Recap

One trade, traced end to end

The shared language: one market, one set of artefacts, one explainable trade path from raw data to final outcome.

The practical goal

Any trader, engineer, or operator should be able to answer three questions quickly: what happened, why did it happen, and would we make the same decision again in replay? A single decision packet links registry, book hash, feature vector, fair-value score, execution estimate, risk approval, order IDs, fills, and final inventory state.

One trade, end to end

Registrytick 0.01 · fee 50 bpsStatebid 0.62 · ask 0.64Canonicalevent_id · book_hashSignalsimb 0.31 · vol 0.014Fair valuep_hat 0.67Executionq_exec 0.645Decisionedge 0.01 · BUY 2,000OMSORD-992 · filled 1,200Inventory+2,000 YES · avg 0.645Replayparity cleanMonitoringhealthy · canary on

Shared operating rule

\[ net\_edge = p\_hat - q\_exec - fees - uncertainty\_buffer \]

No stage may skip the one before it. That is what makes a trade explainable and safe enough to scale.

What a decision packet links

Registry record (market_id, tick_size, fee_bps)
Book hash at decision time (from the canonical event log)
Feature vector ID + signal_version
Fair-value score (p_hat, LB, UB, buffer, model_version)
Execution estimate (q_exec, slippage, P(fill), book_hash it was computed against)
Risk approval (all check IDs, votes, reason codes if any)
Order IDs (client_order_id, exchange ref, tx hash)
Fills (price, qty, ts, builder code)
Inventory delta (free, reserved, redeemable change)
Correlation_id binding everything above