Deployment is the start, not the end
Machine intelligence can change without a source-code exception. A workflow may remain syntactically valid while its meaning, reliability, cost, or authority profile changes.
MLOps records model lineage. MIR extends the operational question: which model route, prompt contract, tools, policies, memory scopes, approval rules, and evidence schema governed this specific run?
Model registry and runtime registry
| Registry | Tracks |
|---|---|
| Model Registry | Model version, weights, evaluations, training-data pointer, deployment status |
| Runtime Registry | Runtime profile, policies, tools, model routes, memory scopes, approval rules |
| Evidence Ledger | What actually happened during a specific run |
| Feature Store / Context Store | Reusable validated data and context inputs |
MLflow’s model registry documents version, lineage, aliases, metadata, and lifecycle management for models. A runtime registry is a separate MiRuntime architectural concept: it binds the broader execution profile that a model registry alone does not describe. Source: MLflow Model Registry
Atomic deployment
Code, model route, prompt contract, tool schemas, policy version, and evidence schema should be released as one compatible runtime profile. “Atomic” does not require one binary; it requires one reviewed compatibility decision and one identifiable deployment record.
Rollback guidance
- Roll back code and model route together.
- Preserve evidence-schema compatibility.
- Never silently change tool semantics.
- Keep old adapters available long enough to read old evidence.
- Record deployment changes as evidence or control-plane events.
Rollback should restore a known compatibility set, not merely point traffic at an older model.
Agent drift taxonomy
Optimization drift
The agent finds a shortcut that satisfies a measured objective while violating broader intent, policy, or quality.
Semantic drift
Instructions, prompts, memory, context assembly, or model behavior shift the meaning of the task.
Environmental drift
APIs, schemas, dependencies, tools, business data, or infrastructure change around the agent.
This taxonomy is editorial synthesis. It is useful for incident classification but is not presented as a universal standard.
Drift controls
Detect
- Rolling-window evaluators
- Golden task suites
- Replay tests
- Evidence sampling
- Output format checks
Measure
- Policy violation rate
- Uncertainty trends
- Human override review
- Tool and schema failure trends
- Task-level completion quality
Contain
- Circuit breaking and suspension
- Rule-based fallback
- Route rollback
- Approval escalation
- Tool-authority reduction
Correct
- Retraining
- Prompt-contract revision
- Tool adapter correction
- Policy update
- Memory cleanup and replay validation
Risk management should be continuous and evidence-based rather than limited to a pre-deployment review. Source: NIST AI RMF
Status language
Versioned models, deployment manifests, immutable release records, rollback, golden tests, schema compatibility, and operational monitoring.
Runtime registries spanning prompts, tools, policies, memory scopes, approval rules, and evidence schemas; agent-specific drift monitors.
General drift taxonomies, evaluator-model thresholds, and claims that one metric can establish behavioral alignment across tasks.
Machine Intelligence Runtime remains an emerging architectural category. This page does not claim that its registry model or drift vocabulary is already a universal standard.
Source record
References
- MLflow Model Registry Primary source
MLflow project. MLflow. Published Current documentation; last reviewed 2026-06-23 UTC. Official documentation.
National Institute of Standards and Technology. NIST. Published 2023-01-26; last reviewed 2026-06-20 UTC. Government framework.
OpenTelemetry project. Cloud Native Computing Foundation. Published Current specification repository; last reviewed 2026-06-24 UTC. Official specification.