A strategy has to prove it before it trades your money.

• Risk labeler → flags observable risk events; strategies that enable the risk-flag gate refuse buys on fresh high-severity flags
• Retro narrator → nightly trade-level review, never on a request path
• Macro narrator → writes the macro narrative; the regime itself is computed deterministically
• Theme generator → search content, never a signal

• Rank → LightGBM ensemble over the feature substrate
• Gate → deterministic rules, pre-registered
• Size & exit → rule-based sizing, rank-percentile and stop exits
• Qualify → the §19 card, then a second window, before real money

• LLM reviewer on every decision → net-negative alpha, removed 2026-06-14
• Crypto LLM verdicts → inverted, killed 2026-08-12
• LLM picker → forward-graded daily against a quant control since 2026-09-11; grades exist, a verdict does not yet
• Registry → a decision date and a kill rule for every A/B comparison

Nothing long-only that was tested beat simply buying every eligible name equally on the long-history research store. The bar was never the S&P 500; it was the equal-weight eligible universe, and that bar has held so far.

Over a 39-trading-day window the top-ranked decile returned less than the middle-ranked names over the following 20 sessions, on 36 of 39 days. A five-hypothesis study ruled out sector, volatility, extension, single names and data artifacts: the score itself inverts inside the top half. The gap faded across the window, and overlapping returns leave roughly two independent observations — a well-checked description of one episode, not a statistically tested permanent property.

The rule holds Bitcoin above its 100-day moving average and cash otherwise. In a historical simulation from 2014-07-20 through 2026-09-29, maximum drawdown was 64.9% versus 83.6% for BTC-hold before costs. With modeled costs of 0.6% per trade side, strategy drawdown was 69.3% and annualized return was 43.1%, versus 49.5% for uncosted BTC-hold. Drawdown was not lower in every period. The simulation assumes fills at the same final close used to calculate the signal; it does not establish executable live prices. Book 42 is paper only, and its original execution soak failed because of a late entry. These historical results do not constitute forward validation.

An LLM reviewer sat on every decision and produced net-negative alpha, so it was removed on 2026-06-14. The crypto LLM verdicts came out inverted — the “buy” bucket (646 calls) performed worst — and were killed on 2026-08-12. A self-audit on 2026-09-11 then found one of our own claims 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.