The pyramid: narrow trusted core at the top, widening rings of less-trusted machinery below. LLMs propose; code disposes. Nothing below the green line can move money.
GATE — pure code, zero LLM
Pass/fail thresholds: ≥100 trades, exp ≥1.0t, PF ≥1.3, maxDD ≤$1500, walk-forward OOS Sharpe ≥0.5. Stamps every verdict. Unbribable.
RISK & EXECUTION — pure code
Topstep 50k eval rules, % risk/trade sizing, kill-switch, Tradovate adapter (paper first, 20 clean days gate). LLMs never touch orders or sizing. Mostly UNBUILT.
DETERMINISTIC ENGINES
Backtest: signals on bar close, fills next bar, ≥1 tick slippage, real fees — never mid-price. Signal judge (Nautilus), walk-forward splitter, metrics. Families are pure signal functions — one file each in backtest/families/, registered in FAMILIES.
ADVISORY LLM RING — local models, $0/day
Feeder proposes hypotheses (graveyard-aware) · Chair (ministral-3-14b-reasoning) decides tweak/abandon/new, exploration floor enforced in code · Council (gemma-26b, glm-4.7-flash, nemotron-3-nano) votes to gate new hypotheses · Coder (glm-4.7-flash) drafts new families — static checks + human review before registration. Every decision journaled with per-model attribution.
DATA
Databento GLBX.MDP3 → MES 1m bars, 1.4M bars (Aug 2022 → today), parquet by month. $6.45 spent all-in. License withholds newest ~8h.
The loop: feeder proposes → backtest → chair decides → council gates new families → gate stamps verdict → graveyard/notebook → feeder reads graveyard → repeat. 20-tweak cap and $1/day budget enforced by the harness, not trusted to LLMs.