Alpha is a theorem.
We prove it live.

H₀ = 0²⁵⁶ — the genesis block

Every signal is appended to a hash chain. Each entry commits to the entire history before it: change one byte of any past trade and every subsequent hash breaks. The track record cannot be rewritten — only extended.

ER → 1: straight-line trend · ER → 0: pure noise

Before voting, the brain measures how efficiently price has travelled. ER > 0.35 declares a trend regime (momentum skills weighted 1.5×), otherwise range (mean-reversion 1.5×). If volatility exceeds 2.2× its median, the regime is chaos and every vote is discounted to 0.6×.

11 skills × parameter variants vote; wᵣ = regime weight

Eleven independent strategies — momentum, breakout, reversion, VWAP, multi-timeframe — vote on every bar. A trade only exists when at least 55% of regime-weighted votes agree on direction. One indicator's noise can never move the book.

m_mem ∈ {0, 0.5, 1, 1.2} · m_coach ∈ [0.7, 1.2]

Raw ensemble confidence is multiplied by the agent's memory of its own recent performance (5 straight losses → trading halted) and its LLM coach's modifier. The result must clear the agent's temperament threshold — a calm agent needs 0.50, an aggressive one only 0.25.

E = live equity · ρ ∈ {5%, 10%, 15%} by temperament

Position size scales with the agent's live equity — winners compound, losers shrink. Stops sit at 2× the 14-bar Average True Range, targets at 4×: every trade is structured at a minimum 2:1 reward-to-risk before entry.

S = return − ½·maxDD, scored on unseen data only

Every night each agent's parameters are re-optimized on the first 70% of recent data and selected purely on the held-out 30% — curve-fitting is structurally impossible to reward. If a rival skill scores 1.5× better on the agent's own market, the agent evolves: the weak skill is replaced. Natural selection, nightly.

11 deterministic strategies vote on real OHLC bars, weighted by the detected market regime. ≥55% directional agreement required.

The agent reads its own ledger. Losing streaks throttle confidence to 0.5× or halt trading entirely; proven streaks earn a 1.2× boost.

Three parallel LLM analysts — technical, macro, sentiment — each cast an independent vote. 2 of 3 must agree with the ensemble.

A final risk officer sees everything the layers produced and holds absolute veto power: APPROVE, REDUCE size, or kill the trade.

Before any vote is cast, the market is classified as trend, range or chaos using the Kaufman efficiency ratio and volatility percentile. Momentum skills are weighted 1.5× in trends, mean-reversion 1.5× in ranges — everything is discounted to 0.6× in chaos. The right weapon for the weather, automatically.

Each agent's specialty skill is grid-searched on its own market: parameters train on the first 70% of recent data and are selected purely on the held-out 30%. Curve-fitting is structurally impossible to reward — only parameters that survive unseen data go live.

All eleven skills are backtested on every agent's market and ranked. An agent stuck in the bottom 40% whose best rival scores 1.5× better gets its specialty replaced — natural selection over strategies. The roster never changes; the strategies inside it do.

A language model reviews each agent's closed trades, names the strongest recurring pattern — good or bad — and writes a confidence multiplier from 0.7× to 1.2× into the brain. Losing patterns get throttled before they compound; clean execution earns more size.