A record before an action.
The interface is one part of a system designed to make changes and failures traceable. These are operating properties, not evidence of an investment edge.
Every release has a way back.
Each release lives in its own directory and is kept intact. Deployment records the old and new targets. A temporary symbolic link is atomically renamed over the public link, so a request sees a complete release. Rollback points that link back to the recorded previous release. Frontend changes do not restart the trading system.
Permission belongs to an exact protocol.
A protocol digest identifies the files covered by trading authorization. Changing a hashed file, even a comment in one, changes that identity and invalidates the authorization. The next version must be explicitly authorized rather than inheriting permission accidentally.
Record first. Reconcile before retrying.
Before asking the broker to change an order or position, the system writes its intent into a durable journal. Transaction keys identify requests so the same instruction is not treated as a new one on retry. If a response is ambiguous, the system checks broker transactions and positions before attempting anything else. A lost response is not proof that an order failed.
The observer cannot place an order.
A separate observer reads broker health, positions, transactions and account history. Its broker interface exposes read operations rather than order mutation. It can keep reporting when the forecasting process is stopped; seeing a connected account does not imply permission to trade.
A narrow public window.
The public API opens SQLite read-only, accepts only named datasets, and bounds the number of rows returned. Private order journals are not on its allowlist. Public account snapshots show balance, equity and open P/L in the account’s currency, alongside exposure percentages. Private account identifiers and order journals remain excluded.
What this does not prove
These safeguards make the research auditable. They do not eliminate execution risk, make model scores calibrated probabilities, or establish profitable trading. The experiments and failures address those empirical questions.