The receipts.
Marketing claims are cheap. This page renders the same numbers our authenticated dashboard does, fetched live from the audit log. If you spot a discrepancy between what we claim and what this shows, email hello@polyquantlab.com and we'll fix the claim.
How the audit works
Three rules. No exceptions.
The instant the engine surfaces a row, model_yes_prob, fill_price, expected_pnl_per_share and the tier are written to arb_audit_log (Postgres). Before any user sees it.
A settler joins each detected row to its market's resolution outcome and computes realized_pnl_per_share = payoff − fill_price − entry_fee. Atomic. Idempotent.
No survivorship bias. Every detection — including the embarrassing ones — stays in the log. The aggregate you're reading above is unfiltered, just SUM and AVG.
What this actually proves
And, more honestly, what it doesn't.
When yes_ask + no_ask < $1.00, buying both sides guarantees the $1 payoff at resolution regardless of outcome. The 100% win rate isn't luck — it's math. They're rare and HFT-contested in practice, but the audit captures them all the same.
A specific strategy — buy the near-certain side in the final 0-120s of a 5m/15m market when Binance AND Chainlink both agree the underlying is > ~2σ past strike — wins ~95% in our log. The 5% tail loses big ($0.9/sh), and the sample is still under 500. We label it 'empirical edge' rather than 'guaranteed'.
The stable/stale tiers — the model claiming +EV against the Polymarket book — accumulate thousands of detections and roughly cancel out (~50% win, tiny negative drift). Polymarket maker bots appear well-calibrated for crypto Up/Down books. We surface them as research data, not as a signal.
What this page doesn't include
- Polygon gas (~$0.10–0.30 per trade) — the realised PnL above is before on-chain execution cost.
- Execution lag — the engine logs detection at scan time (every 30 s); a user clicking the row on the live page arrives later, when the book has often moved.
- Adverse selection on logical arbs — when one does fire, it's usually because HFT bots haven't marked the cheap side yet; by the time a 30 s detector emits it, the window may be closed.
Treat the "Realised / share" column above as a GENEROUS upper bound on what an executing user can capture.
Run your own signals against the same log.
Strategy Builder lets you compose entries / TPs / SLs and audit them with the same infrastructure. Free tier — no card required.