Parity is not a claim, it is a number
One time source, a real matching engine instead of a slippage guess, and selection that is bit-exact against the simulator. The same discipline, pointed at trading.
Ask a quantitative developer why their backtest does not match live and you will get a shrug and a list: slippage, latency, fills, luck. It is the most common question in the field and almost nothing answers it honestly, because answering it requires giving up a convenience that every backtesting framework depends on.
Bar close plus a constant is an assumption wearing a result’s clothes
Almost every backtest in this market prices a fill as the bar’s closing price plus a slippage constant. It is fast, it is simple, and it silently decides the thing you were trying to measure.
A matching engine that works against historical quotes and ticks answers a different question: would this order have filled, and at what price? Not what a plausible fill looks like on average — whether this order, at this size, against the book as it actually stood, would have been done.
That is a harder engine to build and a slower one to run. Everything below follows from having it.
Consequence one: parity becomes achievable
If the simulator and the live path run the same execution logic against the same shape of data, you can compare them and expect the answer to be zero. Ours is $0.00 across thousands of trades at the platform level for a single backtest run.
That number is only meaningful because of what sits under it. One TimeSource abstraction with
two implementations: in backtest, time advances strictly on market-data timestamps; live, a wall
clock with a timer thread that advances timeout state machines even when no prints arrive. The
same execution-algorithm interface serves entry, stop-loss and end-of-day flatten — one contract,
three roles, so there is no second code path to drift.
Selection-side parity is checked separately and more strictly: bit-exact against the simulator, down to the model’s own intermediate values on a given bar.
Two different claims, and worth keeping apart. Bit-exact means the numbers are identical. Zero-delta at the platform level means the money came out the same.
Consequence two: execution findings become discoverable
This is the part that pays for the engine.
Patience pays on the way in and costs on the way out. An adaptive limit adds +6.45 percentage points on entry by capturing the spread — and costs −1.51 on exit, because once a stop has fired, price is by definition moving against you and waiting only widens the loss.
So: adaptive in, cross out. It is a small rule and it is worth real money, and it is invisible to anyone using a slippage constant — because a constant cannot distinguish an entry from an exit. It prices both the same by construction. The asymmetry is not hidden in that model; it is not representable in it.
85 trading days, January–May 2026. Backtest figures about the platform, not a track record: past performance does not indicate future results.
Consequence three: execution quality becomes a number
With real fills you can bin realised prices against the decision reference price by time of day and get a slippage surface. Execution quality stops being a feeling about the broker and becomes a measurement you can act on — including the finding that the cost is concentrated in a window you could simply avoid.
Why this is the same argument as the software one
We spend most of our writing here on verification in code: that generation is cheap now and checking is what remains expensive. This is the same claim, pointed at a different artefact.
A backtest is a generated assertion about the future. The slippage constant is the point where that assertion stops being checked and starts being assumed. You can make the assertion look excellent by choosing the constant well, and nothing in the process will tell you that you did.
Owning the matching engine is what converts the assertion back into something checkable. The number it produces can disagree with you — and did, repeatedly, which is how the entry/exit asymmetry surfaced at all.
That is the whole discipline: build the thing that is able to tell you that you are wrong, then report what it says. A parity figure you cannot fail is not evidence. A parity figure measured against a simulator that runs the same code, on the same clock, through the same fills, is.