Classify

Intent

Assign one or more labels from a defined taxonomy to items based on their observed characteristics. This capability bridges detection and reasoning by providing semantic categorization.

Success criteria:

• Item assigned at least one label from taxonomy
• Label assignment supported by evidence
• Confidence scores reflect classification certainty
• Ambiguous cases explicitly flagged

Compatible schemas:

• schemas/output_schema.yaml

Inputs

Procedure

1. Examine the item: Gather characteristics relevant to classification

   • Identify distinguishing features
   • Note structural patterns, content type, metadata
   • Collect evidence for each observed characteristic

2. Understand the taxonomy: Clarify the classification scheme

   • If taxonomy provided, use those labels exclusively
   • If no taxonomy, infer appropriate categories from context
   • Define clear boundaries between categories

3. Match characteristics to labels: Evaluate fit for each potential label

   • Score how well item characteristics match each category
   • Consider edge cases and borderline classifications
   • Note which features drive each potential classification

4. Assign labels: Select the most appropriate label(s)

   • For single-label: choose highest confidence match
   • For multi-label: include all labels above confidence threshold
   • Flag if no label is a strong match

5. Ground classification: Document evidence supporting each label

   • Reference specific characteristics that drove classification
   • Note any characteristics that contradict the assignment

Output Contract

Return a structured object:

labels:
  - label: string  # Assigned category/label
    confidence: number  # 0.0-1.0 for this specific label
    evidence: array[string]  # Characteristics supporting this label
probabilities:  # Optional: full probability distribution
  label_1: number
  label_2: number
primary_label: string  # Highest confidence label
ambiguous: boolean  # True if classification is uncertain
confidence: 0..1  # Overall classification confidence
evidence_anchors: ["file:line", "characteristic:value"]
assumptions: []

Field Definitions

Examples

Example 1: Classify Code File

Input:

item: "src/services/payment_processor.rb"
taxonomy: ["service", "model", "controller", "helper", "concern"]

Output:

labels:
  - label: "service"
    confidence: 0.92
    evidence:
      - "Located in services/ directory"
      - "Class name ends with 'Processor'"
      - "Contains business logic methods without persistence"
probabilities:
  service: 0.92
  model: 0.03
  controller: 0.02
  helper: 0.02
  concern: 0.01
primary_label: "service"
ambiguous: false
confidence: 0.92
evidence_anchors:
  - "src/services/payment_processor.rb:1"
  - "src/services/payment_processor.rb:15-45"
assumptions:
  - "Directory structure follows Rails conventions"

Example 2: Classify with Ambiguity

Input:

item: "User reported issue: 'App crashes on login'"
taxonomy: ["bug", "feature-request", "question", "documentation"]
multi_label: true

Output:

labels:
  - label: "bug"
    confidence: 0.75
    evidence:
      - "Reports crash behavior"
      - "Describes unexpected failure"
  - label: "question"
    confidence: 0.40
    evidence:
      - "Lacks reproduction steps"
      - "May be user error or configuration"
probabilities:
  bug: 0.75
  question: 0.40
  feature-request: 0.05
  documentation: 0.10
primary_label: "bug"
ambiguous: true
confidence: 0.65
evidence_anchors:
  - "issue:title"
  - "issue:body"
assumptions:
  - "Crash is not expected behavior"
  - "User has attempted normal login flow"

Verification

• At least one label assigned with confidence > 0.3
• Evidence exists for each assigned label
• Labels are from specified taxonomy (if provided)
• Ambiguous flag set when confidence < 0.7
• Probabilities sum to ~1.0 (if provided)

Verification tools: Read (to verify evidence references)

Safety Constraints

• mutation: false
• requires_checkpoint: false
• requires_approval: false
• risk: low

Capability-specific rules:

• Do not invent labels outside the provided taxonomy
• Flag uncertainty rather than forcing low-confidence classifications
• Do not access data beyond what's needed for classification
• Note when item characteristics are insufficient for reliable classification

Composition Patterns

Commonly follows:

• detect - Classify items after detecting their presence
• observe - Classify based on observed characteristics
• retrieve - Classify retrieved items

Commonly precedes:

• compare - Classification enables comparison within categories
• plan - Classified items inform planning decisions
• generate - Classification guides content generation

Anti-patterns:

• Never use classify for binary detection (use detect)
• Avoid classify when precise measurement needed (use measure)

Workflow references:

• See reference/workflow_catalog.yaml#capability_gap_analysis for classification usage