← TERMS

Meta-Loop Control

Second-order discipline for control loops — when sensors become targets, monitors need monitors

Meta-loop control is second-order discipline for closed-loop control: treating declared sensors, comparators, and actuators in a decision hypothesis as proxies under control pressure — not as ground truth — and wiring additional feedback about whether the primary loop still measures what the bind claimed.

First-order cybernetics closes the loop: observable → trigger → supersession. Second-order concern (cybernetics): the people and processes inside the loop know they are measured and optimise the measureGoodhart’s Law, Campbell’s Law. Meta-loop control is how an organisation stays humble while still closing loops.

The problem meta-loops address

(A)DR hypothesis tables require sensor · monitor · comparator · actuator. That is necessary — open-loop confidence until catastrophe is the anti-pattern of omitting them. It is not sufficient:

FailureMechanism
Green proxy, rotting aimMetric substitution — headline metric improves while harm rises elsewhere
False calmComparator never fires because the proxy was gamed, not because the assumption holds
Wrong supersessionActuator fires on proxy noise — expensive revision on a bad measure
Recorded false certainty”Trigger did not fire” read as proof the bind is still valid

Closed-loop on a Goodhart-vulnerable proxy can fail faster and with more legitimacy than open-loop ignorance.

Meta-loop elements

Meta-loop does not mean “never close the loop.” It means each primary-loop element carries tier honesty and its own review path:

Primary loopMeta-loop question
Sensor / observableIs this still a valid proxy for the assumption? Labelled proxy vs aim? Multiple independent signals?
MonitorWho watches override rate, contest overturn rate, reconciliation volume, incidents despite green metric?
ComparatorShould the threshold or metric definition be superseded — not only the decision?
ActuatorDoes firing create new gaming pressure? Is anyone rewarded for keeping the comparator quiet?

Practical meta-sensors already in this corpus: contestability, manual recency sampling, seeded review cases, disclosed model disagreement — human and adversarial checks that the automated loop still means what bind claimed.

Open loop “above” the closed loop

Some correction cannot be fully captured in declared metrics — qualitative judgment, novel failure modes, outsider signal, contest on grounds the table did not anticipate. Retaining legitimate open paths (contest, human override, executive supersession on record) is not a return to open-loop confidence until catastrophe when those paths are declared, tiered, and logged.

The risk is weaponization: rank-driven reopening, panic pivot, or “we’re being meta-principled” as cover for decisions as one-way streets in reverse — perpetual relitigation without observables. Meta-loop and contest path must be mechanised, not rhetorical.

What the literature says

SourceClaimRelevance here
Goodhart (1975)Statistical regularities collapse under control pressureComparators tied to actuators/incentives stop measuring what they measured ex post
Campbell (1979)Social indicators corrupt and distort processes when used for decisionsHuman organisations game declared ADR metrics
Strathern (1997)Measure → target → bad measureAudit/ratings culture; popular aphorism
Double-loop learning (Argyris & Schön, Organizational Learning, 1978)Question governing variables, not only adjust actions within fixed assumptionsMeta-loop = revisiting the metric and assumption frame, not only the tactical decision
Second-order cybernetics (von Foerster, etc.)Observer is inside the system; measurement changes the systemExplains why sensors are never passive
Perverse incentives / Cobra effectReward for proxy produces opposite of aimExtreme form of Goodhart
MOE vs MOP (program evaluation, defence analysis)Measure of effectiveness vs measure of performance — easy to target performance while effectiveness driftsADR: distinguish aim from operational proxy in the assumption table

A2 — working context: No settled formula guarantees a metric stays honest under incentive pressure. Meta-loop control is risk reduction and epistemic humility, not a proof of non-gameability — same stance as Event Sourced Science on incentive realignment without new proxies to game.

Corpus stance

A2 — working context: Decision-making principles and closed-loop control are not a magic wand — they do not eliminate Goodhart, politics, or weaponization. Skipping them remains the worse alternative (unprincipled bind epilogue). The honest posture is: apply principles, close loops, and maintain meta-loop humility — proxies tiered, contested, multi-signal, never rewarded for green alone.