how a signal earns the right to change behavior — the full methodology, in public
skeleton shipped 2026-07-02 — the trust ladder is complete; remaining sections are being written from the committed docs and will land dated
The ladder
Every lane runs the same engine, and every signal climbs the same ladder. A number on this site never carries more authority than the rung it stands on.
Every fire is recorded with its features frozen at fire time. No claims yet — counting.
A feature needs 150+ graded fires before its correlation with outcomes is even read aloud.
The read must hold its direction across three consecutive checkpoints. Trust in the measurement — not yet permission to act.
Candidate thresholds are backtested on a chronological split: fit on the past, judged only out-of-sample. To pass: OOS lift, Bonferroni-corrected p<0.05, OOS n≥30, and an OOS span long enough to cross regimes — 72 hours for the meme lane, 20 trading sessions for equities. Per-lane rails by design.
A licensed signal still doesn't act. It arms in shadow: two consecutive confirms to arm — slow to arm — and any breach disarms instantly. Asymmetric on purpose. Every armed decision is self-graded out-of-sample in an arm ledger.
Behavior changes only when a confirmed-licensed signal is armed AND the master switch is flipped — a manual, witnessed act. Auto-promotion must earn its own license: three-plus armed episodes graded majority-correct with zero false arms. Every flip posts its full evidence chain. The kill switch outranks everything.
SHADOW = the machine decides, records, and grades itself. LIVE = the machine changes what you see. The two are never blurred.
In progress