A one-person lab building custom neural networks to predict short-term stock moves. Every test that decides anything is written down before it runs; every verdict is posted unedited, losses included.
Where it stands today. Lookout, the first model, ranks the market well and hasn’t made money yet. The program now aims to grow the account faster than an index fund, after every cost.
A machine-kept record, hosted at gradedtape.com. Every morning, code scores the latest session’s signals against rules fixed in advance and appends them to the tape. The one correction, on Aug 21, 2026, is documented there. Signals are grouped into bands by share price. Three of four price bands sit below zero, and they stay on the page.
The winners are listed, the losers are gone, and the strategy changed the month it stopped working.
Exhibit B is a track record assembled the usual way. The shaded months were left out of the brochure.
Each experiment’s write-up names which stocks, how long the trade holds, what counts as a win, and how it’s scored. The experiment runs once, on data the model has never seen. The verdict is posted the same way every time, no matter what it says.
The bar: beat the incumbent by more than +0.02 on 71 matched walk-forward windows, interval clear of zero.
Verdict: LOST. −0.0059, interval −0.0086 to −0.0032. Retired in writing.
The bar: beat the best single model, interval clear of zero.
Verdict: PASSED. +0.0287, interval +0.0231 to +0.0350. Adopted.
The bar: beat the incumbent by more than +0.02, interval clear of zero.
Verdict: MISSED. +0.0012, interval 0.0000 to +0.0024, short of the bar by more than an order of magnitude.
Three of the nine tests from the July campaign. All nine, with their bars and verdicts, are in the letters. A win only counts against the full count of tries.
Every claim rests on data the model never trained on. Scores on the training data itself are shown only as a health check, never as a result.
Every verdict in the letters comes with its range of uncertainty and the count behind it, so a real effect can be told from luck.
When to enter, when to exit, how long to hold and how much to stake are all locked before any signal goes out. Changing a rule starts a new experiment; it never edits an old one.
Grading runs unattended each morning and publishes what it says. There is no editorial step between the result and the page.
The writing is open today, and every verdict stays free. Anything that follows a model’s signals waits until a model passes its sealed test.
Every letter’s story and every verdict, plus The Graded Tape. Read the misses before you pay for anything.
FreeEverything free, plus a “How I built it” section at the end of each post, the working papers that close each arc from arc 2 on, and the monthly research log. This is what funds the lab.
$20 / mo · $200 / yrEverything in Letters, your rate locked for as long as you subscribe, and first claim on anything that opens later.
$400 / yrOnly if a model passes its sealed test, a single final reading on data it has never touched. Priced then, in public, against the evidence. No date.
Not yetStart with the record and the failures, then read the letters.