AI coding assistants write plausible trading code in seconds—and silently leak future data, ignore bid-ask spreads, or size positions that blow through prop-firm daily drawdown floors. quant-swarm gives your AI agent 46 read-only, derived-only verification tools organized into four sequential checkpoints.
The problem in plain English: Over 80% of retail and AI-generated backtests look profitable only because they accidentally peek at future timestamps, assume zero execution friction, or rely on two lucky outlier trades for all their net profit.
How Quant Swarm verifies it: Your agent calls
audit.preflight, audit.run, and
audit.symbolic_check locally on your machine (0 credits;
your raw trades and formulas never leave your laptop).
| Check | Tool Called | Technical Verification Method | Pass / Fail Criterion |
|---|---|---|---|
| Point-in-Time Leakage | warden.audit_features |
Verifies every feature column carries an explicit publication timestamp (available_at ≤ as_of) and flags suspiciously smooth autocorrelation or forward-looking names.
|
0 timestamps with available_at > bar_close |
| Execution Cost Survival | audit.run |
Re-evaluates net Sharpe and expectancy under 1.0x, 1.5x, and 2.0x spread + slippage + borrow cost stress. | Positive expectancy retained at 2.0x friction |
| Outlier Concentration | audit.run |
Computes the share of cumulative PnL contributed by the top-1 and top-5 trades, plus cryptographic SHA-256 input manifest hashing. | Top-5 trades contribute < 55% of total PnL |
| Formula Overfit Hygiene | audit.symbolic_check |
Parses mathematical signal expressions into an Abstract Syntax Tree (AST), counting free constants, nesting depth, and rolling-window alignment. | AST depth ≤ 6; magic constants ≤ 2; lookback ≥ 5 bars |
The problem in plain English: Testing 20 alternative datasets (insider buying, congressional trades, 13F whale flows, short interest, macro surprises) at 4 forward horizons creates 80 overlapping statistical tests. Standard t-stats inflate significance by 2×–4× and produce false discoveries.
How Quant Swarm verifies it: Every dataset in our 82-collector
point-in-time lake is dated by its public SEC/exchange release timestamp and must pass
a strict Two-Stage Statistical Gate via
pit.analysis before your agent ever weights it:
Each observation is standardized against that asset’s trailing point-in-time
distribution (available_at ≤ as_of z-scores, percentile
ranks, and YoY deltas). Any feed with schema breaks, restatement contamination, or
staleness beyond its cadence cap is quarantined as
UNSCORABLE—never zero-filled.
Every L0-clean series is tested for residual return power after size and sector
neutralization across 5d, 21d, 63d, and 252d horizons using Newey-West HAC standard
errors (lag = h − 1), Benjamini-Hochberg FDR across
all tested families, and a 200-draw shuffled-date placebo.
Query derived point-in-time convictions directly via
pit.insider_cluster (Form 4 C-suite clusters),
pit.congress_trades (disclosure-lagged House/Senate
tiers), pit.13f_whales (45-day-lagged institutional
accumulation), and pit.short_squeeze (utilization &
days-to-cover).
The problem in plain English: Even a genuine signal fails if one outsized position or a 3-day volatility spike breaches your account’s daily loss limit or prop-firm consistency rule.
How Quant Swarm verifies it: Before any order or rebalance is staged,
swarm-warden-mcp enforces hard house floors and continuous
crash-protection multipliers:
| Guardrail Layer | Tool Called | Deterministic Enforcement Rule |
|---|---|---|
| House Exposure Floors | warden.validate_order |
Caps single-name weight at 25% of equity, gross exposure at 60% of equity, and rejects dust orders below 0.05% of equity. |
| 4-Tier Defensive Crash Shield | warden.defensive_posture |
Maps live volatility, breadth, and drawdown telemetry into
Normal (1.00x),
Cautious (0.68x),
Defensive (0.42x), or
CrisisShield (0.20x) with automatic stop
tightening.
|
| Prop-Challenge Rule Guard | warden.prop_challenge_check |
Enforces TwoStep (5% daily / 10% max loss) and OneStep (4% daily / 6% trailing) buffers, 30%–40% single-day profit consistency caps, and the 0.55x Target-Lock Taper once progress reaches ≥70% of target. |
The problem in plain English: Strategies tuned in a low-volatility bull market routinely collapse during rate shocks or liquidity crunches.
How Quant Swarm verifies it: Using your 100 free signup bonus credits
(or a £19 one-time Starter pack), swarm-gym-mcp runs paired,
block-bootstrapped stress episodes in the cloud:
gym.probe_fragilityReplays your strategy across Low-Vol Bull, High-Vol Correction, Liquidity Crunch, Rate-Shock Rotation, and Sideways Chop windows to surface your worst-regime tail drawdown before live capital is at risk.
gym.prop_challenge_simSimulates up to 500 multi-day evaluation paths on your R-multiples to output exact Pass Probability (%), Daily Breach Probability (%), and the Optimal Risk-Per-Trade (%) that maximizes expected net payout after challenge fees.
tournament.submit &
forward_register
Runs 100 paired episodes against our active champion genome with a Wilcoxon signed-rank gate, or registers an immutable SHA-256 weight vector into the 8-week out-of-sample Forward League.