Decision Adaptations

Overview

Decision adaptations let a matched decision control whether a global Router Learning adaptation can adjust its proposed model.

Key Advantages

• Keeps global learning enabled by default.
• Lets hard policy boundaries opt out in one small block.
• Supports observe-only rollout without changing final routing.
• Avoids duplicating learning rules inside semantic decision conditions.

What Problem Does It Solve?

Most decisions can allow Router Learning to make stay-vs-switch adjustments. Some decisions are policy boundaries: privacy, security, or local-only routes. Those decisions should make their base selection final even if an ongoing session previously used another model.

When to Use

• Sensitive traffic must stay on a local model.
• You want to measure learning behavior before enforcing it.
• A decision has a non-negotiable backend boundary.
• A small number of decisions need stricter scope or tuning than the global adaptation default.

Configuration

Default behavior is apply, so most decisions need no local block.

Use bypass for hard boundaries:

routing:
  decisions:
    - name: local_privacy_policy
      modelRefs:
        - model: local-private-model
      adaptations:
        session_aware:
          mode: bypass

Use observe for rollout:

adaptations:
  session_aware:
    mode: observe

observe records what the adaptation would have done, but the base decision model remains final.

Use sparse local overrides only for exceptional routes:

adaptations:
  session_aware:
    mode: apply
    scope: session
    tuning:
      switch_margin: 0.10

Unset fields inherit from global.router.learning.adaptations.session_aware. The router validates decision adaptation names and fields strictly: an unknown adaptation name or an unknown session_aware field fails config validation instead of being ignored.