An autonomous system, deliberately constrained
The most important engineering in Alpha Yazan AI is not what the system can do — it is what the system cannot do without evidence and explicit human authority. This page is the doctrine we build to.
Seven governing principles
-
Authority separation
The engine that creates strategies is architecturally separate from the machinery that grants authority. Research can propose; only governance — with a human owner above it — can promote. The system cannot expand its own permissions.
-
Paper before live
Prospective paper evaluation is a mandatory stage, not a demo mode. Live-money authority is not enabled anywhere in the platform, and enabling it would be a governed, owner-level decision gated on evidence — not a configuration flag.
-
Fail-closed
When data degrades, a component misbehaves, or the system encounters a state it does not recognize, it stands down. Uncertainty reduces authority; it never expands it.
-
Frozen identities
A validated strategy is frozen: its definition is versioned and immutable. The thing that earned trust is exactly the thing that runs. Any change creates a new identity that must earn its own evidence.
-
Evidence gates
Promotion between stages requires passing predefined evidence gates — falsification attempts, out-of-sample tests, shadow comparison. Gates are set before the evaluation, so results cannot quietly move the goalposts.
-
Rollback as a first-class operation
Every promotion is reversible by design. Rollback paths are built and rehearsed before they are needed, so retreat is always cheaper than damage.
-
Negative-result preservation
Failed hypotheses are recorded with the same care as successes. A research program that deletes its failures cannot be trusted about its wins.
Engineering success ≠ strategy profitability.
A platform can be superbly engineered and still hold strategies that make no money. We treat these as separate questions with separate evidence standards, and we never let pride in the first stand in for proof of the second.
Governance is the product's foundation
Autonomous systems earn trust through constraints that hold under pressure. We publish our doctrine so it can be held against us — and because we believe governed autonomy, not maximal autonomy, is what makes AI research systems useful.