Specification: GenOps Governance Specification v0.1
GenOps is an open, runtime-agnostic governance specification for AI workloads.
GenOps defines the minimal interoperable governance surface for AI workloads. It specifies the semantic contract required for consistent runtime governance across independently developed systems.
This repository contains the GenOps specification document. Governance model: GOVERNANCE.md. Contribution process: CONTRIBUTING.md.
GenOps standardizes governance semantics for AI workloads. It defines:
- The AI Workload Unit (AWU), the atomic boundary of governed work.
- A Deterministic Accounting Invariant, ensuring reservation-before-execution resource tracking.
- A canonical Enforcement Decision Model.
- A fixed reason_code taxonomy for governance outcomes.
- A minimal set of required telemetry attributes and events.
- Clear compliance criteria.
GenOps does not define APIs, SDKs, user interfaces, policy languages, or control planes. It defines the semantic surface that a compliant runtime MUST emit.
AI workloads consume measurable resources and operate under governance constraints:
- Budget limits
- Rate restrictions
- SLA boundaries
- Attribution requirements
These controls are implemented inconsistently across vendors and runtimes. GenOps establishes a shared vocabulary and invariant model so that:
- Governance decisions are machine-readable.
- Accounting is deterministic.
- Telemetry is interoperable.
- Independent runtimes can be evaluated consistently.
GenOps is intentionally narrow. It follows these principles:
- Runtime-agnostic. No dependency on a specific architecture.
- Telemetry-first. Governance must be externally verifiable.
- Deterministic. Identical inputs produce identical reservation outcomes.
- Closed taxonomy. No freeform reason strings.
- Minimal surface area. Only what is required for interoperable governance is defined.
GenOps extends OpenTelemetry semantic conventions. All GenOps telemetry:
- Uses the
genops.*namespace. - MUST be representable as standard OpenTelemetry signals (traces, logs, metrics).
GenOps does not modify or replace existing OpenTelemetry semantics.
GenOps defines governance semantics. OpenTelemetry defines observability transport and structure. They are complementary.
The OpenTelemetry GenAI semantic conventions describe execution-level details such as model calls and token usage. GenOps defines the governance layer—policy decisions, accounting, and enforcement—and is designed to coexist with GenAI signals without overlap. Implementations MAY correlate GenOps telemetry with GenAI identifiers for a unified observability view.
A runtime is GenOps v0.1 compliant if it:
- Emits all required attributes on every AI Workload Unit.
- Implements the Deterministic Accounting Invariant.
- Emits required governance events at defined lifecycle points.
- Uses decision states and reason codes exclusively from the specification.
- Maintains attribution immutability and identity uniqueness.
Compliance is verifiable by inspecting telemetry output. Partial compliance is permitted but must be declared explicitly.
See Section 9 of the specification for full details.
GenOps does not define:
- Policy authoring languages
- Control planes
- Budget coordination mechanisms
- Multi-cluster aggregation
- Storage formats
- SDK implementations
- UI requirements
- Model routing strategies
- Authentication or authorization
- Tracing topology, session/conversation identity, instrumentation APIs, or provider-specific telemetry schemas
These concerns may be built on top of GenOps, but are outside the scope of the specification.
The specification follows Semantic Versioning.
- Major versions introduce breaking changes.
- Minor versions may add attributes, events, or reason codes.
- Patch versions clarify language without altering semantics.
The current version is: v0.1.0
spec/
genops-spec-v0.1.md
GOVERNANCE.md
CONTRIBUTING.md
Future specification versions will be added under the spec/ directory.
GenOps is governed by a transparent model based on lazy consensus, with explicit voting and supermajority requirements for breaking changes. The governance model defines roles, decision thresholds, version governance, and conflict resolution.
- GOVERNANCE.md -- Project roles, decision model, specification lifecycle, version governance.
- CONTRIBUTING.md -- Change types, proposal requirements, review process, backward compatibility policy.
GenOps v0.1.0 is a Working Draft. Feedback from platform engineers, runtime implementers, and standards communities is welcome.