The method

Every discipline on this page exists because I got something wrong first. None of them are theory. Each is a rule I adopted after a specific failure made the cost of not having it concrete — which is the only kind of methodology I trust. The kind with a scar behind it.

Here's how I work, and why.

Read from disk before reasoning. Code you didn't just read is code you're guessing about. It's remarkable how often a confident belief about what a system does turns out to describe what it did three versions ago. So before I reason about what happened, I go read what's actually there — the records, the raw log, the real state — and reason from that, not from my memory of how I think I built it. More than once the disk has told me I was wrong about my own system. The day I found I had two strategies' fill quality labeled exactly backwards, it was only because I stopped trusting my read and pulled the actual data.

Built isn't deployed, and deployed isn't live. Three different states that are easy to blur into one, and the blur is where people fool themselves. Something can be written and not running; running in simulation and not touching real money. "It's built" is not "it works," and "it works in sim" is not "it works." Nothing here gets called done until an artifact shows it done, and I keep the three states named and separate so I can't quietly promote one into another without noticing.

Lock the prediction before the result exists. The single most important rule I have. Before a test runs, I write down what I expect and the threshold that would make me walk away — then timestamp and seal it, before there's any result to see. It's the only real defense against the most natural move there is: seeing how something turned out, then quietly adjusting what you claim you predicted. Goalposts hashed in place before kickoff can't move. It's why the live fill experiment I keep mentioning will have its walk-away threshold written into the record before the first real fill — never after.

Fault is not flat. When a system can't read what's happening — a data gap, a stale value, an ambiguous state — the dangerous move is to treat "no data" as "zero" and carry on. An unknown isn't a safe neutral; it's an unknown, and it gets its own named state. A system that silently reads absence as zero is one that will eventually act with full confidence on nothing at all. So "not sure" halts. It never defaults to "keep going."

Negatives and refusals count. A test that refutes my idea is a result, not a broken test. A search that turns up nothing worth keeping counts as an honest answer, not a wasted run. Most of this industry reports only the hits and lets the misses quietly disappear — which is exactly what makes the hits impossible to trust. Here a refuted thesis and an empty result are recorded with the same weight as a success, because a record that only remembers its wins isn't a record. It's a highlight reel.

Nothing is ever overwritten. When I'm wrong, the correction is a new entry that points back at the mistake — and the mistake stays exactly where it was. I don't edit the past into looking better than it was. An honesty you can only see after it's been cleaned up isn't one you can check. The wrong call has to stay visible, with the correction beside it, or the claim to honesty is just words.

The system proposes; a gauntlet and a human dispose. The analysis layer can nominate a candidate — a pattern, a parameter, a correction — but nominating is not promoting. Nothing a machine suggests reaches the live system on its own say-so. It has to survive a validation process built specifically to try to kill it, and then a human has to sign off. Many candidates get nominated. Very few survive. That's the design, not a disappointment: the bar to change the system is deliberately high, and no single component — me included — gets to wave something through.

None of this has produced a proven edge yet, and I want to be exact about that, because the discipline is not the same as the result. What these rules buy you isn't a guarantee that the system works. It's the guarantee that when it does or doesn't, you'll know which — from the record, not from my word for it. That's the thing I'm actually offering here. Everything else is downstream of it.