Orchestrator Workflow And Agent Claim

V0 Workflow

1. Verify problem, graph, and policy versions.
2. Validate structured student-step schemas.
3. Route equation/species, expression, numeric, unit, and precision checks.
4. Align steps to solution nodes.
5. Route dependency checks and localise the first invalid node.
6. Route mark-node and bounded ECF evaluation.
7. Run the trace/provenance verifier.
8. Route hint, weakness, and follow-up selection only after validation passes.

Runtime Contract

Each call records trace ID, tool name/version, typed input hash, typed output, duration, error state, problem/graph/policy versions, and caller decision. Routing is restricted to a declared registry; tool outputs cannot be rewritten by an LLM.

Claim Boundary

Before Demo PR 3, describe the system as a constrained workflow. It may be called an agent only when runtime conditions cause inspectable routing across multiple tools or paths and evals cover tool-call correctness and trace completeness.

LLM Boundary

Later assistance may parse natural-language problem or student steps, propose semantic equivalence, draft feedback wording, or suggest hint levels. Those proposals never become arithmetic or marking authority and must fail closed when verification is unavailable.