METHOD · 6 OCTOBER 2026

Guard defaults study: the method

Every result comes from Guard’s own engine (zunder-risk) through the same replay this site’s backtest uses, run natively for speed and checked against the WebAssembly build on every input. The protocol was pre-registered and committed before any replay ran.

Protocol summary

Pre-registered; corrected after unblinding · Frozen train selection and original validation/test splits
Question
Does any of 5,940 pre-registered rule sets beat Guard’s defaults on accounts it was not tuned on?
Sample
262 Hyperliquid accounts (train 136, validation 63, held-out test 63); each account’s last 180 days to 6 October 2026.
Measure
The pre-registered score, with its constraints, on accounts the rule set was not chosen on; replayed through Guard’s engine.
Trials counted
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.
Result
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.
Principal limitation
A historical what-if replay: resizing or skipping an order can change a bot’s next decision. Fees and funding follow recorded history. Stop execution and missing history are limits detailed below.
On this page

5,940 pre-registered rule sets tried on 262 accounts (train 136, validation 63, held-out test 63); no supported switch in the corrected replay. The original test was rerun after unblinding. On that split the defaults cut median max drawdown by 43 points and let 19% median account pass-through at full size. All the studies.

Method (as run)

  • Engine: the registered deposit-aware replay and judge, run natively from the same Rust function used for WebAssembly. The original parity check reproduced the WebAssembly KPIs on all 608 inputs. That check predates the correction; it does not by itself validate the corrected data inputs. This study retains that registered replay engine, rather than adopting later stop-model changes.
  • Data path: cached public account history is converted into replay inputs. The corrected path uses the most recent 10,000 fills when history exceeds the cap, deduplicates historical orders by order ID using the latest status, starts capped order coverage just after the oldest returned status timestamp, and infers prices from funding payments only when they belong to an hourly funding interval. These repairs alter inputs, not rules or risk limits.
  • Window: 180 days ending 6 Oct 2026, 01:00 UTC. Requests and the observation cutoff stay fixed; this is not an additional market period.
  • Costs: the registered assumption is 4.5 bps fees plus 1 bp slippage per side (11 bps round trip). Deposits, withdrawals, funding and settlements enter the deposit-aware replay. Costs are assumptions, not a promise about fills.
  • Population: 262 accounts: ordinary directional traders M1 80, M2 63, M3 71; long-term winners A 11; recently liquidated B 37. Ordinary traders are leaderboard ROI terciles with account value at least $1,000 and monthly volume at least $100,000. Maker share must be below 70%; at least 50 judged entries are required when drawn. The original membership and account splits are frozen for the correction.
  • Screening: 596 ordinary accounts were tried; M2 and M3 reached the 200-try cap. A excluded 10 and B excluded 43 accounts for fewer than 50 judged entries. Market makers and accounts with too few fills were excluded at the pre-screen. Nine original members now fail the filters (A 2, B 1, M1 3, M2 0, M3 3). The primary correction retains them. Excluding the two affected test members leaves 61 test accounts: guarded flat at 61 of 171 events (35.7%, 95% interval 15.5–54.1), compared with 123 / 260 (47.3% [21.3, 62.9]) in the frozen primary split. The excluded pair contributes 62 of the 123 flat events. This material eligibility sensitivity is post hoc and does not replace the frozen split.
  • Account splits: seed 20261008 within each group; train 136, validation 63, test 63. The test split was first read in the original study and read again for the correction. It is no longer blind.
  • Time split: the registered early and late 90-day analysis is distinct from the account split. A corrected time-split result is not present in the current aggregate; no original time-split estimate is carried forward as corrected evidence.
  • Objective: P is the median account change in return divided by max drawdown, with drawdown floored at 2%. Challenger differences use paired account bootstraps (10,000 resamples). C1 requires at least the defaults’ median pass-through; C2 at least their pooled flat-at-liquidation count; C3 no higher p90 guarded drawdown. The safety floor is applied before recommendation.
  • Pass-through: judged entry orders taken at full size, including an attachment-only resize with at least 99.9% of requested quantity. A smaller but executed trade still counts as friction. Entries before known order coverage remain as traded under every variant.
  • Trials and correction: 5,940 distinct grid variants plus 6 explanatory ablations; the correction repeats the registered analysis on repaired inputs. Repeated analyses and the deviations below must be disclosed alongside the configuration count. No fresh test claim or new preset selection follows from this rerun.

Bias checks

  • Native against WebAssembly: the original check matched KPIs on all 608 inputs. It establishes engine parity for those inputs, not completeness of historical data. The corrected replay keeps the registered engine.
  • Unseen stop touches: 1-second market candles, aggregated to one minute, are used to find crossings between the account’s own fills. The denominator below is rested, unfired attached stops with candle coverage; missing candle coverage is excluded, not counted as an untouched stop.
Rule setRested, unfiredWith candlesTouched unseenShare of covered stops
defaults5,2974,9714599.2%
careful6,2825,86970812.1%
active5,3584,94962812.7%

The narrower 1.5% stop has a higher unseen-touch rate than the defaults’ 2% stop. Under the registered rule its result remains unconfirmed. A real Guard could stop out between the fills this engine sees and miss a later recovery; this can flatter the replay. Candle gaps add uncertainty. A separate later stop-model diagnostic is not substituted into these results.

  • What-if: every account’s later decisions remain as traded, even after a resized or refused entry. This does not model how a trader would adapt.
  • Liquidation distance: not judged because past liquidation prices are unknown; it was not tuned. “Flat at liquidation” means no guarded position in that coin at the recorded event, not proof that a real liquidation was prevented.

What could still be wrong

  • Post-unblinding correction: splits and parameters remain frozen, but the original test was already seen. Data repair is necessary; it cannot restore blindness. New confirmation needs a newly registered population or period.
  • Small, selected sample: 63 test accounts, only two long-term winners. A selects survivors, B liquidations, M today’s active leaderboard traders with at least 50 judged entries. The mixed sample is not representative of every Guard user.
  • Regime: the accounts share one six-month market period. No corrected time-split estimate is claimed here.
  • History caps and coverage: only the most recent 10,000 fills and capped order history are available. Correcting the cap direction does not restore omitted history. The judged window can be short for very active traders; candle coverage is incomplete.
  • Replay assumptions: realised rather than continuously marked equity, stops sampled at account fills, fixed fee/slippage assumptions, retained future decisions and unjudged liquidation distance. Flow handling and early replay endings can affect coverage.
  • Friction and pooled events: full-size pass-through treats useful smaller trades as friction. Liquidation counts pool events from accounts with repeated liquidations; uncertainty is bootstrapped over accounts.

Deviations

The following historical changes preceded the original test. They are retained as the audit trail; the final correction below occurred after results had been seen.

  1. The engine and data path moved to main’s twice, before any selection result.

    • At 05:40 UTC, main’s replay left HIP-3 and outcome markets out, marked both curves at the window start and booked settlements. Cohort M was judged again with it (study.ts refilter).
    • At about 12:50 UTC, main’s replay started reading the ledger and taking the starting equity from the account’s history on the deposit-aware engine. An earlier train grid on the previous engine was stopped with no result read.
    • A third train grid on the deposit-aware engine was interrupted and resumed. The replays that end early at a withdrawal were aligned (a runner fix: one trade per fill before the replay’s end). No result was read between the cut and the fix.
  2. Requests beyond the compliance cache. The current path also asks for window-start candles, funding from the first served fill and the ledger. These were fetched (A and B ledgers at 300 a minute), and cohort M’s ledgers were read from the survival study’s cache.

  3. Sub-window starts. A sub-window may also start from the account’s first deposits when the account was empty at its start. A start backed out from today’s value is still refused.

  4. A cheaper pre-screen bound during the cohort M fetch: orders that open or add to a position after the oldest order served. It can only skip accounts that cannot pass the 50-entry filter.

  5. Analysis added afterwards, outside the selection: the 6 ablation variants, the per-entry reasons and the flow estimate.

  6. Cohort M re-judged with the new engine. M1 then had 71 passes in 173 tries, so it was fetched further until 80 passed (196 tried), as the protocol’s stopping rule requires. M2 and M3 stayed at their 200-try caps. Members are the first 80 passing in draw order.

  7. The original test list was fixed before the test: defaults plus careful and active, because no challenger passed validation.

  8. 7 Oct: rerun after unblinding to repair the data path. Four defects were corrected: wrong end of capped fill history, duplicated order statuses, order coverage inferred from order creation rather than the oldest returned status, and prices inferred from non-hourly funding payments. The registered deposit-aware replay, observation cutoff, cohort membership, split seed, grid, original train top 10 and originally frozen test variants were retained. The repeated analysis replaces the report’s results; it is not a fresh held-out look. Repeating variants does not create new parameter configurations, but the rerun is an additional analysis and deviation. A separate later stop-model experiment is diagnostic only. Corrected time-split confirmation remains unavailable in the current aggregate.