Human governance
What the platform proposes on its own, and what only the human is allowed to govern.
The human always owns the strategy. The multi-agent research system explores, scores and proposes continuously — but the moves that reshape the lab or reach the market belong to the human. The boundary is hard-coded in the tool registry and the permission flags, never in a prompt: the platform proposes, the human governs.
The boundary
| Action category | Who decides | Execution |
|---|---|---|
| Research, scoring and shadow-mode observations | The agents, continuously | AUTO |
| Operating a position inside granted permissions (entry, reduce, exit) | The agents, within human-granted permissions | AUTO + notification |
| Strategy, deployment, capital, execution permissions, scoring or learning-rule changes | Human validation required | Manual |
| Arbitrary order placement by the LLM | Forbidden | Not in the registry |
What the human governs
The lab can generate hypotheses, build features, score opportunities and observe markets without supervision — none of that risks capital. What stays with the human is the layer that changes the rules or touches money: which strategies become active, which scoring formulas go live, how much capital is allocated, and when execution permissions are granted or widened. Each of those reshapes the agents’ mandate rather than acting within it.
“Safe enough” is a permission, not a vibe
The system never asks the model whether an action is safe enough to take alone. Safety is a property of which tools exist in the registry and which permissions the human has granted. The agents can act only inside those bounds; there is no arbitrary order-placement tool to call — so there is nothing to misjudge, and during research the no-real-money key means even a compromised agent reaches no real money.