Safety model

This page explains why you can trust what the dashboard is doing — the guardrails are built into the structure of the system, not just promised in a policy. If you only read one thing about safety, read the five rules below.

The five invariants

These are enforced by the code and by an automated test suite that fails the build if any of them is broken.

1. No AI in risk, sizing, or execution

Stages 7 (risk & sizing), 9 (execution), and 11 (monitoring) are pure code. The modules that do this work are physically forbidden from importing the AI/LLM code — a build test scans them and fails if they ever do. So an AI can suggest a trade, but it can never size it, route it, or compute your P&L.

The same wall applies to the deterministic feature layer added in M3 — the indicators, ATR/level anchors, and the price forecast that ground the agents. That math is code-only, and the risk and execution modules are forbidden from importing it too. So even the forecast is evidence the AI reasons with; it can never reach the code that sizes or routes a trade. In fact the trade's entry, stop, and target are now derived by code from the ATR/level anchors, not chosen by the model — and a thesis with broken geometry (stop on the wrong side, or too far from the anchor) is rejected before it can reach you.

2. The human approval gate cannot be bypassed

Order execution requires a signature that can only be created inside the human approval module. There is no code path to the broker that skips it. Trying to place an order without a valid signature is not a policy violation you could override — it is a hard error. No order leaves without a human approving it.

3. Execution and audit are append-only and idempotent

Every order attempt is written to a hash-chained audit log before the broker is called, and re-submitting the same order is a no-op rather than a double-trade. Entry and exit use the same audited path. The record cannot be silently rewritten.

4. The kill switch is human-wired and out-of-band

The emergency stop (stage 10) works independently of the rest of the pipeline. One click flattens positions and blocks new orders, and it does not depend on the AI or the normal run flow being healthy. See Portfolio & kill switch.

5. Paper mode is the default and always visible

The system starts in paper (simulation) mode, and the orange PAPER badge is shown on every screen. Live trading is never silently on. In this build, live trading is not wired at all.

What this means for you

• You are always the final decision-maker. The AI cannot trade around you.
• A risk check you see is real — it was computed by code the AI can't touch.
• The emergency stop will work even if something else is misbehaving.
• You cannot accidentally trade real money from this dashboard.

What is not covered

Real brokerage connection, KYC, funding, and regulatory compliance are deliberately out of scope and human-owned. This tool is decision-support and paper-trading first — a place to research, propose, review, and rehearse the discipline of the approval gate, not an autonomous live trader.