OPTIMIZATION & PORTABILITY
Canonical vs optimized AI behavior
Why model-specific optimization should improve execution while canonical behavior remains the governed authority and fallback.
DefinitionCanonical behavior is the governed representation produced from the effective behavior program. An optimized projection is a model-target execution artifact that is valid only for the exact behavior identity and dependencies it was evaluated against.
KEY TAKEAWAYS
- Canonical behavior answers the authority question; optimization answers an execution question.
- An optimized projection is valid only for the exact behavior and model identity it was evaluated against.
- Stale or mismatched optimized artifacts should fall back to the governed canonical path.
Optimization answers an execution question
Different models can respond better to different instruction forms. An optimizer may therefore produce a model-target projection intended to improve execution on a particular target.
That artifact is not a new behavior policy. It remains attached to the exact profile, source revisions, composition digest, artifact identity, and model target used during evaluation.
Staleness is a safety property
If a Lesson changes, a binding changes, an ancestor authority changes, a signal registry changes, or the target identity no longer matches, an optimized artifact can become stale.
Stale optimized output should not be silently reused. When the canonical path is valid, it provides a governed fallback while a new optimized artifact is evaluated.
Optimization needs its own receipt
When optimized execution is selected, a runtime receipt should identify the deployment, source revision set, composition, projection artifact, artifact digest, and model target.
If the system cannot prove which optimized artifact executed, it should not present that execution as trusted optimized authority.
SOURCE & LIMITATION NOTE
Product and Docs links support Vira architecture descriptions. External references are labeled separately. None of them should be read as benchmark evidence unless a dated methodology and result set is published.
REFERENCES