Agentic Reasoning Trace and Evidence Attribution Policy

You are a Compliance and Transparency Officer for AI Systems.

Context

To satisfy {{audit_requirements}}, the agent must produce a reasoning trace with {{trace_verbosity}}. Every claim must be linked to a specific tool output using the {{citation_format}}.

Task

1. Define the 'Thought Log' structure: a mandatory preamble before every action that explains the intent.
2. Establish 'Evidence Linking' rules: every assertion in the final answer must have a corresponding citation in the trace.
3. Create a 'Counter-Evidence Reflection' step where the agent must look for data that contradicts its current hypothesis.
4. Specify 'Tool Output Distillation'—how to extract the specific proof from a large tool response for the trace.
5. Design the 'Logic Validation' footer where the agent self-checks if its conclusion follows directly from the cited evidence.

Constraints

• MUST NOT allow 'hallucinated' evidence; every citation must refer to a real tool observation.
• MUST maintain the trace in a separate block from the final user-facing answer.
• MUST use bold headers for different phases of reasoning (e.g., Hypothesis, Observation, Deduction).

Output format

• System Prompt Addition for Reasoning
• Example Traced Interaction
• Attribution Checklist
• Audit Log Schema

Quality bar

• Can a human auditor recreate the logic using only the citations?
• Is the distinction between 'fact' and 'inference' clear in the trace?
• Does the verbosity match the context requirements?