METHOD · 7 OCTOBER 2026

Guard on likely bots: the method

A dedicated cohort of likely bots, identified by public order and fill patterns. The protocol preceded replay outcomes; source recovery, restored-input sensitivity and starting-equity metric corrections followed them. The original-cache result is primary, the restored-cache result a separate sensitivity, and humans descriptive context. This fixed historical engine does not establish current live parity.

Protocol summary

Pre-registered; corrected after outcomes · Dedicated cohort; separate cache sensitivity
Question
How did fixed Guard defaults change modeled losses, drawdowns and returns for accounts classified as likely bots?
Sample
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.
Measure
Loss of half the flow-adjusted stake, paired relative worst-day and drawdown reductions, and full-size entry pass-through. Public order/fill heuristics identify likely automation, not AI or verified bot identity.
Trials counted
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.
Result
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.
Principal limitation
Exploratory subgroups inferred from public orders, with no ground truth. These measurements are additional trials; results for likely bots are not robust yet.
On this page

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. All the studies.

Diagnosis and correction

The research wrapper used raw input starting equity to build threshold-crossing indices and normalize worst-day losses. The frozen replay engine can add pre-window PnL to that raw input when it was backed out of a current account value. Using the raw denominator then starts the index at effective/raw rather than one. This is a wrapper metric bug; engine accounting, risk limits, fills and cached histories are unchanged.

An audit-only pass first replayed both cache profiles and the 84 frozen humans with the existing wrapper, retaining raw/effective starts and old/corrected summaries from identical engine outputs. Six of 148 original-profile analyzed accounts differ, in both defaults and active arms. Two additional original-profile candidates with too few judged entries also differ; they remain excluded by the unchanged filter. All 152 restored-profile accounts and all 84 humans have matching raw/effective starts. Actual and guarded effective starts agree throughout. H1 counts are unchanged in every audited arm, but index ratios and worst-day normalization must still be corrected.

The complete corrected comparison confirms that all reported H1–H6 headline point estimates and intervals are unchanged. Original-profile aggregate statistics still change (22 scalar fields for defaults, 20 for active), chiefly starting-equity and absolute worst-day means plus secondary index statistics. Both restored-profile full headline blocks and the entire human block reproduce their prior values exactly. This numeric stability does not excuse the faulty normalization.

The wrapper now reads validated finite positive out.kpis.actual.start_equity, requires guarded effective start to match, and retains raw input, backed-out flag, both effective starts and first indices in private provenance. Corrected curves start at index one. Offline reruns use separate metric-corrected directories; normal replay refuses to overwrite the preserved raw-start runs directories.

The comparison helper requires exact historical aggregate reproduction against the preserved pre-correction comparison. It separately checks that corrected attempts, exclusions and membership are unchanged and all engine-derived outcomes are identical; only wrapper-normalized statistics and their provenance may differ. It does not claim corrected aggregates equal the original frozen raw-start aggregates. No new requests, S3 writes, draw, classification or engine/risk changes were made for this correction.

Frozen method and two input snapshots

  • Protocol: research/business/bot-cohort-protocol.md; seed 20261011, 10,000 bootstrap resamples, percentile 95% intervals, Wilson share intervals, paired binary bootstrap and exact McNemar. All intervals and p-values remain unadjusted.
  • Engine/data path remains 9aecc6d/a00af40; WASM 662,578 bytes (214,001 gzip), version 0.1.0. The version string alone does not identify this source snapshot. The metric correction is in the research wrapper only.
  • Prior flow semantics are frozen. Deposits move both curves by their amount; withdrawals by the same share of equity. Time-weighted returns/drawdowns remove those flows, but the old halt semantics let a deposit raise equity and hide a day’s loss while leaving withdrawal halts unchanged. This study establishes no parity with the new flows-v4 release or current live Guard accounting; a new engine requires a separately disclosed rerun.
  • Defaults: attach stops, leverage 5×, position 200%, risk at stop 2%, stop distance 2%, open risk 6%, daily stop 6%, drawdown halt 25%. Active sensitivity: 10×, 500%, 3%, 1.5%, 15%, 4%, 35%. No production default is selected or changed.
  • Recorded fills, fees, funding and ledger flows feed the prior replay; guarded fills/stops use its fixed execution/cost assumptions. Main-dex perpetuals only; HIP-3, spot and outcome activity are outside both curves. Counterfactual execution and unchanged subsequent bot decisions are modeling assumptions, not measured live outcomes.
  • H1 is crossing below 0.5 in the flow-adjusted stake index. H2 is the account median relative worst-day reduction for actual losing days, with effective starting equity and the registered −100% loss floor. H3 uses actual positive drawdown; H4 compares return/max(drawdown,2%). H5 pools recorded liquidation events and bootstraps accounts: “flat in the coin at a recorded event,” not proven live prevention. H6 is an account median, not pooled order share.

The original cache profile analyzes 148 bots (M1/M2/M3=56/51/41; strict n=104). The restored cache profile analyzes 152 (56/53/43; strict n=107). These are separate input snapshots. Restored compliance portfolios make four previous starting-equity failures replayable: three input failures and one nonpositive reconstructed start. A four-account ablation makes them valid with portfolios alone; metadata alone preserves the failures. The portfolios were acquired 6 Oct 2026 01:42–02:03 UTC, shared metadata at 01:12 UTC. No fresh state, portfolio or metadata was requested.

Restored portfolios, metadata and candles also change existing-account accounting/execution: before metric correction, 75/148 original defaults rows and 59/148 active rows differed, including nine starting equities and twelve actual curves. The 152 profile is therefore not just four extra members or a code-only effect. Cached metadata is one fixed instrument snapshot, not historical venue rules. The common-member diagnostic below isolates this input change without dropping the four valid additions from the restored profile.

Frozen classifier summary

The earlier classifier was kept fixed for this cohort. Its feature distributions were inspected when its thresholds were originally chosen, so descriptive agreement with those earlier groups is weak support. This is a summary of its frozen rules, without ground truth about bot or AI identity:

Bot evidence:

  • client order ids on orders not routed through a builder (+2, or +1 above 10%);
  • Ioc market orders off builder routes (+1);
  • post-only-heavy order flow (+1);
  • sub-second sequences of separate actions (+2, or +1);
  • trading round the clock (+1);
  • ≥ 100 actions per active day (+1);
  • actions on the minute (+1);
  • ≥ 90% of orders cancelled (+1).

Human evidence:

  • the official web app’s market orders, tif: "FrontendMarket" (−2, or −1);
  • a third-party front end, i.e. a builder fee on most fills (−1);
  • a daily sleep gap (−1);
  • under 180 filled orders in 180 days (−1).

Classes: score ≥ +2 is likely bot, ≤ −2 likely human, anything else unclear. Strict cut-off: ±3.

Corrected defaults and human context headlines

All intervals are 95%; pp means percentage points. Positive H2/H3 indicates lower modeled losses/drawdown.

ProfilenH1 actual → guarded count; paired difference (95%)H2 relative worst-day reduction (95%)H3 relative drawdown reduction (95%)H4 return/drawdown improved (95%)H5 flat/recorded events (95%)H6 median full-size pass (95%)
Original cache, corrected14839 → 10; -19.6 pp (-26.4–-13.5)22.3% (0.04–41.9%); n=13726.4% (13.3–38.4%); n=14870/148, 47.3% (39.4–55.3%)344/588, 58.5% (36.2–66.9%); 19 accounts53.6% (30.1–81.0%); n=148
Restored cache, corrected15240 → 10; -19.7 pp (-26.3–-13.8)24.6% (0.04–41.9%); n=14126.4% (14.2–39.2%); n=15271/152, 46.7% (39.0–54.6%)366/588, 62.2% (43.6–69.3%); 19 accounts54.0% (30.3–82.7%); n=152
Common original members, restored cache14840 → 10; -20.3 pp (-27.0–-14.2)22.3% (0.04–41.9%); n=13726.4% (15.7–38.4%); n=14869/148, 46.6% (38.8–54.6%)366/588, 62.2% (43.8–69.1%); 19 accounts53.0% (29.9–81.0%); n=148
Frozen humans, defaults context8441 → 5; -42.9 pp (-54.8–-31.0)54.1% (37.7–66.7%); n=8462.5% (58.3–66.0%); n=8453/84, 63.1% (52.4–72.6%)49/82, 59.8% (31.0–79.6%); 20 accounts22.8% (5.7–40.0%); n=83

H2’s small positive lower interval is shown with extra precision rather than rounded to zero. One human replay has no judged entries, so H6 uses 83 humans; H1–H4 use all 84. Liquidation event counts have different denominators from account shares.

Costs and follow-up

ProfileMedian return change, guarded − actual (95%), ppDrawdown haltAt least one daily haltMedian follow-up days (95%); IQRCapped / ended early
Original cache, corrected0.019 (-0.3–3.0)43/14860/14884.9 (40.2–90.0); 21.2–180.00/148 capped; 5/148 ended early
Restored cache, corrected0.019 (-0.3–2.7)44/15261/15288.2 (41.3–90.0); 22.0–180.00/152 capped; 5/152 ended early
Frozen humans, defaults context25.860 (8.4–44.9)27/8452/84180.0 (169.8–180.0); 85.3–180.030/84 capped; 7/84 ended early

Halts, lower full-size pass-through, adverse return outcomes and modeling limits belong beside protection figures. Medians do not imply average improvement; severe relative outcomes occur when original losses or drawdowns are small. Flow-adjusted return means can be extreme and are not a reliable headline summary here.

Bot fitted windows end at each account’s own order-history read; human windows end at 6 Oct 2026 01:00 UTC. Bots pass an activity/classifier funnel and ≥50 judged-entry filter. Humans were randomly drawn under survival eligibility, with previously seen outcomes. Duration, endpoints, activity, input coverage and selection differ; context differences cannot be attributed to automation or a causal difference in Guard’s effects.

Exploratory return trade-off, requested after outcomes were known

Jonas asked why smaller losses do not establish higher final returns. An additional descriptive calculation reads the existing corrected rows, grouped by the sign of as-traded return. It adds no replay, data request, rule, population choice, causal attribution or p-value. The groups are defined by observed outcomes, so this is an explanation of this sample, not a strategy selection rule or a pre-registered confirmation.

Original-cache defaults groupnHigher / lower guarded returnMedian paired return change
All analyzed bots14874 / 74+0.019 pp
As-traded return below zero8058 / 22+9.4 pp
As-traded return above zero6816 / 52−4.5 pp

The original sample’s median return improved from −1.12625% as traded to +0.36125% guarded: a difference of +1.4875 percentage points between the two medians. This differs from the median of each account’s paired return change, +0.019 pp (unadjusted 95% bootstrap interval −0.34 to +2.97 pp), because the accounts occupying the middle can change when each return distribution is ranked. The observed sample median improved; the paired estimate remains inconclusive. No interval for the difference between the two medians was calculated.

The restored 152 profile has the same pattern: 76 higher / 76 lower; 60 of 83 originally losing accounts improve (median paired +8.1 pp), while 53 of 69 originally winning accounts give up return (median paired −4.9 pp). These account counts do not identify how many profitable trades Guard blocked or isolate which rule caused the changes. Caps, resizing, stops and halts can alter both profitable and unprofitable paths. Risk protection does not itself establish an expected return advantage.

This is one additional post-outcome analysis, raising the conservative minimum from 70 at correction completion to 71 currently. The preserved correction comparison/helper retain their original completion-time count 70; publication metadata must include this subsequent diagnostic. No confidence interval or causal conclusion is inferred from these outcome-conditioned groups.

Corrected active and robustness sensitivities

Profile/armnH1 actual → guarded; paired difference (95%)H2 median reduction (95%)H3 median reduction (95%)H4 improved (95%)H5 flat/recorded (95%)H6 median full-size pass (95%)
Original cache primary14839 → 10; -19.6 pp (-26.4–-13.5)22.3% (0.04–41.9%); n=13726.4% (13.3–38.4%); n=14870/148, 47.3% (39.4–55.3%)344/588, 58.5% (36.2–66.9%); 19 accounts53.6% (30.1–81.0%); n=148
Original cache active14839 → 14; -16.9 pp (-23.6–-10.8)8.4% (0.0–24.7%); n=13711.9% (0.9–27.0%); n=14867/148, 45.3% (37.5–53.3%)378/588, 64.3% (40.0–75.0%); 19 accounts87.9% (73.1–94.4%); n=148
Restored cache primary15240 → 10; -19.7 pp (-26.3–-13.8)24.6% (0.04–41.9%); n=14126.4% (14.2–39.2%); n=15271/152, 46.7% (39.0–54.6%)366/588, 62.2% (43.6–69.3%); 19 accounts54.0% (30.3–82.7%); n=152
Restored cache active15240 → 14; -17.1 pp (-23.7–-11.2)12.2% (0.0–29.0%); n=14111.9% (1.1–27.5%); n=15270/152, 46.1% (38.3–54.0%)378/588, 64.3% (40.7–75.3%); 19 accounts87.9% (73.1–94.5%); n=152
Defaults profileStrict H1 difference (95%), ppStrict H2 (95%)Strict H3 (95%)Equal-strata H1 (95%), ppEqual-strata H2/H3 (95%)
Original cache, strict n=104-17.3 (-25.0–-10.6)9.9% (0.0–41.9%); n=9517.4% (1.3–31.3%); n=104-19.4 (-25.7–-13.2)H2 24.6% (0.09–42.4%); H3 27.4% (17.3–40.6%)
Restored cache, strict n=107-17.8 (-25.2–-11.2)14.0% (0.0–42.4%); n=9819.6% (2.2–32.9%); n=107-19.7 (-26.2–-13.7)H2 28.6% (0.2–43.8%); H3 27.4% (17.5–40.6%)

The registered robustness rule requires the main interval to exclude zero (H4, 50%) and its direction to persist under strict/equal-strata cuts; it does not require every sensitivity interval to exclude zero. H5/H6 remain descriptive. Registered-engine ab953f8 n=139 sensitivities are preserved in earlier aggregates but were not rerun with effective-start normalization; they cannot serve as a corrected confirmation. No new active human arm was computed.

Frozen human source and context comparison

Source recovery reproduced the full 212-member survival-M block with the unchanged frozen feature extractor and the frozen scoring function: primary bot/human/unclear 19/84/109, strict 10/32/170, zero missing/insufficient, every rule share and every feature median (tolerance 1e-12). A fixed acquisition cutoff at original feature/protocol commit 1e14dbb, 6 Oct 2026 20:39:20 UTC before the first outcome split, excludes later responses mixed into Guard’s backup. Initial unions failed two fill-order medians; no subset or outcome selection repaired labels. Aggregate/features equality supports recovery but is not independent row-by-row identity proof of lost labels.

Missing private classifier input initially produced empty human arms; later RAM cache disappearance had an unconfirmed cause. The cached-only retry yielded 70/84 valid; Jonas approved fixed missing historical fills/funding/ledger/candles at ≤600 weight/min. Final defaults-only context has 84/84 valid. That historical fetch was a post-outcome deviation from the cached-only protocol, completed before this correction. No requests were made during the audit/correction; the classifier and cached inputs remained frozen.

Corrected bot − human contextOriginal-cache bots vs 84 humans, pp (95%)Restored-cache bots vs 84 humans, pp (95%)
H1 paired difference minus human paired difference23.3 (10.0–36.4)23.1 (10.3–36.6)
H2 median reduction minus human median-31.8 (-60.2–-3.9)-29.5 (-58.7–-4.7)
H3 median reduction minus human median-36.0 (-49.5–-22.3)-36.0 (-48.2–-22.2)

Positive H1 means a smaller measured loss-of-half reduction for bots; negative H2/H3 likewise mean smaller measured reductions for bots. Even an interval excluding zero is descriptive context, not a matched/causal test or evidence about automation. Different follow-up and sampling are explicit above. No common-window adjustment, matching or causal identification was performed.

Funnel, multiplicity and remaining decisions

The frozen leaderboard frame was 47,471 accounts; 11,579 remained eligible. Pre-screen pools M1/M2/M3 were 623/407/664; tried 346/190/400; fetch-time members 50/50/43. Strata are all-time ROI terciles; monthly volume is an eligibility threshold. The original corrected engine considered 162 candidates: the 143 fetch-time members plus eligible prior judged-filter failures. Cache profiles change replayable inputs, not the draw or bot classifier. No ground-truth bot/AI identity is claimed, and leaderboard survivorship/activity selection limits generalization.

Conservative minimum 71: 61 registered trials + five pre-outcome deviations (66), plus post-outcome frozen-source reconstruction, replay cache expansion, approved missing historical human fetch, this wrapper metric correction, and the subsequent user-requested return breakdown. Failed source proofs, timestamp diagnosis, repeated replays, cache ablations, audit-only replays and common-member computations are disclosed additional work, not independent confirmations. Bootstrap intervals and p-values are unadjusted. Old report assertions are not grandfathered through this correction.

This publication choice does not turn the restored input snapshot into an independent confirmation. The subsequent return breakdown is labelled exploratory; a separate reviewer checked subgroup counts and medians against the saved corrected rows without replay or requests, with all values matching. Fixed subsequent decisions, static instrument metadata, variable windows, counterfactual execution, pooled liquidation concentration and frozen prior flow semantics remain limits. No live prevention, guaranteed returns or claim that Guard benefits bots more than humans is supported.

Registered population and limitations

The following sections reproduce the registered selection plan; the corrected report above discloses the actual counts and subsequent deviations.

Population

Frame

The leaderboard the compliance study fetched at AS_OF (47,471 rows), with cohort M’s eligibility:

  • account value ≥ $1,000;
  • month volume ≥ $100,000;
  • a finite all-time ROI;
  • not a vault (the stats API’s vault list, which includes HLP);
  • not one of the excluded system addresses.

Excluded besides: every account an earlier Guard study measured or screened by fetch. That is any row of the compliance study’s results-*.jsonl and of the defaults study’s fetch-M.jsonl (640 accounts), so no outcome seen before can enter this cohort.

  • Frame: 11,579 accounts.
  • ROI strata: cohort M’s tercile bounds of all-time ROI (M1 ≤ −6.05%, M2 up to +46.7%, M3 above), so the strata mean what they meant in the survival study. Frame: M1 3,820, M2 3,325, M3 4,434.

Pre-screen 1: archive (free, no requests)

  • Source: our copy of Hydromancer’s per-user perp fills (archives/hydromancer-reservoir/by_dex/hyperliquid/fills/perp/all), the main dex, over the last 90 archive days (6 Jul to 3 Oct 2026). 10,081 frame accounts traded in it.
  • Per account:
    • the bot split’s fill features: api_cloid_share, builder_share, fill_orders, and the taker-fill timing fallback (quiet6, subsec, actions_per_active_day, minute_edge);
    • the maker share of notional;
    • the number of opening orders.
  • Active: ≥ 25 opening orders in the 90 days and an archive maker share < 70%. That leaves 6,746 accounts.
  • Pool: active accounts with a fill-only score ≥ +1: M1 623, M2 407, M3 664.
    • The order rules can add up to +3 (B2, B3, B8) or take away 2 (H1).
    • A score of +1 is therefore one point below the primary cut-off.

Pre-screen 2: one order history (requests)

For each pool account in draw order, round robin over the strata, two requests:

  • historicalOrders: the same body fetchAccountHistory sends, so it is reused;
  • extraAgents.

The provisional score is the bot split’s rules on:

  • the archive’s fill features;
  • timing from the orders where they hold ≥ 50 entry orders, else the archive’s taker fills (the bot split’s own fallback).

An account with a provisional score ≥ +2 is fetched in full.

Full fetch and filters

As in the Guard defaults study (screenAndFetch, filterAccount):

  1. the first fills page: under 50 perp fills in a complete window, or a maker share ≥ 70%, excludes;
  2. fetchAccountHistory;
  3. a maker share of perp fill notional < 70% over all fetched fills;
  4. the replay under the current defaults runs;
  5. ≥ 50 judged entries in the window (entries whose stop is known from the order history).
  • Primary: score ≥ +2, likely bot.
  • Sensitivity: score ≥ +3, strict.

Members:

  • pool accounts, not audits;
  • passed every filter;
  • final class likely bot (≥ +2).

Size (stopping rule): per stratum, until 50 members or 400 pool accounts tried, whichever comes first. The target is 150. A stratum that runs out is reported as short, and the pooled analysis uses what was reached. A pooled n under 100 is labelled underpowered.

Audits: what the funnel misses (reported, never members)

  • In-pool audit. Every 5th pool account per stratum with a provisional score below +2, at most 30 in all, is fetched in full and classified anyway.
  • Outside audit. The first 30 of the active non-pool accounts (fill-only score 0 or −1), shuffled with random.Random(20261012), are pre-screened and fetched in full.
  • Reported: the share of audit accounts whose final class is likely bot and that pass the filters, with Wilson intervals. From these follows a rough estimate of the bots the funnel leaves out.
  • What it means for the population. The population is “likely bots that the funnel finds”. If audits show many missed bots, that is a limit of the cohort’s coverage. It changes nothing in the analysis.

Approved API wallets (extraAgents)

What it can show: that the account has approved, and still has, one or more agent wallets that may sign orders for it.

What it cannot show:

  • that an agent signed any given order;
  • that a program, rather than a person, decided. The official web app’s one-click trading also approves an agent, and an agent can be driven by hand.
  • revoked or expired agents.

Use:

  • descriptive: the share with ≥ 1 agent and with ≥ 1 named agent, by provisional class across every pre-screened account and by final class across every fully fetched one;
  • one sensitivity population: members with ≥ 1 approved agent, for the main measure only (below).

Known limits (stated now)

  • No ground truth. The classifier reads how orders were placed, not who decided. A person can run a script; an agent can click.
  • The funnel defines the population. “Likely bots” here are those that pass an archive fill-score pre-screen and a provisional score. The audits estimate what that misses.
  • Directional bots only. Market makers (maker share ≥ 70%) are excluded, as Guard’s sizing rules are not built for quoting.
  • Short histories for busy bots. At most 10,000 fills and 2,000 orders.
    • A high-frequency bot is followed for days, not months.
    • It may fail the ≥ 50 judged entries filter when its 2,000 orders cover little time.
    • This is reported as an exclusion.
  • Survivorship: drawn from the leaderboard at AS_OF with ≥ $1,000.
  • A what-if that keeps the bot’s later decisions. A real bot behind Guard would see refusals and resizes and react: its state machine would differ.
  • Replay assumptions: web/README.md, “Backtest replay”.
  • Main-dex perps only; the excluded share is reported.
  • One period: the 180 days to 6 Oct 2026.