Behavioural standards across teams

Enterprises often accumulate successful pilots with incompatible conduct. One assistant is proactive and proposes plans before being asked. Another never challenges impractical constraints. One remembers every correction forever. Another forgets by design. Each team invented informal norms inside prompts. Users experience one brand and many personalities. Incident reviews blame "the model" when the real failure is missing organisational conduct policy attached to shared capability.

Helpful context: Cognitive Posture defines conduct architecture. The Collaboration Contract allocates shared judgment. From experimenting to governing explains when standards become organisational policy. This page covers conduct that must travel with shared cognitive units and agents.

Why standards must travel with libraries

A library of cognition without behavioural standards produces a platform that feels intelligent and behaves inconsistently. Functional substitution rules say which cognitive unit can replace another with equivalent judgment. Conduct standards say how the system behaves when using that judgment across products.

Without shared standards, nobody can say with confidence which judgments have been delegated, how escalation works when authority is missing, or whether temporary corrections require endorsement before they influence the next run. Two teams may import the same AskTargetedQuestion version and still produce incompatible clarification behaviour because agent strategy and local prompts diverged. Standards are how from experimenting to governing survives reuse. Publishing AssessMealPracticality to a domain library is a promise about behaviour and evidence, not about answer format alone.

Shared substrate

Model endpoints, retrieval, tools

Shared cognitive units

Named judgments, suites, substitution

Shared conduct

Posture, authority, escalation, explanation

Shared cognitive units need shared conduct. Endpoints alone do not produce a single organisational character.

What to standardise first

Not every conduct question needs a library on day one. The starting points are where inconsistency hurts trust or consequence. Reversible planning work may proceed proactively when assumptions are visible. Consequential medical interpretation deferrals must be consistent across agents. Purchase proposals require the same approval bridge everywhere. Memory discipline must prevent silent promotion of temporary corrections to endorsed knowledge.

The priorities are posture expectations for reversible versus consequential work, authority and escalation norms per decision class, explanation and contest patterns for shared cognitive units, and evaluation that includes conduct alongside answer quality before promotion.

What this looks like in practice

An organisation publishes standards as importable cognition. AskTargetedQuestion in the organisational library forms the smallest useful human question when information or authority is missing. All agents use it rather than inventing ad-hoc clarification prompts. Conduct cases test whether the question is targeted, not whether it is grammatical alone. Refusal and deferral templates define when to stop unsuccessfully, when to defer medical interpretation, when to bridge to human approval rather than guessing. Memory discipline standards require endorsement before temporary corrections become durable knowledge, so "avoid pasta this week" cannot silently become household policy.

Promotion gates reject a cognitive unit that answers well while escalating wrongly, challenging never, or remembering without policy. Answer quality without conduct quality recreates the pilot problem inside the library. This matters because the same household or clerk meets a product through more than one surface. A meal companion on mobile, a grocery copilot in chat, and an internal nutrition assistant should not teach three different habits for supplying allergy authority. When AskTargetedQuestion is organisational policy, the question shape, materiality threshold, and escalation copy match even when domain cognitive units differ.

Conduct as release criteria

Standards are release criteria owned by named roles. Domain owners maintain judgment standards for their library cognitive units. Knowledge stewards maintain endorsement and scope for what may influence decisions. Evaluation owners maintain rubrics and release gates including conduct cases.

This is governance without bureaucracy when standards ship as packages teams import, not as PDFs they ignore. An agent team imports AskTargetedQuestion v1.3 and inherits conduct tested on organisational cases. Regression on conduct becomes a library semantic versioning question with an owner.

Adopting cognitive units in existing code must include conduct parity during shadow eval alongside answer parity. Premature discipline warns against publishing standards before judgments stabilise. The balance is real: standards attach before promotion, standards arrive after the first prototype has named a judgment.

Rollout pattern

Standards roll in three passes. First, posture and authority expectations are documented in the spec for one product slice. Second, conduct cases are encoded in suites for shared cognitive units already in shadow eval. Third, promotion gates attach so non-compliant versions cannot enter domain libraries. That sequence respects timing while ensuring that by the time a cognitive unit reaches shared use, conduct travels with it.

Invoice intake shows the same pattern at enterprise scale. A field interpretation cognitive unit may live in the domain library while refusal templates for uncertain tax classification live in organisational conduct packages. Clerks learn one deferral voice. Auditors read one escalation record. Product teams keep domain judgment local without reinventing posture on every agent.

Audit questions before promotion

Before a shared cognitive units version promotes, reviewers answer conduct questions alongside judgment questions. Did the package ask only when information or authority was missing? Did it challenge an impractical constraint when the spec requires challenge? Did it defer medical interpretation with a named bridge rather than guessing? Did it treat chat preference as context until endorsement completed? Did it explain contest paths when the household or clerk could reasonably disagree?

These questions belong in release checklists, not in post-incident regret. A cognitive unit that passes answer suites while failing conduct teaches users that fluent wrong escalation is acceptable. Standards exist so promotion gates can reject that combination before it spreads through substitution.

Scaling signals

Scaling adds questions beyond individual product conduct. Portfolio observability asks whether autonomy rates, eval regression, library versions, and cost attribution remain visible as products multiply. Scaling signals include duplicate cognitive units detected in inventory audits, conflicting memory endorsement policies across products, rising token spend without decision-level attribution, and incident reviews that cannot name the owning library version. When those signals appear, the next work is reuse and governing, not a larger model tier.

Standards fail when treated as one-time policy PDFs. They succeed when imported as cognitive unit and agent hooks with conduct suites attached to promotion gates. Conduct review accompanies library semantic versioning review: a minor version bump that changes clarification behaviour is a conduct event even when answer format remains unchanged. Standards decay when products change faster than library promotion gates, making quarterly review of organisational conduct packages a necessary discipline.

When behavioural standards vary by team, nobody can say with confidence which judgments have been delegated.

Introduction to Thoughtware . Ch. 34

Common mistakes

Treating conduct as UX copy. Posture is architecture with eval hooks. It governs when systems ask, challenge, refuse, and defer, not how marketing describes them.

Letting each agent invent clarification. Users learn inconsistent habits for supplying authority when clarification patterns vary by product surface.

Promoting on answer suites alone. Fluency hides escalation failure until consequence arrives in production.

Standards without owners. A document nobody maintains becomes fiction within one quarter. Standards need semantic versioning and named maintainers like any other shared artifact.

What to do next

Three conduct failures that would destroy trust though answers looked fluent form the starting point for organisational standards. One standard per failure, attached as conduct cases to the next library promotion gate, makes the policy concrete and enforceable. The pattern extends as shared cognitive units mature: each promotion carries conduct alongside judgment, and each regression triggers review at the layer that owns the standard.

See Cognitive Posture, authority is granted, adopting cognitive units in existing code, and premature discipline.

Read next: Premature discipline.