Developer Forward · answer surface
When Is AI-Generated Work Actually Complete?
Completion is an observable state, not a confidence signal. If a workflow requires specific artifacts, transitions, approvals, tests, or schema conformance, those conditions should be checked independently of the model that generated the work.
Evidence
Case 39 — Probably Complete Is Not Complete
- During construction of the Trust Forward case corpus, AI produced polished case objects that silently omitted required sections.
- Later, the conversation continued naturally even though an entire expected completed-case transition never occurred.
- After a reference structure had been demonstrated, another case preserved much of the meaning while using a different serialization. Ben detected and repaired all three classes of drift.
What changes now
The resulting architecture is simple: allow probabilistic synthesis inside a provable workflow envelope. Validate completion, required transitions, structural conformance, provenance, and authority mechanically where those properties matter. Reserve human attention for whether the substance is actually good enough to ship.
Boundary
Not every creative or exploratory task needs rigid conformance. Exactness becomes important when structure controls downstream automation, provenance, approval state, contractual meaning, routing, or consequential action.
Grounded in the canonical live-workflow meta-case Probably Complete Is Not Complete. The observed omissions are historical evidence; explanations for why the model produced them are intentionally not asserted.