PRE-REGISTERED · FORWARD ONLY · PUBLIC KILLS
A strategy has to prove it before it trades your money.
Quant strategies trade paper money live, under rules written down in advance with a decision date and a kill rule. Watch them compete. Nothing graduates to real money until it clears six gates over 90 trading days, repeats that in a second window, and is signed off.
SYSTEM DESIGN · ONE ENGINEER
One decision row feeds every surface.
Market data becomes features, features become a ranking, the ranking becomes a typed decision. That single row drives the feed, the paper books, the graduation gate and the broker. Nothing renders that didn't come through it.
per (ticker, date) feature cache
WHERE THE AI SITS
The AI here doesn't pick stocks. It labels the world so the math can.
We tried the other way; it lost, and we published that.
- Risk labelerblocks buys on fresh high-severity flags, for strategies that use the gate
- Retro narratornightly trade-level review, never on a request path
- Macro narratorwrites the macro narrative; the regime itself is computed in code
- Theme generatorsearch content, never a signal
- RankLightGBM ensemble over the feature substrate
- Gatedeterministic rules, pre-registered
- Size & exitfixed lots, percentile and stop exits
- Qualifysix gates over 90 trading days, then a second window
- LLM reviewer on every decisionnet-negative alpha, removed Jun 14
- Crypto LLM verdictsinverted, killed Aug 12
- LLM pickergraded daily against a quant control since Sep 11 — grades exist, no verdict yet
- Registrya decision date and a kill rule for every experiment
Generative AI never reaches an order without passing a pre-registered gate.
What I've learned so far
Four findings from the research store and the paper books. None of them are flattering.
- 01
The classic factors lost to doing nothing clever.
Nothing long-only that was tested beat simply buying every eligible name equally. The bar was never the S&P 500; it was the equal-weight universe, and that bar has held so far.
- 02
The model's top decile trailed its own middle.
Over a 39-day window this summer the top-ranked names returned less than the middle-ranked ones over the following 20 sessions, on 36 of 39 days. A five-hypothesis study ruled out sector, volatility, extension and data artifacts — the score itself inverts inside the top half. The gap faded across the window, so this describes one episode, not a permanent property.
- 03
Bitcoin trend: historical drawdown reduction, not forward validation.
A simulation of holding BTC above its 100-day moving average and cash otherwise, from 2014-07-20 through 2026-09-29, showed maximum drawdown of 64.9% versus 83.6% for BTC-hold before costs. At modeled costs of 0.6% per trade side, strategy drawdown was 69.3% and annualized return was 43.1%, below uncosted BTC-hold's 49.5%. Results varied by period, and the simulation assumes execution at the signal's final close. Book 42 is paper only; a late entry failed its original execution soak. No forward validation has been established.
- 04
We fired the AI stock-picker. Twice.
An LLM reviewer sat on every decision and produced net-negative alpha, so it was removed on June 14. The crypto LLM verdicts came out inverted — the buy bucket (646 calls) performed worst — and were killed on August 12. We also audited our own claims and found one of them false: the picks were being logged, but nothing was scoring them. The scorer now exists and runs daily. It has produced grades, not a verdict on any picker, so the honest answer today is that we do not know yet.
Nothing has graduated to live trading yet. That's the point of the gates.
How the discipline works
The same four steps for every strategy. Nothing runs indefinitely.
- 1
Backtest
Every policy idea replays offline against historical decisions and bars before it touches a book.
- 2
Pre-register
A survivor gets a registry row: the question, a decision date, and the rule that kills or promotes it.
- 3
Paper soak
It trades paper against live prices, after modeled costs.
- 4
Graduation read
On the date, the pre-committed gates decide — and a winner must replicate in a fresh window before real money.