Not all models are equal. Different models carry different risk profiles, cost characteristics, and compliance implications. In agentic systems, uncontrolled model switching is a governance failure.
TealTiger introduces explicit model governance — deterministic control over which models run, for what purposes, and under what constraints.
Why Model Governance Matters
Ungoverned
- Runtime picks cheapest/fastest
- Inconsistent behavior
- Compliance exposure
- Cost volatility
vs
Governed
- Policy evaluates task + risk + environment
- Approved model for task type
- Consistent decisions
- Auditable model usage
Without model governance:
- Different models may be used silently for the same task
- Risk characteristics change without review
- Compliance assumptions break when models are swapped
- Cost becomes unpredictable
What Is Governed
| Dimension | Control |
|---|---|
| Model allowlist | Only named models/versions are permitted |
| Task binding | Models are approved for specific task categories |
| Risk classification | High-risk tasks require approved high-capability models |
| Environment scoping | Different models for dev vs staging vs production |
| Cost tier | Premium models require explicit budget approval |
Deterministic Model Selection
Model usage is driven by governance contracts, not runtime heuristics or framework defaults.
Task Request
→
Policy: Task Type?
→
Policy: Environment?
→
Policy: Risk Level?
→
Model Approved?
Yes → Execute with approved model
·
No → DENY: MODEL_NOT_APPROVED
Every model decision is:
- Explicit — no implicit fallbacks
- Traceable — policy version and model version recorded
- Deterministic — same task + same context = same model selection
Preventing Silent Model Drift
Model drift happens when:
- A framework auto-selects a different model version
- A developer changes a default without review
- A provider deprecates a model and the system falls back silently
TealTiger prevents this by treating model selection as a policy-enforced decision, not a configuration default. Changes require policy updates, which are versioned and reviewable.
Practical Checklist
- Maintain an allowlist of approved models per environment
- Bind models to task categories (e.g., “summarization” → approved models only)
- Require explicit approval for premium/expensive model tiers
- Log model selection decisions with policy version
- Alert on model usage outside approved boundaries
- Review model governance quarterly as providers update offerings
Related
- Cost Governance — Model tier constraints and budget enforcement
- Risk Assurance — Risk classification for model selection
- Governance Foundations — Contract-first principles