How do tool-constrained agents like Diagrammatica recover when the trusted backend rejects an invalid diagram specification?

Based on the preprints provided, I can give you a direct answer about Diagrammatica's error recovery.

When the trusted backend rejects an invalid diagram specification in Diagrammatica, the agent employs tool-constrained computation as its primary recovery mechanism [AG-2026.03-1733]. Here's how it works:

The architecture concentrates the agent's action distribution onto tool calls with "convention-fixing semantics," meaning the agent specifies a compact, human-auditable diagram specification, and the backend performs exact symbolic or numerical manipulations [AG-2026.03-1733]. This design choice is deliberate: it offloads the burden of mathematical correctness from the LLM (which is error-prone) to a trusted computational system that enforces the rules.

When a diagram specification is invalid, the rejected specification itself becomes informative. Because the diagram syntax is compact and human-auditable, the agent can inspect *why* it failed—the mathematical or notational conventions it violated are encoded in the backend's error feedback. The agent can then reformulate the specification to comply with those constraints and resubmit [AG-2026.03-1733].

This is complemented by the toolkit's two calculation paths (Naive Dimensional Analysis and Exact Diagrammatic Analysis), which consume the same diagram specification [AG-2026.03-1733]. If one path fails, the agent can fall back to the other or adjust the specification based on which path succeeded.

The key insight: the agent doesn't need to "understand" physics conventions; it only needs to learn which diagram specifications the backend accepts, via iterative rejection and refinement.