double loop learningdouble-loopArgyris double-looporganizational double-loop learning

OrganisationEpistemology

Double-Loop Learning

Revise governing variables — not only actions within fixed goals, metrics, and assumptions

MARKO ŽIVKOVIĆ In review

Double-loop learning (Chris Argyris and Donald Schön, Organizational Learning, 1978) is learning that questions and revises governing variables — goals, policies, metrics, role definitions, and the assumptions that frame what counts as success — instead of only adjusting actions while those variables stay fixed.

Single-loop learning detects error against a fixed reference and corrects within the frame: the thermostat sees the room is cold and turns on the heat; the target temperature does not move. Double-loop learning asks whether the reference itself should change: whether this room should be heated, whether the target is wrong for the season, whether the sensor measures the wrong thing.

In this corpus, double-loop is the organisational-learning name for what meta-loop control and second-order cybernetics express in control vocabulary: when closed-loop control wires comparators to actuators and incentives, first-order correction is not enough — someone must be able to revise the hypothesis table, not only fire supersession on a row that still assumes the wrong proxy.

Single-loop vs double-loop

	Single-loop	Double-loop
Detects	Deviation from setpoint	Deviation and whether setpoint is right
Changes	Actions, tactics, execution	Governing variables — goals, metrics, policies, assumptions
Keeps fixed	Goals, metrics, policy frame	Nothing sacred except what bind still requires until superseded
Corpus pattern	Closed-loop control — sensor → comparator → actuator on declared observables	Meta-loop control — revise proxies, thresholds, MOE/MOP split, assumption frame
Failure mode if stuck	Optimise the wrong thing faster — local optimization, metric substitution	N/A — double-loop is the escape from single-loop on a bad frame
Risk if abused	—	Perpetual relitigation, decisions as one-way streets in reverse, “meta” cover for politics

Single-loop is necessary — organisations that never close tactical loops drift in open-loop confidence until catastrophe. Double-loop is necessary when the frame is wrong — when Goodhart’s Law and Campbell’s Law show the measure has become the game.

In (A)DR and decision principles

At bind, (A)DR records assumptions, observables, monitors, and falsification triggers — a single-loop contract: if evidence diverges, actuators fire. Double-loop enters when the assumption table itself must change:

The sensor was a valid proxy at bind but is gamed or decoupled from aim
The comparator threshold was wrong for stakes class
The MOE/MOP split collapsed — performance metric green, effectiveness flat
Governing policy (delegation, reward, escalation) produces perverse response to the loop

Decision Making Principle 4 requires explicit assumptions with falsification plans — single-loop discipline. Meta-loop control and contestability supply double-loop paths: supersede the metric definition, not only the decision; log human override and contest overturn as signals about the loop, not noise to suppress.

Double-loop learning is not permission to reopen every closed debate (Decision Making Principle 7). It is recorded revision of the frame when declared meta-sensors say the frame failed — same honesty as epistemic humility at the level of policy and measurement, not only facts.

Relation to cybernetics and meta-loops

Vocabulary	Emphasis
First-order cybernetics	Feedback closes error between state and reference
Double-loop learning (Argyris & Schön)	Organisational — who may change governing variables, and under what legitimacy
Second-order cybernetics	Observer inside the system — measurement changes behaviour
Meta-loop control (this corpus)	Mechanised double-loop — meta-sensors, multi-signal proxies, never reward green alone

Higher-order stacks (governance of governance, “triple-loop” in some literature) appear in cybernetics as declared depth, not a fixed count. A3 — open question: how many meta-loops before bureaucracy dominates remains context-dependent — see cybernetics limits section.

What the literature says

Source	Claim	Relevance here
Argyris & Schön (Organizational Learning, 1978)	Single-loop: error correction within fixed governing variables. Double-loop: change the variables themselves	Names the distinction this corpus maps to closed-loop vs meta-loop
Argyris (Overcoming Organizational Defenses, 1990)	Model I / Model II — defensive routines block double-loop	Explains why organisations stay single-loop despite pain (face-saving, bypass)
Goodhart (1975) · Campbell (1979)	Proxies collapse under control pressure	Why double-loop must include metrics and incentives, not only strategy slides
Second-order cybernetics (von Foerster, etc.)	The measured system responds to measurement	Mechanism behind gaming declared ADR sensors

Corpus stance

A2 — working context: Adopt double-loop as vocabulary for legitimate frame revision — paired with closed-loop control for tactical correction and meta-loop control for mechanised second-order discipline. Do not treat double-loop as rhetorical cover for unprincipled reopening; wire it through declared meta-sensors, contestability, and recorded supersession of assumption tables, not silent slide edits.