What are the kill gates?

Pre-registered conditions, published before any results are in, that decide the fate of the forward book. If the strategy underperforms the market by enough, or the model's curation fails to beat a dumb mechanical baseline over a meaningful sample, the book gets de-scaled — or shut down entirely and reverted to passive. They are the commitment that this experiment ends if it doesn't work.

Why publish them in advance?

Because a kill criterion written after you see the results isn't a kill criterion — it's an excuse. Fixing the conditions up front, in public, is what makes the eventual verdict trustworthy. The gates are tighten-only: they can be made stricter, never loosened, once outcomes start landing.

When are they checked?

Quarterly. Most gates need at least twenty closed theme-bets before they can return a real verdict — until then they read as pending or insufficient data, which is the honest state of a book that has just opened. The first review that can return a verdict is 2027-01-08; 2026-10-01 is a checkpoint that publishes the instruments as they stand.

Kill gates · first verdict-capable review 2027-01-08 · checkpoint 2026-10-01

The conditions that end this.

Forward kill gates, pre-registered and reviewed quarterly from 2026-10-01. Gates tighten-only — a criterion can be made stricter, never looser, after results are in.

K1 insufficient data n = 0 closed

Trailing-6mo book < SPY − 5pp → de-scale the sleeve one step (90→60→30%).
K2 pending n = 0 closed

At n≥20 closed theme-bets: mean bet excess vs matched-window SPY < 0 → sleeve floored at 30%; < −5pp → thesis declared dead end-to-end, published, and the book reverts to passive pending a new mandate.
K3 pending n = 0 closed

The AI's curation must beat the raw mechanical cluster baseline (shadow-logged for every bet) at n≥20 — else curation is decoration and gets removed.

Curation instrument · not a gate

What the model turned down, scored anyway.

Every cluster the radar triggers is decided in public — shaped, rejected, or parked — and most never enter the book. Since 2 September 2026 those decisions are scored on price alone: the equal-weight cluster from the radar's closes, 60 sessions on (or the latest close while the window is open), against the S&P 500 over the same dates. A rejection is right when the cluster trails the index. This reads the curation long before the kill gates can; it decides nothing.

Rejected clusters n = 39 · 0 at 60 sessions: Cluster vs index: -1.7 pp mean, 36% beat it. Rejecting was right 64% of the time.
Shaped, no free slot n = 38 · 0 at 60 sessions: Cluster vs index: -0.3 pp mean, 50% beat it. Named cores vs equal weight: +1.4 pp.
Parked n = 11 · 0 at 60 sessions: Cluster vs index: -1.2 pp mean, 55% beat it.

Why this page exists

A strategy with no off-switch is a religion.

The forward book is an experiment; no performance claim is made. Backtested mechanical variants did not clear the benchmark, which is precisely why the terms of failure are fixed in advance: if the book trails the market by a stated margin, or curation fails to beat an equal-weight baseline over a real sample, it de-scales or reverts to passive on the published criteria rather than on judgement. The results feed the public track record; the forecasts are the calls these gates judge.

The model's bets This week's radar Track record

Common questions

What are the kill gates?: Pre-registered conditions, published before any results are in, that decide the fate of the forward book. If the strategy underperforms the market by enough, or the model's curation fails to beat a dumb mechanical baseline over a meaningful sample, the book gets de-scaled — or shut down entirely and reverted to passive. They are the commitment that this experiment ends if it doesn't work.
Why publish them in advance?: Because a kill criterion written after you see the results isn't a kill criterion — it's an excuse. Fixing the conditions up front, in public, is what makes the eventual verdict trustworthy. The gates are tighten-only: they can be made stricter, never loosened, once outcomes start landing.
When are they checked?: Quarterly. Most gates need at least twenty closed theme-bets before they can return a real verdict — until then they read as pending or insufficient data, which is the honest state of a book that has just opened. The first review that can return a verdict is 2027-01-08; 2026-10-01 is a checkpoint that publishes the instruments as they stand.