Bounded AI · Defined risk · Auditable abstention

An AI that must earn permission to trade.

Evidence Governor separates market detection, option construction, AI criticism, and execution authority. The language model can review evidence. It cannot invent a contract, change size, or reach an order tool.

Final total equity
official $100,000 baseline
Paper P&L
total account equity basis
Round trips
4 wins · 3 losses
Final state
0 positions · 0 open orders

Authority map

One pipeline.
Five hard boundaries.

The agent is autonomous only inside a deterministic safety envelope. Each component has one job, and the AI has less authority than the risk governor.

30 real model calls

Inspect every decision.

Choose a historical case to see exactly what the AI selected, which evidence IDs it cited, why four answers were rejected, and what the outcome revealed only afterward.

Loading cases…

Outcome audit

Less loss is not an edge.

The AI improved the paired result by abstaining or switching in four economically different cases. But every available candidate in the accepted sample was non-positive, and the AI still lost money.

Cumulative net P&L across accepted reviews

AI governorDeterministic baseline
X-axis: accepted review index · Y-axis: cumulative net P&L (USD). Source: audited exposed historical replay, 26 accepted cases from the 30-case 2026-09-01 stratified model run.

Bounded execution

Positive Paper P&L, without an edge claim.

The competition workflow executed, but the system keeps simulated fills, data quality, statistical promotion, and order authority as separate claims.

Paper workflow

COMPLETE

Seven authorized multi-leg option round trips completed and the account finished flat at $100,125.91.

paper_evidence=true · live_claim=false

Deterministic risk

ENFORCED

Per-trade risk targeted 0.10% of current equity; daily loss was capped at 0.25% and $250; one spread at a time.

ai_sizing_authority=false

Market-data limit

INDICATIVE

The free options feed was Indicative, not OPRA. Simulated Paper fills do not prove live executability.

strategy_efficacy_claim=false

Order boundary

SCOPED

The runtime could submit only code-locked Paper candidates and exact owned-leg closes. This public demo has no account or order access.

final_positions=0 · open_orders=0