Live–backtest parity: the same system, run twice
Most strategies die leaving the lab. Our live trading and backtests are the same system — with identical, auditable results.
A beautiful backtest is easy. A live book that matches it is not. The usual failure is not a bad idea. It is a different system in production: another data path, another fill model, another chance to condition on information the backtest never had.
Live–backtest parity is the requirement that the agent trading founder capital is the same program as the agent you can replay. Same inputs, same decision rule, same auditable trail.
What parity means here
- Point-in-time data only. No restated fundamentals sneaking into yesterday’s decision.
- The live agent and the replay share a configuration that was locked before deployment.
- Orders, holdings, and timestamps can be compared trade for trade.
- When a metric cannot be loaded from the brokerage path, the page stays empty rather than filling in a zero.
That last point is easy to get wrong. A missing return is not 0%. A failed embed is not a quiet holiday. Parity includes how failure is shown.
Why it matters for Sophon
Continual-learning systems are especially exposed to leakage. If the update step can see the future, the live agent will never catch the backtest. Sophon-3 is built so the learning step and the trading step consume the same clock.
Simulation and live trading reproduce the same decisions trade for trade.
The annual tables and growth charts on each agent page are there so anyone can watch that claim in public. Methods paper to follow. Until then, the live book is the argument.