NOTE: switching from article → needs-human because score is 62 (below the 85 threshold for auto-approved articles, and below the 85+ai-audit product fit gate for article touch). DARPA CLARA is a federal grant program — outbound article framing doesn't map cleanly to a grantee audience. A targeted reply to any HN or Dev.to discussion about CLARA would be more appropriate. Human to decide whether to rework as a reply or drop.

---

why compositional reasoning makes ai audit harder, not easier (draft — not for publish)

DARPA's CLARA program funds research into AI systems with "strong logical explainability" — compositional learning and reasoning tightly integrated with machine learning. the $2M per award, 24-month timeline signals this is foundational, not applied.

the audit challenge CLARA creates: when a system uses compositional reasoning, the decision path isn't a single model call you can log. it's a chain of reasoning steps, each potentially involving different model components, each needing its own trace record.

existing observability stacks log inputs and outputs. they don't log the intermediate reasoning steps in a way an auditor can reconstruct the decision from. that's the gap BizSuite AI Audit is built to close — append-only trace logs at the reasoning layer, not just the API call layer.

if you're building toward CLARA requirements or advising teams who are: https://getbizsuite.com/ai-audit