Projections

Overview

Projections sit between signal extraction and decision matching. They resolve competition among signals, combine several signals into a score, and turn that score into named outputs a decision can reference.

The route pipeline is:

1. routing.signals extracts facts.
2. routing.projections coordinates or derives facts.
3. routing.decisions matches policy and chooses candidate models.

Key Advantages

• Reuses one coordination or threshold policy across several decisions.
• Keeps numeric aggregation out of boolean decision trees.
• Preserves named, explainable outputs for replay and debugging.

What Problem Does It Solve?

Individual signals are intentionally narrow. Real routing policy often needs one winner from a competing intent group or one reusable difficulty score from several weak indicators. Without projections, that logic is duplicated across decisions and numeric thresholds become hard to audit.

When to Use

Use projections when:

• only one member of a domain or embedding group should remain active
• several signals should contribute to one continuous score
• several decisions should share the same named threshold bands

Skip projections when one raw signal expresses the route condition clearly or when multiple matches should remain independently visible.

Projection Types

Type	Goal	Decision-visible?	Guide
partitions	Keep one winner from competing domain or embedding signals	No; decisions still reference the winning raw signal	Partitions
scores	Combine signal values with weighted_sum	No; scores feed mappings or other scores	Scores
mappings	Convert a score into named threshold outputs	Yes, through type: projection	Mappings
trace	Explain partition, score, and mapping results in Router Replay	Operational only	Projection Traces

Current methods are:

• partition semantics: exclusive and softmax_exclusive
• score method: weighted_sum
• mapping methods: threshold_bands (first matching output) and multi_emit (every matching output; requires at least two outputs)
• optional mapping calibration: sigmoid_distance

Configuration

routing:
  signals:
    embeddings:
      - name: technical-support
        threshold: 0.75
        candidates: [installation help, troubleshooting]
      - name: account-management
        threshold: 0.72
        candidates: [billing issue, subscription change]
    context:
      - name: long-context
        min_tokens: 4K
        max_tokens: 200K

  projections:
    partitions:
      - name: support-intents
        semantics: exclusive
        members: [technical-support, account-management]
        default: technical-support
    scores:
      - name: request-difficulty
        method: weighted_sum
        inputs:
          - type: embedding
            name: technical-support
            value_source: confidence
            weight: 0.5
          - type: context
            name: long-context
            weight: 0.5
    mappings:
      - name: difficulty-band
        source: request-difficulty
        method: threshold_bands
        outputs:
          - name: support-fast
            lt: 0.5
          - name: support-escalated
            gte: 0.5

  decisions:
    - name: escalated-support
      description: Route difficult support requests to the larger model.
      priority: 150
      rules:
        operator: AND
        conditions:
          - type: projection
            name: support-escalated
      modelRefs:
        - model: support-large

Only mapping output names are referenced with type: projection. Decisions do not reference partition or score names directly.

The DSL exposes the same three concepts through PROJECTION partition, PROJECTION score, and PROJECTION mapping blocks. The Dashboard exposes them under Config > Projections.

Dependencies and Limitations

• Projections make no additional model or storage calls; they consume signal results already computed for the request.
• A partition default is a fallback, not evidence that its member matched.
• Weighted sums do not automatically calibrate inputs from different signal families. Evaluate weights and mapping bands together on labeled traffic.
• Cycles between derived scores are rejected during validation.
• See the balance recipe for an end-to-end example.