Authors propose a formal framework assigning AI participation and assurance per workflow unit, executed on a 17-unit wing-spar workflow: AI-generated CAD passed all 23 deterministic checks at first attempt, while stress clauses were dispositioned ADJUDICATE for insufficient convergence evidence

The framework gives each workflow unit a participation and assurance record containing its engineering requirement, approved operational formalization with declared scope and version, applicable mechanism, declared fallback, assurance obligations with explicit evidence kinds, deployment-readiness status and regime guarantee clause. Prior work addresses generation capability, assurance of machine-learning constituents, organizational risk management, assurance cases, deterministic verification, statistical risk control, human oversight and tool qualification separately; the difference here is integrating these mechanisms and their limits at the level of individual workflow units, with explicit treatment of the step from requirement to operational formalization. Presented as definitions and propositions, including Type A per-instance logical implication, Type B population-level statistical statement, Type C procedural release property, and Type D/E participation boundaries; the authors state they did not identify a comparable integrated scheme in the sources reviewed.

A staged classification and readiness procedure separates participation, assurance and readiness, and keeps current regime classification, target regime attainable in principle, and deployment readiness as distinct statuses; Type A or Type B is assigned only where the verification or calibration mechanism actually exists. The procedure guarantees each workflow unit receives exactly one regime, and classification does not depend on readiness; readiness is a status, not a guarantee, and no guarantee clause is claimed for a NOT READY record. The authors describe these as consequences of the procedure's structure rather than substantive results about engineering work; decision predicates such as whether AI adds justified value, whether delegation is permitted and whether an approved formalization exists remain engineering and organizational judgements recorded with their justification.

In the executed 17-unit wing-spar structural-analysis workflow, the AI-generated CAD program entered the analysis only through deterministic admission and passed all 23 geometric and validity checks at the first attempt; downstream meshing, solving and extraction retained established deterministic tool paths; AI-produced drafts were released only through authorized human acceptance. The case demonstrates selective AI participation and unit-level evidence handling: 1 Type A, 6 Type D, 4 C[adjudicate] and 6 C[assist], with neither Type B nor Type E instantiated. A single author-led end-to-end execution with regimes, criteria and tools fixed before the run; a 61-entry hash manifest pinned specifications, criteria, material data, analysis plans and the solver binary; the solver analyses were executed once and not independently repeated.

Favourable numerical values did not automatically yield PASS: the flange and stiffener stress metrics did not converge, all four stress clauses returned NOT_CONVERGED at limit and ultimate load in both senses, and the final disposition was ADJUDICATE, while the buckling clauses received PASS because the evidence their rule required was present. The outcome surfaces the requirement-to-formalization gap in execution: the predeclared stress rule represented the root-clamp region but not the tip load-introduction region revealed by the run; the authors recorded this as a limitation and changed no criterion. Convergence tolerances were frozen before the run (1% for displacement and buckling factors, 5% for the stress metric); the flange metric rose from 182.6 MPa to 246.5 MPa (25.9%) and the stiffener from 98.4 MPa to 117.8 MPa (16.4%); a post-hoc diagnostic showed about 0.2% change in the interior flange but was not passed to the evaluator.

The framework is intended for engineering organizations and researchers who must decide how AI participates in safety-critical workflows, by which mechanism an AI-produced artefact may be admitted, and what happens when evidence is insufficient. It applies to engineering processes that can be decomposed into workflow units, each with a declarable requirement, formalization, mechanism and fallback; the case applies to a tapered, stiffened, cantilevered aluminium wing-spar segment idealized as a mid-surface shell model, evaluated against project-defined stress and buckling criteria under positive and negative tip loading. The authors state the framework derives no workflow-level dependability, independence between units, error-propagation bounds or system safety, and establishes no certification, regulatory compliance or productivity improvement.

A careful reader would still watch: the decision predicates rest on engineering and organizational judgement, and different decompositions may yield different allocations; formalization fidelity for natural-language requirements cannot be established by proof; Type B guarantees depend on a declared recurring population, a fixed generator version and stated calibration assumptions, and are unavailable for one-off units; Type C guarantees are procedural and do not ensure correctness of judgement; readiness depends on judging the adequacy of evidence. On the case, the structural model is idealized (mid-surface shell, linearized rigid tip-load relation without established finite-rotation validity, perfect-geometry linear buckling, project criteria rather than certification requirements); OS-level isolation of AI-generated code was not demonstrated, mesh determinism was shown on the same host only, the solver analyses were not independently repeated, and no comparable manual baseline exists, so no efficiency claim is made. This parse is a fast read, and some details in figures and supplementary material were not checked item by item; readers citing specific values or clauses should consult the original and its supplement.