Spicy Human / Tools / Regime Agent

Regime Agent

A trading research agent that has to earn the right to trade real money, and hasn't yet. It runs on paper, nightly, and grades itself against a bar it cannot move.

Two tiers: a sealed nightly search that only nominates, and forward paper races that make the claim A timeline runs left to right. The left part is history, sealed 180 days back, where the nightly search happens. Past the seal is unseen data, then today, then the forward paper races. Only the forward record feeds the graduation gate, and the gate is the only door to real money, which is still closed. Nightly search five strategy families · up to 600 trials · two hours over history sealed 180 days back can nominate. cannot claim. nominates Paper sleeves a champion and challengers each under a charter hashed before the race began history, sealed the seal unseen today forward, on paper The gate 42 counted days · deflated Sharpe at target drawdown under a fifth · vs. holding the index real money no path exists in the code
The screen is never the claim. Search produces nominations; only the forward paper record is allowed to say a strategy works.

What it does

Every night it searches five families of strategy against a slice of market history that is sealed 180 days back: regime detection with a hidden Markov model, trend following, mean reversion, breakout, and volatility targeting, which is the do-nothing control made tradable. The search can only nominate. The claim that a strategy works is made forward, by live paper sleeves racing under charters that were hashed before the race began, on data the search has never seen.

The book is capped, not the choice. Ten positions at most, no more than a third of them in one sector, no leverage. A wider list of instruments gives the signal more to choose from, not more to hold; more than a hundred are watched across sixteen sectors, and every one is re-checked for eligibility each cycle.

What comes out is a journal, not a ticker: a local dashboard and a small desktop window share one editorial record of what was tried, what was seated, what was voided and why.

How it works

The seal
History is cut 180 days back and the search never sees past the cut. Anything found in the sealed slice has to prove itself on the days after it.
Charters
Before a challenger is seated it gets a charter: what it trades, how it sizes, when it is voided. The charter is hashed first, so nobody can quietly rewrite the rules after a good week.
The race
A champion and its challengers trade the same paper days side by side. A challenger has to beat the champion over a full ninety traded days to take the seat.
The gate
Forty-two counted paper days, a deflated Sharpe ratio at the target, drawdown under a fifth, all measured against simply holding the index. The count restarts from zero whenever the gate is found to have been softer than it should be, and it has restarted more than once.
Eligibility
Every instrument is checked each cycle and every exclusion gets a written reason, because an asset silently missing from a cycle looks exactly like one the signal declined to buy. That ambiguity once hid a data fault for twelve days.
The lock
The nightly cycle, the search and the weekly review queue behind one process lock so two jobs never write the same ledger at once.

Where it stands

Paper only, by rule, restated more than once. The graduation clock is running and has been reset when a gate turned out to admit variants it should not have. A second instance runs the same code with a higher volatility target on its own ledger, and counts toward nothing.

Not public. It runs on one machine and trades no one's money, including its maker's.