Hold the ceiling at A3 for discovery research. Literature synthesis and hypothesis generation are exactly where agents outperform — they read everything. A scientist validates before anything becomes a program. Raise it only after two consecutive quarters where the override rate stays under five percent and every override has a written cause.

Keep the repeatable research thread volume in the spine and pull the exception tail back into the function. The spine is cheaper per unit; the tail is where the context that prevents rework actually sits.

Hold the ceiling at A2 for innovation portfolio. Agents assemble evidence and even recommend killing a program. Stopping or funding research is a capital decision with careers attached, so a committee owns it. Raise it only after two consecutive quarters where the override rate stays under five percent and every override has a written cause.

Keep the repeatable program volume in the spine and pull the exception tail back into the function. The spine is cheaper per unit; the tail is where the context that prevents rework actually sits.

Hold the ceiling at A3 for experiment design. Protocol drafting and power analysis are agentic. Ethics clearance and scientific validity judgments are personally owned by a qualified researcher. Raise it only after two consecutive quarters where the override rate stays under five percent and every override has a written cause.

Hold the ceiling at A4 for lab operations. Scheduling, replenishment and instrument monitoring are pure logistics and run touchless. Calibration certificates carry a technician signature. Raise it only after two consecutive quarters where the override rate stays under five percent and every override has a written cause.

Agents in this lane may act without a human when the research thread sits inside the stated value, risk and confidence envelope. Outside it, the unit holds at a gate with a named approver and a running clock.

Every autonomous decision writes inputs, the rule version applied, the model and prompt version, the output and a reversal path. An action with no evidence record is treated as a control failure, not a fast decision.

Escalation is mandatory when confidence falls below the floor, when two options sit inside the confidence band, or when a break condition on a ratified position fires. Director, Discovery Research owns the response clock.

Agents in this lane may act without a human when the program sits inside the stated value, risk and confidence envelope. Outside it, the unit holds at a gate with a named approver and a running clock.

Every autonomous decision writes inputs, the rule version applied, the model and prompt version, the output and a reversal path. An action with no evidence record is treated as a control failure, not a fast decision.

Agents in this lane may act without a human when the protocol sits inside the stated value, risk and confidence envelope. Outside it, the unit holds at a gate with a named approver and a running clock.