how to make agent decisions auditable?

To make agent decisions auditable in the Liminality architecture, you need a structured logging and routing mechanism that captures the "why" and "how" of every decision point. Here is the integrated approach based on your existing flows and assets:

1. **Use the Hypothesis-Branch Map Structure:** Treat the flow *"How should a research program's hypothesis-branch map be structured..."* as the foundational schema for decision logging. Record every agent decision as a node in this map, capturing: * The initial hypothesis or intent. * The branch taken (the specific sub-question or tool route selected). * The outcome (proven, falsified, or inconclusive). * The confidence score or evidence used to make the choice. This creates a persistent, queryable history of decision logic rather than just a log of actions.

2. **Integrate with the Multi-Task Prioritization Layer:** Use the flow *"We are adding a persistent multi-task prioritization layer to the Liminality engine..."* to log the context of *why* a specific decision was prioritized over others. The audit trail must include: * The set of competing tasks/subjects at the time of the decision. * The criteria used for prioritization (e.g., urgency, resource availability, strategic alignment). * The final selection and the discarded alternatives. This ensures that the *order* and *selection* of decisions are also auditable, not just the decisions themselves.

3. **Ground Decisions to Real Tools via Liminality MCP:** Since Liminality reduces questions to sub-questions and grounds them to real tools, the audit log must explicitly record: * The original high-level question. * The decomposed sub-questions. * The specific tool (from your connected toolkits: airtable, asana, figma, gmail, etc.) invoked. * The input parameters and output results of that tool invocation. This links the abstract decision to the concrete action, providing a complete chain of custody for the agent's behavior.

4. **Ensure Persistence Across Sessions:** Address the flow *"What is the best enforcement design so multi-session agent work always gets published (pushed/merged) instead of strandi…"* by ensuring that the decision logs from the hypothesis-branch map and prioritization layer are persisted to a durable store (e.g., Airtable or Google Drive) after each session. This prevents loss of decision context when agents switch sessions or contexts.

**Pending Action:** To implement this, you need to define the specific schema for the decision log nodes in your hypothesis-branch map and determine the storage location (e.g., Airtable base, Google Sheet) for the persistent multi-session logs. Would you like to proceed with designing this schema or setting up the storage integration?