What is inherited, and what is new

Cybernetics and control

Stafford Beer (Brain of the Firm, 1972; The Heart of Enterprise, 1979) specified what an organisation must contain to remain viable, and insisted that the function facing outward and forward — System 4 — hold a genuine model of its environment rather than a set of indicators. An indicator can be out of range; only a model can be surprised. He also identified the standing tension between System 3 and System 4, and the tendency of the former to capture the latter. Baarool's distinction between a measure that describes and a measure that decides is that argument in different clothes.

W. Ross Ashby (1956) established that a regulator must reduce the variety of its environment to something it can process, and that the reduction discards according to categories fixed before the signal arrives. Every organisation's blindness is therefore a function of its attention. This is the first of the five mechanisms.

Conant and Ashby (1970) proved the result the rest of this rests on: every good regulator of a system must be a model of that system. Beer's insistence that System 4 hold a model rather than a set of indicators is not a stylistic preference — it is that theorem applied to organisations. An indicator can be out of range. Only a model can be wrong.

Claude Shannon (1948) gave the vocabulary for the cadence problem. A channel has a capacity; information arriving faster than the channel can carry it is lost, and no improvement in the quality of the message recovers it. A quarterly reporting rhythm against a risk that moves in hours is a channel-capacity failure, not an attention failure.

Nancy Leveson (Engineering a Safer World, 2011) built the formal version: accidents as inadequate control rather than chains of component failure, with explicit control loops, feedback channels, and the specific claim that missing or delayed feedback is the mechanism. STAMP and STPA are the closest existing formalism to what Baarool proposes, applied to engineered systems rather than to governance information.

Safety science

Andrew Hopkins (Lessons from Longford, 2000; Failure to Learn, 2008) documented the finding this work generalises: Esso Longford and BP Texas City both held excellent personal-injury records while process safety degraded. A measure that described one thing was mistaken for a measure that decided about another. Hopkins reached it first, and reached it through narrative case study.

Erik Hollnagel on the gap between work-as-imagined and work-as-done; Sidney Dekker on why organisations produce accounts rather than understanding; Diane Vaughan (1996) on the normalisation of deviance at NASA; Charles Perrow (1984) on tight coupling and interactive complexity; Karl Weick on sensemaking and high-reliability organisations.

Inference under uncertainty

Karl Friston's free-energy principle and the active-inference literature supply the machinery for a system that maintains a model, predicts, and revises when the prediction fails. Baarool's technical work builds a trust-gated active-inference architecture for System 4 — the point of the gate being that a signal's admission to the model should depend on the credence placed in its source, and that this dependency should be explicit and adjustable rather than buried in an interface.

Lamport, Shostak and Pease (1982) named the harder half. In an adversarial setting, a component that lies consistently and plausibly is more dangerous than one that fails loudly. The same holds for organisational signals: the patient, agreeable lie defeats filtering, because filtering can only reject what looks wrong. That result is why the design problem is strategic rather than one of better detection — a conclusion the technical paper reaches and does not resolve.

Thomas Schelling (1960) on credible commitment, which is the game-theoretic sibling of Elster's pre-commitment: a constraint is only worth anything if the other party — including your own later self — can see that you cannot escape it.

Organisational learning

Chris Argyris and Donald Schön distinguished single-loop correction within an existing frame from double-loop learning that questions the frame, and named the defensive routines that prevent the second — arrangements that avoid embarrassment and simultaneously prevent anyone learning what produced them, with their own undiscussability as a second layer. Argyris also supplies the sharpest objection to Baarool's remedy: a recorded rationale will be written in the same defensive register as everything else unless it is required to state what would falsify it.

Pre-commitment

Jon Elster (Ulysses and the Sirens, 1979) formalised the device that every working stopping rule uses: acting at a moment when judgement is reliable to constrain a later self whose judgement will not be. The clause that matters is the one usually forgotten — that a signal to be released must be refused. A constraint that can be quietly revised is not a constraint.

Measurement and security economics

Dan Geer on the limits of what security can know about itself, and on risk as a function of dependency — which is transitive and largely invisible from inside a boundary. The Index of Cyber Security, which he co-founded, is the model for a recurring published measure held by a small independent team. Ross Anderson founded the study of security failures as incentive failures rather than technical ones; the Workshop on the Economics of Information Security continues it. Douglas Hubbard and Richard Seiersen argue for better quantification — a position orthogonal to this one in a way worth staging rather than eliding.

Judgement

Philip Tetlock established that predictions unaccompanied by scoring do not improve, and that accountability changes reasoning. Gary Klein's premortem is the practical instrument for the same problem.

Baarool draws on public investigative records rather than on privileged access. The bodies that produce them are the reason this can be done at all.

Transport safety. The Australian Transport Safety Bureau, the US National Transportation Safety Board, the UK Air Accidents Investigation Branch and Rail Accident Investigation Branch, and their counterparts under ICAO Annex 13. These are the closest existing institutions to what Baarool describes: statutory, no-blame, and concerned precisely with how a signal failed to reach someone able to act.

Process and industrial safety. The US Chemical Safety Board; the UK Health and Safety Executive; Australia's NOPSEMA, which publishes improvement and prohibition notices individually; the Longford Royal Commission and the Baker Panel report on Texas City.

Coronial jurisdictions. Eight in Australia, several of which publish recommendations alongside the responses to them; the UK's Prevention of Future Deaths regime. The recommendation-and-response pair is the scarce asset — findings are published widely, and what was done about them almost nowhere.

Audit and integrity. Nine Australian audit offices and their follow-up audits, which are documented evidence of an organisation's response to a warning it had already received; the UK National Audit Office; the US Government Accountability Office.

Corporate disclosure and law. Since December 2023, US registrants have been required to describe their cybersecurity governance annually and to disclose material incidents. No equivalent requirement exists in Australia or the United Kingdom. The Delaware oversight line, and the EU’s NIS2 and DORA regimes, are treated here as evidence about the argument rather than as background to it — external attempts to install these conditions by statute. The method page sets out what each jurisdiction contributes.

Privacy and data protection. The UK Information Commissioner's monetary penalty notices, which set out in detail what an organisation knew and when; the Office of the Australian Information Commissioner's notifiable breach reporting.

Teaching an Organisation to Watch the World

The plain-language edition of the design. How the outward-facing intelligence function of a viable organisation can be built as a working mechanism, and why its hardest problem is deciding what to believe. ISBN 978-1-7648504-0-7 · Read

A Trust-Gated Active-Inference Architecture for System 4

The technical companion: the mathematics, a linear-Gaussian reference implementation, and the experimental results. The architecture makes the interface’s hidden trust assumption explicit and adjustable, and demonstrates the system rejecting an injected false signal. It also establishes the limit of the approach — the patient, agreeable lie that no filtering catches — which is why the remaining problem is strategic rather than one of better detection.

Reference implementation

system4.py — a computational sketch of the perceive, plan and act loop, published alongside the papers so the claims can be run rather than taken on trust.

Essays

Collected here, and written under the author’s own name rather than the practice’s. Recognition attaches to people; methods attach to institutions.