Notes from the desk.
How the debate works, how the models learn from every session, and the small design choices that separate a research desk from a signal group.
Self-improving trading models: a closed loop that only keeps measured gains
A failed replay hands the model back to the agents that compiled it, with every trigger's outcome. They diagnose, edit, re-validate — and the change is kept only if the numbers improved. Method, a worked case, limits.
Read post →Backtesting with AI, honestly: what the validator counts and what it refuses to
What counts as a trigger, how outcomes are scored, why an unfilled setup is not a loss, and how the P&L card sizes contracts so a 1-tick stop cannot post a fantasy trade.
8 min read →Why Sentio doesn't send you signals
The single design choice that separates a decision-support desk from every Discord group and Telegram channel — and why the CFTC agrees.
5 min read →The four honest failure modes
Losses 3–5 of the day. The revenge trade at 14:47. Size creep on a green day. Trading the wrong session. These are the trades that kill accounts.
6 min read →How the discipline layer works
From a rule written in plain English at 9am to a gate enforced in code at 2:47pm — the plumbing that keeps you honest with yourself.
7 min read →From a paragraph to a live scanner in ten minutes
A walkthrough of the model builder — describe a setup in plain English, watch it compile, replay-validate against 3000 bars, activate.
8 min read →