06 — Classification

How Triage Classifies

Triage evaluates every incoming piece of evidence across multiple dimensions:

Dimension 1: Source and Provenance

• Where did this come from? (MCP, conversation, agent, manual)
• Who produced it? (user, AI, system)
• Is the source trusted? (known connector vs unknown)

Dimension 2: Semantic Category

• What type of knowledge is this?
• Categories: insight, fact, decision, action_item, preference, relationship
• Domain: executive, product, engineering, design, sales, marketing, etc.

Dimension 3: Entity Linking

• Which known entities does this relate to? (accounts, contacts, deals, etc.)
• Is this a new entity or an update to an existing one?
• Confidence of the link

Dimension 4: Duplication and Conflict

• Does this duplicate existing canonical memory?
• Does it conflict with established truth?
• Is it stale (old information presented as current)?

Dimension 5: Sensitivity and Materiality

• How sensitive is this information?
• Is it material (business-significant) or trivial?
• Does it require governance review?

Dimension 6: Memory Scope

• Should this be personal, conversational, agent, or organizational scope?
• Does it meet the threshold for organizational memory?

Classification Output

Every triage evaluation produces:

{
  "classification": {
    "doc_type": "lead|document|policy|code|... ",
    "workflow_type": "sales|marketing|engineering|... ",
    "intent": "description of intent"
  },
  "linking": {
    "primary_entity": "entity_id",
    "related_entities": ["entity_id_1", "entity_id_2"]
  },
  "canonical_review": {
    "is_duplicate": false,
    "has_conflict": false,
    "conflict_reason": null
  },
  "memory": {
    "scope": "organizational",
    "candidate": true
  },
  "governance": {
    "decision": "stage",
    "reason": "High-value decision with clear entity links"
  }
}

Routing Decisions

Decision	Meaning	Action
stage	Candidate for review	Write to triage_results as pending
reject	Below threshold or spam	Terminal, no memory
defer	Needs more context	Return to review queue
repair	Incomplete, fixable	Route for repair then re-triage

Confidence Dimensions

Confidence is six things, and one number must never represent all six:

1. Pattern confidence — How well does this match known patterns?
2. Prediction confidence — How likely is this to be accurate?
3. Input completeness — How complete is the source data?
4. Data freshness — How current is the information?
5. Historical precision — How accurate has this source been?
6. Policy eligibility — Does this meet policy requirements?

Small samples use a named conservative lower bound, not raw accuracy. 3/3 is not better evidence than 470/500.