Neuro-symbolic execution: inference proposes, codegen commits

A ReAct loop interleaves model reasoning and tool calls in one control path. That pattern is useful for exploratory tasks, but it does not by itself establish a transaction boundary. EGI separates neural inference from symbolic execution. Inference may misclassify intent; the execution layer must still reject an unauthorized write. A schema failure returns to a defined recovery path rather than allowing the model to reinterpret policy.

System 1 — neural inference

Operational input is often inconsistent: email threads, supplier PDFs, exception messages, and ERP notes may disagree with the open purchase order. Neural inference identifies the job step, entity, and likely intent. This component is replaceable, so a model change is evaluated against the same labeled intents and job specifications before it reaches the write path.

System 2 — symbolic codegen

Symbolic execution produces code or a typed invocation constrained by the tenant schema, acting identity, and control plane. If a required field was not resolved at the binding checkpoint, the action fails closed. Idempotency keys, legal entity, posting period, and approval state are enforced as executable constraints rather than advisory prompt text.

• Inference may propose. Codegen commits—or refuses.
• Refusal is a first-class outcome: missing bind, policy miss, schema miss, approval required.
• Retry is recovery in the execution graph, not a longer ReAct trace with the same illegal args.
• Audit logs the codegen artifact and the system change, not the chain-of-thought.

Enforcing the transaction boundary

When one model selects the tool, supplies arguments, and judges the result, a plausible response can mask an incorrect transaction. Common examples include a valid API response for the wrong object or a coerced value accepted by the ERP. EGI keeps validation on the symbolic side, where the resulting system state is checked against the job specification.

ERP work makes this separation concrete. Unstructured inputs inform the decision, while the committed transaction becomes part of the company’s operational record. Warehouse, operations, and revenue workflows use the same boundary: inference interprets the input; symbolic execution applies the authorized change; workflow state records completion or recovery.