Judgement and Model Minimisation

Neil D. Lawrence

Trent retreat

Opening

The Judgement Layer

  • Intelligence is not only computation — it is communication under constraint.
  • Agentic AI makes organisational questions into engineering questions.
  • The governance question: does the judgement layer hold, migrate, or get bypassed?

A century-old disciplinary wall

  • Computer science: protocols, capacity, architectures, multi-agent systems.
  • Organisational science: hierarchy, escalation, bounded rationality, culture.
  • Same phenomenon, different vocabularies — until now.

Why the wall is falling

  • Agents share a protocol with managers: natural language.
  • Agent graphs encode SOPs, roles, review — org charts as message-passing.
  • Digitally mediated work makes organisational communication observable.

Motivation: communication as design

Conway’s Law

Any organization that designs a system (defined broadly) will produce a design whose structure is a copy of the organization’s communication structure.

Melvin Conway (Conway, 1968)

The memo that restructured Amazon

  • Bezos (2002): reorganise around communication bottlenecks, not org-chart habit.
  • Conway: systems copy the communication structures of the organisations that build them (Conway, 1968).
  • Inverse Conway: reshape communication so the desired architecture can emerge (Skelton and Pais, 2019).

From channels to topography

What Shannon does not explain

  • Information theory: what can pass through a given channel.
  • It takes senders, receivers, and links as given.
  • It does not explain why that structure exists, or how it should change.

Information topography

  • Stores: people, teams, records, agent state.
  • Channels: reporting lines, tickets, APIs, message buses.
  • Dynamics: how stores, channels, and links evolve under constraint.

The judgment layer

Judgment Layer: Organisations as emulsions

  • Organisations are a stable mixture: automatable routines and irreducible human context.
  • In a working emulsion the two phases are inseparable — like oil and water held in suspension.
  • The judgment layer is the interface: what gets challenged, what gets waived, what triggers a halt.

What the judgment layer looks like

  • Approval exceptions: “we usually accept this class of risk without escalation.”
  • Escalation triggers: “this pattern always goes to a human, even if the rule says approve.”
  • Context-dependent waivers: policy says X, but the situation means Y is safer.

Institutional tacit knowledge

  • The “judgement layer” in organisations is often tacit: norms, exceptions, escalation paths, and context that rarely makes it into documentation.
  • It lives in handoffs and approvals: what gets challenged, what gets waived, and what triggers a halt.
  • Ceding this tacit knowlege without making it explicit is how we accumulate agentic debt.

Good regulators and model minimisation

The Good Regulator theorem

  • Every good regulator of a system must be a model of that system.
  • The judgement layer is the regulator: it must still model what it oversees.
  • If the system outruns the model, regulation fails — even if the tools still run.

Judgement as live regulation

  • Routine variation can be summarised away; anomaly must be escalated.
  • The layer is where authority over flow is exercised — approve, contest, overrule.
  • Same object in CS and Org science
    • Ownership, control, exception handling.

Why we minimise models

  • We minimise so that we can maintain a model.
  • Smallest decision-relevant representation that still supports oversight.
  • Perhaps prefer decomposable graphs over monoliths
    • Consistent explorable behaviour of system

Against maximisation by default

  • “Use the largest model” is a default, not a design.
  • Capability without a modellable envelope is how debt compounds.
  • Ask: what is the least system that still earns regulated trust?

Agentic debt

How agentic debt accrues

  • Automate attenuation without preserving the escalation path institutions assume.
  • Volume moves; authority does not — or authority moves without accountability.
  • The formal org chart and the measured topography diverge.

Debt as failed regulation

  • Debt is unpaid explanation: systems that work until they must be justified.
  • Large opaque graphs accelerate debt when outputs are accepted without challenge.
  • Solvency = judgement’s model capacity ≥ system complexity.

Discussion

Thanks!

References

Conway, M.E., 1968. How do committees invent? Datamation 14, 28–31.
Skelton, M., Pais, M., 2019. Team topologies: Organising business and technology teams for fast flow. IT Revolution Press, Portland, OR.