Cross-venue
vol-arb on
DeepBook Predict.

What SVX is

A single-operator, fully-automated cross-venue volatility-arbitrage bot for the Sui Overflow DeepBook Predict track. It uses Predict's SVI surface as a pricing brain, takes the opposing side on Polymarket when the implied probabilities disagree, and samples realized vol from Hyperliquid's BTC perp to gate the expiry-convergence strategy. The delta-hedge leg was built, exercised on mainnet, and disabled after the 2026-07 audit (it sized delta at the oracle expiry). Three venues, one bot.

The dashboard you're on is a read-only window onto the operator's live bot. There is no signup, no wallet-connect, no deposit — the single-operator architecture is intentional and is what keeps SVX out of securities-law territory until the post-audit vault phase.

1. 1

   Build the pricing brain

   LIVE

   Back-solve Predict's implied vol from the on-chain SVI surface, reprice every strike at any expiry, and run an arbitrage-free checker (butterfly violations) on the live surface. The /surface page streams it in real time.

2. 2

   Trade the surface against Polymarket — with real money

   LIVE · REAL MONEY

   When Predict and Polymarket disagree, buy the cheap side on the Polygon-mainnet CLOB. Real fills, UMA settlement detection, on-chain auto-redeem, and a wallet-vs-ledger reconciliation invariant that pauses trading on unexplained cent-level drift.

3. 3

   Kill what real money disproved

   CUT — POST-MORTEM

   The IV−RV perp strategy paid $29.12 in fees for −$1.80 of direction PnL over 5,219 fills (a perp has no vega) — reconciled to the cent and hard-disabled in code. The delta hedge and the margin-lever signal got the same treatment: measured, documented, switched off. The post-mortems live on their pages.

4. 4

   Measure the SVI feeder itself

   THE FINDING

   The brief calls this bot 'a live stress test of the SVI feeder' — so we ran the test. Against every recorded oracle settlement, Predict's favorite is well-calibrated above 90¢ but systematically UNDERCONFIDENT below it — and the gap concentrates exactly where Polymarket disagrees:

5. 5

   Ship the strategies that finding implies

   LIVE ON TESTNET · MAINNET-DAY-ONE

   Divergence-mint (venues disagree ≥8pp: 94% win / +11.9% ROI May, 93.5% / +11.5% July) and calibration-harvest (every other favorite below 90¢: 90.6% / +14.2%) — two disjoint bands of the same edge, both minting live dUSDC via predict::mint, both reproducible from this bot's own ledger via GET /backtest. Predict Sui-mainnet launch day is an address swap and one config flip.

6. 6

   Pressure-test the brief's vault ideas against the same data

   SIMULATED — ONE YES, ONE NO

   The range ladder works — but only at σ/2 widths (+10.1% over 104 replayed oracles; ATM rung +29%), and svx mint-ladder mints the winner live via predict::mint_range. PLP + crash-insurance does NOT work on today's surface (PLP realizes ~0.7%/yr as the counterparty to the calibration edge; 1.5σ insurance costs ~2× its realized payout) — published with the numbers, because the NO is the same smile diagnosis from the house's side.