’s order-history read. These differences prevent attributing the gap to automation or a causal difference in Guard’s effects.

In this separate defaults replay, loss of half was 41/84 → 5/84; paired difference −42.9 pp (95% interval −54.8 to −30.9 pp). Full-size pass-through was 22.8% (95% interval 5.7–40.0%; n=83, one account had no judged entries). This is context, not a matched comparison.

Guard defaults study · an exploratory trade-off

The “active” experiment

The frozen “active” rule set passed more entries at full size in the corrected replay, with deeper drawdowns. Similar flat-at-liquidation counts do not prove equal protection. This did not meet the criteria for a recommended preset.

  • On the original test split: median account full-size pass-through 44% (95% interval 23–57%) against 19% for the defaults; liquidation events with the guarded account flat 121 against 123 of 260; the worst decile’s drawdown 51% against 49%.
  • Measured with a 1.5% attached stop, a result the study could not confirm (the replay misses more stop-outs at 1.5%); the exploratory replay link below uses the defaults’ 2% stop and leverage, so its results are not the measured rule set above.
  • Not the default: it breaks the study’s pre-registered worst-case constraint. Its limits are looser than Guard’s mainnet ceiling, so these findings do not authorize a mainnet configuration.
Defaults
19%2–32%
“Active”
44%23–57%

Median account full-size entry pass-through, original test split, with 95% intervals.

Defaults
49%
“Active”
51%

Max drawdown of the worst decile (90th percentile) of held-out accounts.

The rules that differ
RuleDefaultsActive
Loss at the stop2%3%
Position size200%500%
Open risk6%15%
Daily loss stop6%4%
Drawdown halt25%35%
Leverage, attached stop, liquidation bufferas the defaultsas the defaults

What these numbers can and cannot carry

  • The defaults study was corrected after its test results had been seen. Its original train top ten, rules and account splits stayed fixed; no new blind confirmation or corrected time-split result is claimed.
  • What-if replays: each account’s own trades over the 180 days to 6 October 2026, judged by Guard’s engine; later decisions are kept as traded.
  • The replay sees Guard’s attached stops only at the account’s own fills. One-second prices show that 9% of the default stops the replay never fired, among those with candle coverage, were touched in between. A real Guard could close those positions at a loss and miss a recovery; this can flatter the replay.
  • These replay studies cover main-dex perpetuals; HIP-3, spot and outcome markets are excluded.
  • The dedicated bot cohort uses public order/fill heuristics without verified automation or AI identity. Its original-cache primary and restored-input sensitivity remain separate; human context has different selection and follow-up. The study’s fixed historical engine does not establish current live parity.
  • The liquidation autopsy covers 4 Sep – 3 October 2026, read at a daily snapshot.
  • Past behaviour, not a forecast, and no promise of returns.

Each study’s own limits, deviations and trial count are on its method page: Guard defaults study · How traders compare with Guard’s defaults · Liquidation autopsy · Guard on likely bots.