Common response and PI optimization policy, revision 1.7

This policy applies to every laser and provider when a card admits response or PI optimization. It does not grant additional operations: use only the frozen card's declared knobs and the owner adapter's validated capabilities. A settling-only or FF-only study cannot silently tune PI. Hardware-specific gain units, actuation, timing, continuity compensation and constraints belong to the owner adapter; the scientific decision policy is shared.

Where admitted, consider P, I and the integral-output contribution limit independently, in either direction. Never assume that gains must decrease together. Distinguish an integral tuning limit from absolute device output limits. Preserve the physical envelope, output continuity, foreground priority and unknown-write handling. Do not change a calibration or physical safety bound as a tuning shortcut.

For joint target-response optimization, judge the overall balance of response speed and steady performance. A modest rise in steady RMS can be worthwhile for a substantial improvement in response speed: for example, 10 percent more RMS for twice as fast a response can be a reasonable trade-off. This is an illustration, not a fixed allowance, weight or automatic adoption rule. Do not require every metric, especially RMS, to be non-degrading. Quantify both the benefit and the cost, with their repeatability and uncertainty, and explain why the trade-off suits the actual operating objective. Consider centered noise, mean error, RMS relative to the current target, drift, oscillation, overshoot, saturation, actuator activity, mode-hop evidence and robustness. Explicit experimental precision requirements and physical safety bounds remain binding.

Use comparable baseline, candidate and repeated confirmation observations. Observe enough oscillation/drift periods; a short window cannot establish a long-term optimum. Separate transient records from steady windows and analyze each leg against its own target. Record request, verified actuator-write and acquisition timestamps separately, using full-rate measured data where available. Distinguish existing tracking bias from the response caused by the new step. Preserve gaps, missing samples and unobserved settling as such; an observation timeout is not a measured settling time.

Learn a defensible settling criterion from evidence rather than treating the existing LOCKED/TRACKING tolerance as the optimization objective. Freeze the comparison criterion across candidates; if it changes, re-evaluate all comparable raw traces. Do not loosen a threshold merely to make a candidate pass. A response criterion applies only in its measured operating context, including step size, gains, FF, source cadence and actuator state. Report optical settling and thermal/PZT recentering separately whenever they differ; require recentering only when the admitted task declares it.

The task model chooses proposals and authors the explanation. The host validates and executes bounded trials and returns owner-confirmed receipts; it must not replace the model's trade-off decision with an undisclosed hard RMS gate or fixed scalar weighting. After repeated confirmation, return the new measurements to that same task model before accepting its final adoption decision. A proposal, trial application, runtime adoption and saved configuration are different outcomes. State which occurred, which settings remain and which conclusions still require testing. On interruption, restore only the current intervention's uncommitted settings; never overwrite a newer foreground target.
