Research & Validation

Published test reports for the AQMath Deleverage Modulator. Each report is generated verbatim from the source write-up in the engine repository, so the figures here are the figures the test produced. The point of publishing them is that a drawdown claim is only worth as much as the test behind it.

• OOS Validation — Proteus (v18) on 182 Unseen Baskets2026-09-09 · 中文

  Out-of-sample re-measurement of the live Proteus (v18) engine on the same 182 unseen-token baskets published on /results (those figures were produced by Aegis v14): v18 cuts the worst drop in 182/182 baskets, median -45 pp, and modestly beats v14 (median MaxDD 16.8% vs 18.3%, Calmar +0.55 vs +0.46). No re-tuning; the v14 leg reproduces the published figures exactly.

• Adaptive Exposure Capping — Why Fixed Drawdown Limits Fail in Regime Shifts2026-09-02 · 中文

  Fixed drawdown thresholds are basket-blind: a single number cannot serve both low-vol majors and high-vol alts. Walk-forward OOS on the production path: adaptive (volatility-scaled) wins both baskets on Sharpe, Calmar and MaxDD. Alt basket MaxDD 29.4% vs 36.8% fixed, at identical average exposure. Includes the exposure-matched constant control that separates de-risking from timing, the exposure-minimum vs price-minimum lag diagnostic, and the lookback-truncation inversion that demoted the full-history winner to worst on the production path.

• Locked Out? — What Happens After the Shield Goes Defensive2026-08-18 · 中文

  Crown Test 4 on the production KKT stack: every defensive episode in 6.2 years analysed for re-entry speed and recovery capture. The Shield re-enters fast (median 1 day to 50% exposure) but captures only 32% of the Buy & Hold recovery at 12 months. Zero of seven cycles with 12m data end with Shield equity above B&H. Gates 1/4.

• Static vs. Dynamic — Shield vs the 60/40 Humans Actually Hold2026-08-12 · 中文

  Crown Test 2: Aegis (v14) Shield vs static crypto/stablecoin splits at three levels of human-ness - a robot (yearly rebalance), a drawer (never touched) and a human (capitulates when the pain breaks them, 30 seeds). Gates 2/4: MaxDD 34.2% vs 60.7-81.1% and the best Calmar of all nine strategies pass; the robot 60/40's higher Sharpe (0.621 vs 0.493) and doubled equity fail. But zero of 30 simulated humans held the plan - every one capitulated 3-5 times and gave up $12.5k-$16.4k of equity.

• Human Factor Stress Tests — DCA Jitter, Signal Lag & Frozen Weights2026-08-10 · 中文

  Test 1: DCA jitter costs at most 0.5 pp of MaxDD. Test 2 hits the signals themselves: random delays and missed days pass, but a constant 3-day lag fails the Sharpe gate - $5,058 of damage with zero seed variance. Test 2b re-runs it on the production KKT 60/40 stack: every error scenario breaks the MaxDD gate. Test 2c runs the reader-proposed frozen-weights control and a 1-5 day lag sweep: freezing the weights still fails 3 of 4 scenarios and even a 1-day lag breaks the gate. Same-day execution is the only policy that passes.

• Three Price Feeds, One Strategy — Feed Sensitivity2026-08-09 · 中文

  The pipeline merges three price feeds by per-date median - but 5 of the 8 frozen-plan tokens are collected by the free-tier aggregator alone. Pulling the same eight tokens from three independent feeds shows exchanges agreeing to ~0.1% while the aggregator prints 49 one-day moves above 30% in a year - 41 corroborated by no exchange. The Aegis (v14) replay outcome barely moves, but trade timing and the macro loop's volatility input do.

• How AQMath Works — Signals & What You Execute2026-08-07 · 中文

  The must-read user guide: AQMath produces signal-only trading guidance from frozen KKT risk-parity weights (re-optimised every 180 days) and the daily Aegis (v14) Deleverage Shield - the user executes every trade manually, on their own exchange account.

• Institutional Risk Report — 3,000 Synthetic Histories2026-08-07 · 中文

  Stationary block bootstrap, Deflated Sharpe, White's Reality Check and CVaR on the production Aegis (v14) wiring: the drawdown cut wins in 98.6-99.7% of 3,000 paired synthetic histories per basket and the one-month tail is ~40% thinner everywhere - but excess return over buy & hold is statistically zero.

• Graveyard Gauntlet — Who Saves the Basket When Tokens Die?2026-08-07 · 中文

  A worst-case basket holding CEL, LUNC and FTT through their 2022 deaths: the system draws down 18% vs 62% for buy & hold, and the honest attribution shows the optimizer dodging slow deaths, the Shield absorbing fast ones, and the liveness screen cleaning up afterwards - not before.

• Robustness — How Far Can Reality Deviate Before the Shield Breaks?2026-08-07 · 中文

  129 walk-forward re-runs perturb every user-facing knob one at a time and stress fees to 20x and execution to 2 days late: no cliff found - two knobs are true plateaus, one is a clean protection dial, and the drawdown cut survives late execution almost intact.

• Regime Autopsy — Does the Shield Survive Its Worst Regimes?2026-08-07 · 中文

  Six named crash regimes - May 2021, LUNA, FTX, Aug 2024, 2025-26 - sliced post-hoc on frozen-weight walk-forward series, with the ADV-K2 liveness screen active: 15 of 15 covered regime-basket combinations show a positive out-of-sample drawdown cut, LUNA contagion held to 16.8% vs 44.1% Buy & Hold.

• Liveness Screen — Fixing the Dead-Token Weight Flaw2026-08-07 · 中文

  The KKT risk-parity loop can hand a dead token a real allocation, because a collapsed chart looks low-vol: a Celsius-style zombie won a 24.7% frozen weight in our stress basket. We reproduce the flaw, reject three naive filters, and ship the ADV-K2 cap-zero screen — +8.5% final value, Sharpe 0.24 to 0.32, zero cost on live-token baskets.

• Three-Basket Comparison — Aegis (v14) KKT MACRO + Shield2026-08-05 · 中文

  Walk-forward E2E comparison of three structurally different baskets (majors+gold, veterans+gold, new-gen alts) on the shipped Aegis (v14) stack: shared-window max drawdown 9.8% vs 25.4% Buy & Hold; the Shield cut drawdown 30-36 pp in every window.

• E2E Study — TIA / QNT / XRP + PAXG Anchor2026-08-03 · 中文

  Walk-forward E2E test of a candidate basket (TIA, QNT, XRP, PAXG) on the live dual-speed stack: max drawdown 16.9% vs 24.6% Buy & Hold, beta 0.62; the KKT optimizer zero-weighted TIA on every re-optimisation.

• OOS Validation — Aegis (v14.0) on Unseen Tokens2026-07-17 · 中文

  Out-of-sample test of the shipped AQMath Aegis (v14.0) Deleverage Modulator on 16 baskets of never-tuned tokens: median 54.3 pp drawdown cut, Calmar up in 16/16.

Interactive version: run the same engine on your own price data at /backtest, or see the live paper-trading log at /results.