Reason

The Reason layer executes agents and rules against artifacts, producing replayable decision traces with full model metadata.

Responsibilities

• Agent planner with goals, tools, constraints
• Policy guardrails (OPA/Cedar)
• Prompt templating with tests
• Deterministic execution (pinned models, seeds, params)

Decision trace (YAML)

trace_id: <uuid>
inputs: [<artifact_id>]
policy: gdpr.redact.v1
model:
  name: llm-x
  temperature: 0.3
  top_p: 0.9
  seed: 12345
prompt: |
  <prompt text>
tools:
  - search
  - calculator
outputs:
  summary: <text>
  citations: []
cost:
  tokens: 1234
  usd: 0.004
timestamp: 2025-10-03T12:34:56Z

Testing & replay

• Golden prompts and fixtures
• Scenario sweeps (seed grid, temperature bounds)
• Re-run under pinned artifacts/models