I was sure which of my strategies had honest numbers. I had it exactly backwards.

Every backtest is a negotiation with one question you'd rather not ask: when your stop-loss triggers, what price do you actually get filled at?

It sounds like a detail. For a system that takes a lot of stops, it's the whole ballgame. Assume your stops fill exactly at the level you set them, and every losing trade looks as small as it possibly could — a clean, capped loss. Assume they fill a few ticks worse, the way real stops in a fast market usually do, and a stop-heavy strategy that looked profitable can flip negative. Same trades, same signals; the only thing that changed is an assumption about fills, and the assumption moved the P&L across zero.

So when I recently compared two of my strategies — call them the fixed-stop version and the volatility-scaled-stop version — the first thing I did was decide which one's recorded numbers I could trust. I concluded, with confidence, that the fixed-stop strategy was the honest one: its stops filled right at the level, clean and consistent. The vol-scaled one I decided was the flattered one — its fills looked suspiciously generous. I was confident enough that I built the next stage of analysis on top of that judgment. I even locked a hypothesis into the record that depended on the fixed-stop strategy being the clean, trustworthy baseline.

Then I did the thing that makes this worth writing about: instead of trusting my read, I had the system pull the actual fill data off the disk — the real recorded depth of every single stop — and check.

It was backwards. Both labels. Exactly reversed.

The fixed-stop strategy's stops filled at precisely the level, every time — to the point that its worst losses were all identical, a perfectly capped tail. That is not the signature of honesty. It's the signature of an idealization: a fill model that quietly assumes zero slippage, the single most optimistic assumption available. Its clean numbers weren't clean because the strategy was disciplined; they were clean because the fills were a best case that can't actually be achieved — real stops are never better than the level. That strategy was the most flattered thing in my book, and it was the one I'd called honest.

The vol-scaled strategy was the reverse. Its stops triggered on a bar closing past the level, so its fills landed well beyond it — its recorded losses ran deeper than the stop distance, not shallower. It hadn't been getting generous fills. It had been getting punished, and its numbers were pessimistic, not optimistic. The one I'd flagged as too-good-to-be-true was the one carrying the honest, painful fills.

I had ranked them by exactly the wrong dimension, and my confidence had nothing to do with the truth. Confidence rarely does.

Here's the part I actually want you to notice, because it's the point — and it isn't the mistake. Everyone makes mistakes. It's what happened to the mistake.

I didn't quietly re-label the two strategies and move on with the corrected version, the way you'd never know it had happened. The original judgment stayed exactly where it was, byte-for-byte, in the record. The correction was written as a new entry pointing back at it: here is what I claimed, here is what the disk actually said, here is the reversal. And the hypothesis I'd locked on the wrong premise didn't get silently rebuilt on the right one — it got flagged, in the open, as resting on a refuted assumption and needing to be redone. The wrong thing is still visible. It has to be. An honesty you can only see after it's been cleaned up isn't one you can check.

This was not a special occasion — it's the design goal. The system is built to catch claims that don't survive the data, mine among them, and to log each reversal as a correction that points back at what it overturned: still on the record, never quietly cleaned up. This one sits there alongside the others from the same stretch, and if I handed you the log you could read every one. I'm not telling you this to perform humility. I'm telling you because it's the whole point: I don't build systems that are right. I build systems that catch me being wrong and won't let me hide it — from you or from myself. The measure isn't how often the first guess is correct. It's whether the wrong guesses survive contact with the record. They don't.

And there's a punchline that matters for everything I'll write here later. That backwards fill judgment wasn't a footnote. It turned out that the entire question of whether any of my strategies has a real, cost-surviving edge collapses down to this one thing — how the stops actually fill in live trading — which is a question no simulation can answer, including mine. So a good part of what comes next is going and measuring it for real, with the threshold for whether it's even worth the money written down and hashed before the measurement, so I can't move the goalposts once I've seen the answer.

That's what this is. Not a track record. A method you can audit, run by someone who will show you the reversals with the timestamps still on them.