Specification v2.0.0 · 2026-08-10

CSM 2.0 Specification
Framework & Governance

The versioned, deterministic-by-design governance operating model for AI systems.

Context

Why CSM 2.0 Exists

The original 2025 CSM publication defined four governance domains and sixteen components. It described what governance responsibilities exist but did not formalize how those responsibilities are evaluated, evidenced, transferred between teams or reopened when conditions change.

CSM 2.0 adds an operational specification layer: governance contracts, a determinism boundary, explicit state models, evidence and decision schemas, handoff contracts, reassessment triggers and a reference evaluator. The original domains, components and their names are preserved unchanged.

Continuity

What Remained from V1

Four governance domains: Enterprise, Project, Code, UX
Sixteen original components (4 per domain)
Original component names and descriptions
Proportionality principle: governance depth scales with risk
Handoff visibility as a core value
Original provenance and authorship

CSM 2.0 is based on and evolved from the original 2025 CSM publication. Cognitive System Management: A Framework for Enterprise AI Project Governance” by Subodh KC, AI Governance on Medium, 2025-08-29.

Read original article
WHERE

Four Governance Domains

ENT

CSM-Enterprise

Establish the decision context before individual projects improvise it.

Who has authority, who owns the outcome and risk, and what organizational boundaries apply?

ENT-POLICYENT-RISKENT-DATAENT-MANDATE
PRJ

CSM-Project

Make scaling an explicit governance decision.

What evidence should justify continuing, changing, scaling or stopping an AI initiative?

PRJ-BUSINESSPRJ-TESTINGPRJ-SCALEPRJ-PLAYBOOK
CODE

CSM-Code

AI-generated code is still organizationally accountable code.

How should software engineering governance change when AI contributes to implementation?

CODE-STANDARDSCODE-SECURITYCODE-HUMANCODE-TRACE
UX

CSM-UX

Governance reaches the people relying on the system.

What do humans need to understand, supervise, challenge and appropriately use AI-supported outcomes?

UX-IMPACTUX-EXPLAINUX-CAPABILITYUX-ADOPTION
WHAT

Six Execution Functions

EF1-PURPOSE

Purpose, Scope & Accountability

Establishes the intended purpose, scope, ownership and accountability boundaries for the AI system.

Question: What is the system for, who is accountable, and what boundaries apply?

EF2-MAPPING

System, Data & Dependency Mapping

Maps the system architecture, data sources, model providers, tools and integrations.

Question: What components, data and dependencies comprise the system?

EF3-RISK

Risk, Evaluation & Monitoring

Identifies, evaluates and continuously monitors risks including drift, performance and security.

Question: What risks exist and how are they evaluated and monitored?

EF4-DELIVERY

Controlled Delivery & Change

Governs the delivery pipeline, change management, testing gates and deployment controls.

Question: How are changes controlled and delivered?

EF5-OVERSIGHT

Human Oversight, Feedback & Learning

Ensures human oversight mechanisms, feedback channels and organizational learning loops.

Question: How do humans oversee, provide feedback and learn from the system?

EF6-COMPLIANCE

Compliance, Evidence & Assurance

Collects, organizes and produces evidence for governance assurance and regulatory obligations.

Question: What evidence demonstrates governance and what compliance obligations apply?

HOW

Governance Contract Model

Every CSM component has a governance contract that formalizes its purpose, applicability, objective rules, human judgment points, evidence requirements, handoffs and reassessment behavior.

Purpose & core question
Applicability conditions
Required & optional inputs
Objective rules (deterministic)
Human judgment points (interpretive)
Required decisions
Required evidence with expiry
Responsible & accountable roles
Outputs
Blocking conditions
Allowed exceptions with scope & duration
Handoff targets & outputs
Reassessment triggers
Execution function mappings
Source provenance
Component dependencies
View all 16 governance contracts
States

State Model

Requirement States (10)

NOT_ASSESSED

No evaluation has been performed for this requirement.

NOT_APPLICABLE

The requirement has been evaluated and determined not to apply to this system context.

REQUIRED

The requirement applies and has not yet been satisfied. Work has not begun.

IN_PROGRESS

The requirement applies and work is underway but not complete.

SATISFIED

All applicable objective rules pass and required evidence exists.

PARTIAL

Some but not all required evidence or rules are satisfied.

BLOCKED

A blocking condition exists that prevents progress.

EXCEPTION_APPROVED

An approved exception covers this requirement. The exception has a valid expiration.

HUMAN_REVIEW_REQUIRED

A human judgment point has been reached and requires explicit review.

REASSESSMENT_REQUIRED

A reassessment trigger has fired for this requirement.

Domain States (5)

NOT_STARTED

No applicable components have been assessed.

IN_PROGRESS

At least one applicable component is assessed but not all are satisfied or validly excepted.

READY_FOR_HANDOFF

All applicable required components are satisfied or validly excepted and no human review remains.

BLOCKED

At least one applicable blocking component is BLOCKED.

REASSESSMENT_REQUIRED

At least one applicable component requires reassessment.

System States (5)

ASSESSMENT_INCOMPLETE

One or more applicable components have not been assessed.

GOVERNANCE_BLOCKED

One or more applicable components are BLOCKED.

HUMAN_DECISION_REQUIRED

One or more human judgment points require explicit review.

READY_FOR_GOVERNANCE_DECISION

All applicable components are satisfied, validly excepted, or not applicable. No blockers or human reviews remain. A human authority may record an approval decision.

REASSESSMENT_REQUIRED

One or more reassessment triggers have fired.

Determinism

Determinism Boundary

For a fixed CSM specification version, organizational policy configuration, explicit assessment date and normalized input record, objective CSM evaluation must produce the same computed requirement set and statuses.

Deterministic Decisions (10)

Whether an accountable owner is assigned (ENT-MANDATE)
Whether a required evidence artifact exists (any component)
Whether an exception has expired (any component)
Whether a reassessment trigger has fired (system-wide)
Whether a dependency is satisfied (component dependency graph)
Whether evidence has expired (evidence schema)
Whether a component is applicable based on system context (applicability rules)
Whether required inputs are present (input validation)
Whether a domain is ready for handoff (domain state derivation)
Whether a blocker exists (blocking condition check)

Human Judgment Decisions (10)

Whether a use is ethically acceptable (ENT-POLICY)
Whether residual risk is acceptable (ENT-RISK)
Whether a legal obligation applies (ENT-POLICY)
Whether an explanation is adequate for affected users (UX-EXPLAIN)
Whether a business benefit justifies risk (PRJ-BUSINESS)
Whether a scale decision should proceed (PRJ-SCALE)
Whether human oversight is proportionate (CODE-HUMAN)
Whether training is adequate for the system (UX-CAPABILITY)
Whether an exception should be approved (any component)
Whether a system is approved for a specific purpose (system-level approval)
Human Review

Human Review Boundary

When a rule cannot be resolved objectively, the evaluator returns HUMAN_REVIEW_REQUIRED and provides: the question requiring review, the responsible role, the evidence needed and the affected components.

The evaluator does not call OpenAI or another model to resolve human judgment decisions. A human authority may separately record APPROVED_FOR_<PURPOSE> as a decision record, not an engine inference.

Evidence

Evidence Model

Evidence records are versioned, have explicit effective and expiry dates, and link to specific requirement IDs.

EvidenceRecord {

evidenceId, type, title, source, owner

createdAt, effectiveAt, expiresAt?

systemVersion?, artifactVersion?

uri?, hash?

relatedRequirementIds[]

status: active | expired | superseded

}

Hash is optional and supports integrity checking. An artifact with a hash is not called “immutable.”

Decisions

Decision Records

DecisionRecord {

decisionId, requirementId, question

decision, decisionType

reviewer, reviewerRole, rationale

evidenceRefs[]

decidedAt, effectiveAt, expiresAt?

specVersion, policyVersion

}

decisionType distinguishes: objective_computation, human_approval, legal_review, exception, risk_acceptance

Exceptions

Exceptions

ExceptionRecord {

exceptionId, requirementId, reason

approver, approverRole

compensatingControls[]

issuedAt, expiresAt, scope

evidenceRefs[], reassessmentTrigger

}

Expired exceptions must not remain silently valid. The evaluator checks expiresAt against the explicit assessmentAsOf date.

Handoffs

Handoff Contracts

Handoffs are bidirectional feedback loops, not a waterfall.

HANDOFF-ENT-PRJ

Enterprise to Project

Policy and risk boundaries become project requirements.
ENTPRJ
intendedPurposeaccountableOwnerpolicyBoundariesriskContextdataOwnershipstrategicMandate
HANDOFF-PRJ-CODE

Project to Code

Acceptance criteria and approved assumptions become implementation constraints.
PRJCODE
approvedUseCasebusinessSuccessCriteriaevaluationCriteriareleaseExpectationsapplicableConstraintsscaleDecisionCriteria
HANDOFF-CODE-UX

Code to UX / Operations

Actual system behavior and limitations shape user interaction and oversight.
CODEUX
implementationVersionvalidatedBehaviorknownLimitationssecurityTestResultsmonitoringSignalsfallbackBehavior
HANDOFF-UX-PRJ

UX to Project

User feedback and operational behavior trigger product/project reassessment.
UXPRJ
userFeedbackusageFindingsoperatorIssuesincidentSignalsusabilityFailuresescalationData
HANDOFF-PRJ-ENT

Project to Enterprise

Incidents, lessons and newly discovered risks may require policy or risk updates.
PRJENT
newRiskspolicyGapsincidentsstrategicChangesdiscoveredDependencieslessonsRequiringEnterpriseChange
Reassessment

Reassessment Triggers

18 triggers. Each specifies which domains and components reopen. Triggers do not automatically reopen all 16 components unless necessary.

TRIGGER-USE-CHANGE

Intended-use change

The intended purpose of the system has changed.

TRIGGER-SCOPE-EXPANSION

Scope expansion

The system scope has expanded beyond the approved use case.

TRIGGER-OWNER-CHANGE

Accountable-owner change

The accountable owner has changed.

TRIGGER-MODEL-CHANGE

Model change

The AI model has been changed or materially updated.

TRIGGER-PROVIDER-CHANGE

Model-provider change

The model provider has changed.

TRIGGER-CONFIG-CHANGE

Material prompt/configuration change

Material prompt or configuration changes affect system behavior.

TRIGGER-DATA-CHANGE

Major dataset or RAG-source change

Training data, retrieval sources or operational data have materially changed.

TRIGGER-TOOL-AUTHORITY-CHANGE

Tool/agent authority change

The tools or authority granted to an agent have changed.

TRIGGER-NEW-INTEGRATION

New integration

A new integration has been added to the system.

TRIGGER-DEPENDENCY-CHANGE

Material dependency change

A material dependency has changed or been removed.

TRIGGER-SECURITY-INCIDENT

Significant security incident

A significant security incident has occurred.

TRIGGER-EVAL-FAILURE

Significant evaluation failure

A significant evaluation failure has been detected.

TRIGGER-DRIFT

Material drift

Material model or behavioral drift has been detected.

TRIGGER-USER-HARM

User harm/complaint pattern

A pattern of user harm or complaints has been identified.

TRIGGER-JURISDICTION-CHANGE

Jurisdiction/regulatory-context change

The regulatory or jurisdictional context has changed.

TRIGGER-POLICY-CHANGE

Major organizational policy change

Organizational AI governance policy has materially changed.

TRIGGER-EXCEPTION-EXPIRY

Approved exception expiry

An approved exception has expired or is about to expire.

TRIGGER-DECOMMISSION

Retirement/decommissioning

The system is being retired or decommissioned.

Proportionality

Governance Depth & Proportionality

BASELINE

Minimum governance for all AI-enabled systems. Universal requirements apply.

ENHANCED

Additional requirements activated by organization-defined factors beyond baseline.

INTENSIVE

Maximum governance depth for systems with significant consequential, safety, or regulatory exposure.

Proportionality Factors (10)

Consequential decisionsExternal customer impactSensitive dataAutonomous actionDifficult to reverseSafety impactSignificant financial impactLarge affected populationCritical vendor dependencyExternal regulatory exposure

The framework supplies recommended factors. The organization defines actual policy thresholds. Once thresholds are configured, evaluation is deterministic.

Machine-Readable

Machine-Readable Specification

The CSM 2.0 specification is generated from canonical TypeScript source. No manually maintained duplicate JSON.

JSON Specification

Full machine-readable spec with all domains, components, contracts, triggers and crosswalk.

JSON Schema

JSON Schema for validation and tooling integration.
Evaluator

Reference Evaluator

A pure deterministic reference evaluator implements the CSM 2.0 rules. It uses no LLM calls, no randomness, no external API and no implicit clock.

No LLM calls
No randomness
No external API dependencies
No implicit current-time (uses explicit assessmentAsOf)
Stable normalized inputs
Stable lexicographic output ordering
Explicit spec version
Explicit policy version

The evaluator is a reference implementation. It is NOT HAIEC itself.

Test Fixtures

Example Test Scenarios

8 golden test fixtures covering representative scenarios. All are test fixtures, not real customers.

fixture-1

Low-impact internal AI assistant (BASELINE)

fixture-2

External customer-facing AI assistant (ENHANCED)

fixture-3

AI-assisted software-development workflow (BASELINE)

fixture-4

High-impact decision-support scenario (INTENSIVE)

fixture-5

Agent with tool authority (ENHANCED)

fixture-6

Expired governance exception (BASELINE)

fixture-7

Human/legal applicability review (INTENSIVE)

fixture-8

Material provider/model change (ENHANCED)

Limitations

Limitations

CSM 2.0 is a practitioner-developed governance methodology, versioned for public scrutiny and implementation. It is not a regulatory standard, certification, peer-reviewed standard or legal framework.

Does not determine legal compliance
Does not replace qualified legal review
Does not guarantee governance outcomes
Does not produce composite scores or ratings
Human judgment decisions require qualified reviewers
Proportionality thresholds are organization-defined
Versioning

Version History

v1.0

Historical conceptual publication

2025-08-29
Original CSM publication. Four governance domains and sixteen original components.

Added:

  • Four governance domains: Enterprise, Project, Code, UX
  • Sixteen original components (four per domain)
  • Cross-domain handoff visibility concept
  • Implementation guidance (assessment, pilot, scale)
  • Proportionality concept
v2.0.0

Current specification

2026-08-10
Formalizes CSM into a deterministic governance operating model with machine-readable contracts.

Added:

  • Two-axis operating architecture (Domains x Execution Functions)
  • Governance Contract schema for every component
  • Objective rules vs human judgment points
  • Determinism contract
  • Explicit state model (requirement, domain, system)
  • Evidence schema with versioning
  • Decision record schema
  • Exception contract with expiration
  • Formal handoff contracts
  • Reassessment engine with deterministic triggers
  • Component dependency graph
  • 4x6 operational mapping
  • Governance depth / proportionality profiles
  • Semantic versioning (spec, policy, assessment)
  • Machine-readable specification design
  • Informative NIST/ISO crosswalk
v2.0.1

Expanded implementation examples

2026-08-11
Expanded V1 implementation examples and handoffs to address vendor governance, agentic code generation, consent management, and incident response governance.

Added:

  • Enterprise: vendor governance, provider data-processing review, model card requirements
  • Enterprise: incident response authority, escalation protocol, breach notification procedure
  • Code: agentic tool authority policy, agent sandbox boundaries, AI API key management, prompt injection defense
  • UX: AI processing consent, opt-out and human review request, misuse detection monitoring, manipulation guardrail review
  • Enterprise roles: vendor manager, incident response lead
  • Code roles: platform engineer
  • UX roles: privacy counsel
  • Handoffs: Enterprise to Code (security incidents), Code to Enterprise (policy gaps), UX to Enterprise (user-reported incidents)
Provenance

Original Publication Source

Cognitive System Management: A Framework for Enterprise AI Project Governance

Subodh KC · AI Governance on Medium · 2025-08-29
Read original article

CSM 2.0 does not produce legal compliance verdicts. It may produce governance requirements, evidence gaps, review requirements, decision records and framework mappings. Legal applicability or legal sufficiency must remain subject to qualified interpretation where needed.

Forbidden

Forbidden Output States

The evaluator must never produce these states:

LEGAL_COMPLIANTEU_AI_ACT_COMPLIANTCOMPLIANTSAFETRUSTWORTHYCERTIFIEDAUDIT_PROOF

The evaluator must never produce these score names:

compliance_scoresafety_scoretrust_scorerisk_score
Informative Mapping

NIST AI RMF & ISO/IEC 42001 Crosswalk

Informative Mapping. This crosswalk is informative. It does not constitute equivalence, certification, or a compliance guarantee. CSM 2.0 is an independent governance methodology. NIST AI RMF and ISO/IEC 42001 are separate frameworks with their own scopes, definitions and requirements. Organizations seeking certification or compliance with NIST or ISO should consult the original standards and qualified assessors.

CSM ComponentNIST FunctionISO ClauseRelationship
ENT-POLICYGovernClause 5 - Organizational context, Clause 6 - Planningsupports
ENT-RISKMapClause 6.1 - Actions to address risks and opportunitiessupports
ENT-DATAMapClause 7.3 - Data for AI systemspartially-maps
ENT-MANDATEGovernClause 5.3 - Roles, responsibilities and authoritiessupports
PRJ-BUSINESSMapClause 8.1 - Operational planning and controlpartially-maps
PRJ-TESTINGMeasureClause 8.3 - AI system impact assessmentsupports
PRJ-SCALEManageClause 8.4 - AI system change managementpartially-maps
PRJ-PLAYBOOKGovernClause 7.5 - Documented informationsupports
CODE-STANDARDSManageClause 8.2 - AI system development controlssupports
CODE-SECURITYMeasureClause 8.2 - AI system development controls (security)supports
CODE-HUMANManageClause 8.3 - Human oversightsupports
CODE-TRACEMeasureClause 7.5 - Documented information (records)supports
UX-IMPACTMapClause 8.3 - AI system impact assessmentsupports
UX-EXPLAINManageClause 8.5 - Transparency and information for userssupports
UX-CAPABILITYGovernClause 7.2 - Competence and trainingpartially-maps
UX-ADOPTIONMeasureClause 9.1 - Monitoring, measurement, analysis and evaluationsupports

Explore CSM 2.0 Further

Browse all 16 governance contracts, try the reference assessment, or return to the main CSM page.

Discuss AI →