Stable error taxonomy for host systems embedding predict-rlm

[axel-havard]

Summary

Embedded systems need structured error categories, not just freeform exceptions or logs.

Problem

A host integrating predict-rlm may need to distinguish:

• timeout
• cancellation
• capability denied
• capability failed
• model/provider failure
• recursion limit reached
• validation failure
• internal runtime failure

Without stable categories, orchestration and telemetry become brittle.

Proposed direction

Define and document a small stable error taxonomy and include it in:

• event stream
• final structured trace
• terminal result object

Why this matters

Hosts want to make decisions like:

• retry on transient model failure
• do not retry on capability denial
• treat timeout differently from cancellation
• classify runs accurately in dashboards and learning systems

Acceptance criteria

• documented error categories
• errors appear in structured outputs
• categories are consistent across event stream and final result