Closed-Loop Control

Closed-loop control applies cybernetic feedback to committed decisions and platform binds: the organisation acts on a hypothesis at bind, instruments the world for the observables that hypothesis requires, and updates through declared actuators — falsification triggers, contest resolution, supersession, delegation policy revision, escalation threshold changes — when evidence diverges. The original bind stays visible; correction travels forward as L0 facts.

Introduction in decision terms: Decision Principles and the Record — assumption tables with monitors and triggers. At platform scale: One Authority Per Invariant — At Scale — platform domains instrument error rates, contest outcomes, drift, and calibration decay, then act.

The pattern

At bind, declare the loop:

The loop is closed when actuator firing is mandatory and recorded, not optional negotiation. Debate closed at bind (Decision Making Principle 7); observation continues; revision routes through mechanism.

In software systems

Metric alerts wired to ticket creation and spend gates. Model drift monitoring that triggers delegation review. ADR assumption tables with automated checks. Override and contest metrics feeding threshold recalibration. Event-sourced supersession chains when triggers fire — not UPDATE policy SET ….

What good looks like: the system can answer what observation would change our mind and what happens when it crosses — without a meeting.

In human organisations

Quarterly assumption reviews with pre-declared triggers in the (A)DR. Strategy commits paired with market observables and mandatory review dates. Post-bind contestability as formal feedback channel. Platform/compliance domains with published SLAs and signal-driven policy revision when error rates or contest overturn rates shift.

What good looks like: revision depersonalises — the trigger fired, not you attacked me.

In socio-technical systems

Hybrid AI+human binds close the loop: automated commits within recorded delegation policy; human contest and override as sensors; calibration studies and threshold L0 updates as actuators. Centralised cross-cutting authority stays learnable only when consequences update bind policy — not when the platform emits L0 forever without feedback.

Goodhart, Campbell, and second-order risk

Closing the loop creates control pressure on whatever you declared as sensor and comparator. Goodhart’s Law and Campbell’s Law predict the same mechanism: once a measure is wired to mandatory actuators and incentives, it stops behaving like a reliable measure of the assumption. Metric substitution is the dark pattern form — the dashboard improves while the aim rots.

That does not argue for open-loop instead. Open-loop confidence until catastrophe is worse at the first order. The risk is false legitimacy: a green comparator on a gamed proxy reads as the record proves we were right — confident deck with wiring diagrams.

Mitigations — meta-loop control:

• Label observables as proxies with honest tier and known distortion risk
• Use multiple independent signals — outcome-linked, not activity-linked only (MOE vs MOP discipline)
• Meta-sensors: contest overturn rate, override rate, incidents despite green metric, manual recency sampling
• Actuators that can supersede the metric definition, not only the decision
• Never reward “trigger did not fire” as performance success

A2 — working stance T12: No formula guarantees non-gameability under incentive pressure. Closed-loop plus meta-loop is humility with machinery, not a magic wand.

Why this solves the anti-patterns

Open-loop confidence until catastrophe

Missing feedback structure at bind — sensors and actuators wired here replace open-loop omission.

Decisions as one-way streets

Forward-only commits without designed return path. Closed-loop control is the return path — triggers, contest, supersession.

Commitment-as-immunity and unprincipled bind epilogue

Feedback externalised in the record depersonalises late correction; without the loop, repair routes through identity and politics.

Related patterns

• Decision Making Principle 4 — hypothesis schema at bind
• Experiment architecture — pre-committed design, separate observation streams
• Refutation propagation — dependents receive correction events

Corpus stance

Adopt — same architecture for software platforms, organisational strategy, and (A)DR discipline: closed-loop control, not open-loop confidence until catastrophe.