🤖 → exchange · 🤖 → 🤖 · 🎯 → exchange · 🎯 → 🤖. Rows are origins (oracles + visitors), columns are targets. Cells coloured by RTT (green < 100 ms, amber < 250 ms, red ≥ 250 ms, dash = unreachable / not measured).

🎯 → 🎯 and 🤖 → 🎯 are intentionally omitted — browsers behind NAT have no reachable inbound port, so those legs would require a TURN/signalling relay (future work). The 🎯 → 🤖 column lights up only when the 🤖 node has a publicly-reachable edge_pinger on 9876.

Pulls every available oracle TWAP in parallel — Pyth (hermes.pyth.network), Binance kline TWAP, Coinbase candle TWAP, DexScreener h1, CoinGecko 24h, and our own cross-venue TWAP (from SPREADS.duckdb venue_snaps) — and compares each against every other oracle PLUS the live consolidated best bid/ask PLUS the most recent actual trade prints from Binance + Coinbase. Surfaces every dislocation above a configurable bps threshold.

Persistent WebSocket subscriptions to Binance (100 ms depth20), Bybit (orderbook.50), OKX (books5), and Coinbase (level2 incremental) — books updated on the venue's own publish cadence, no REST polling. Aggregated across venues for the selected pair. The detector-stats strip below shows what fraction of candidate signals die at each gate (time-alignment, fee threshold, etc.) — the explicit "is this a real tradeable arb at wire speed?" check.

A live in-RAM order-book replay: every tick_seconds we take a fresh snapshot of all venues, compute intra-tick bid/ask and high/low across the rolling window, run find_arbs() with DOM-aware sizing (CEX ladder depth + AMM-curve capacity), apply round-trip taker fees, and book the net PnL straight to cash. No DuckDB on the hot path — the snapshot ring is pure RAM so the strategy tick lands in single-digit ms.

Live allocator: routes capital to the venues with the most persistent, contemporaneous (time-skew < 3 s) edges over the selected window. Updates with the snapshot history that the 15-second background scheduler writes to data/SPREADS.duckdb.

Fees applied to every arb's net PnL come from the venue's fee schedule at the tier appropriate for cumulative 30-day filled USD. Cold-start = $0 = base tier. As the live engine writes fills via volume_tracker.record_fill(), volume rolls past each min_vol_usd threshold and the next-tier rate auto-applies.
Mode: Executed (real fills) Assumed (all opportunities)   Refresh

Per (pair, buy-venue, sell-venue) edges seen in the selected window — sorted by hit count. This is the same window that drives the allocator above.

Implied cross rates (mid_A / mid_B) on each venue that quotes both. Sorted by spread between cheapest and richest venue. Snapshot is fresh on every refresh.