Hazel runs itself.

Hazel exists to settle an internal argument: whether the AI work we sell survives contact with a domain that keeps score. Markets keep score. A demo can be impressive and wrong indefinitely; a position is open or closed, and the ledger settles it.

Deterministic on purpose

The strategy is deterministic. Given the same inputs, Hazel reaches the same decision every time — no sampling, no run-to-run drift, nothing that has to be described as usually. That is a constraint we chose, and it is the one that makes everything else testable: any behaviour can be replayed against recorded data and asserted exactly.

The model's job is not to have opinions at the moment of execution. It is to help build and maintain the rules; execution follows them, and the rules are written down where they can be argued with.

End to end

• Ingest. Market data normalised into a structured layer the strategy reads from — not a prompt.
• Decide. The strategy evaluates the state and emits an action, or an explicit no-op. Doing nothing is a decision it has to make on the record.
• Guard. Risk limits sit outside the strategy and can veto it. They are not something the strategy can talk its way past.
• Record. Every decision, and the state it was made from, is written down so it can be replayed later without a reconstruction.

Why it counts as evidence

Most AI work is demonstrated in the one condition it was built for. The interesting question is what a system does on the days nobody designed for — thin data, a stale feed, a state the rules do not cleanly cover. Hazel is where we find out, at our own expense rather than a client's.

A strategy you cannot replay is not a strategy. It is a story about one that worked once.

The engine is in testing. What it is already good for is the argument: when Disrupt Solutions says a system ships, this is the standard being described.