PaperlayerView exact revision on arXiv

Source-linked AI summary

Dependable AI-Assisted Engineering: A Formal Framework for AI Participation and Assurance in Safety-Critical Workflows

Puxue Tan

arXiv:2610.04084v1cs.CVcs.AI

TL;DR

Safety-critical engineering needs evidence-based rules for admitting AI-produced artefacts, and those rules may differ across workflow units. The paper develops a unit-level framework for assigning participation and assurance, then applies it in an executed wing-spar analysis workflow. In that case, favourable stress values did not yield a pass without required convergence evidence, and unresolved criteria went to engineering judgement.

  • Problem

    Generation alone does not establish an AI-produced artefact’s acceptability, while workflow units may require different participation and assurance arrangements.

    Source: §1.1 · pp. 1–2, §1.2 · p. 2

  • Method

    The framework records each unit’s requirement, applicable formalization and assurance mechanism, fallback, evidence obligations, guarantee and deployment readiness.

    Source: §7 · p. 30

  • Results

    In the wing-spar case, favourable stress values did not yield PASS without required convergence evidence, and unresolved stress criteria were referred to engineering judgement.

    Source: §5.5 · p. 25, §7 · p. 30

  • Takeaways & Limitations

    The case demonstrates selective AI participation and explicit evidence handling at unit level, with deterministic admission for AI-generated CAD and retained deterministic tools.

    Source: Case-study limitations · p. 28, §7 · p. 30

    The framework derives no workflow-level dependability or error-propagation bound, and its unit-level comparison result applies only under stated recurring-population assumptions, not to one-off decisions.

    Source: Framework-level limitations · p. 27

Abstract

from arXiv · show

Generative AI can produce engineering artefacts, but generation alone does not determine whether or how those artefacts should enter safety-critical workflows. This paper develops a formal framework for assigning AI participation and assurance at the level of individual workflow units. Each unit has a participation and assurance record covering its engineering requirement, an approved operational formalization where applicable, the applicable mechanism, fallback where applicable, evidence obligations and the applicable guarantee, plus a deployment-readiness status. The framework distinguishes deterministic verification, statistically calibrated admission, authorized human judgement supported by AI advice, authorized human adjudication of AI-produced artefacts, retained deterministic tool paths and explicit non-participation; these arrangements carry different kinds of guarantee rather than levels on a common scale. The framework also separates formalization fidelity from verifier soundness, provides a staged classification and readiness procedure, and derives conditions for comparing a gated AI-assisted unit with an incumbent process under recurring-population assumptions. We instantiate and apply the framework in an executed 17-unit wing-spar structural-analysis workflow combining deterministically gated AI-generated CAD, retained deterministic computation and human judgement. The AI-generated CAD program passed all 23 deterministic checks and was admitted at the first attempt. Favourable stress magnitudes did not suffice to pass the stress criteria where the predeclared mesh-convergence evidence was insufficient; those criteria were instead referred to engineering judgement. The case demonstrates selective AI participation and explicit evidence handling at unit level; no claim is made of workflow-level dependability, certification, structural safety or productivity.

1 Introduction

Generation alone does not establish that an AI-produced artefact is acceptable for safety-critical engineering use; different workflow units may require different participation and assurance arrangements. The paper responds with a unit-level framework, a staged classification and readiness procedure, and an executed wing-spar demonstration.

1.1 AI participation in safety-critical engineering workflows

  • Safety-critical workflows need traceable, reviewable evidence linking engineering decisions and deliverables to requirements; generation alone does not supply an acceptance rule or a course of action when evidence is insufficient.The appropriate participation and assurance can differ across units, even for the same AI system.Source: §1.1 · pp. 1–2

1.2 Beyond binary trust

  • The framework treats AI adoption as a unit-level decision, distinguishing deterministic checks, retained tools, human judgement and calibrated statistical admission rather than a binary trust choice.These arrangements carry different guarantees, not positions on one assurance scale.Source: §1.2 · p. 2

1.3 Contributions and paper organization

  • Each unit’s record captures its requirement, applicable formalization and mechanism, fallback, evidence obligations, guarantee and deployment readiness.Formalization approval is recorded but does not prove fidelity to the full informal requirement.Source: §1.2 · p. 2, §1.3 · p. 3
  • The staged procedure separates whether AI participates, which assurance mechanism is available, and whether its obligations are evidenced; assignment of deterministic or statistical admission requires the relevant mechanism to exist.The current regime, attainable target regime and deployment readiness are recorded as distinct statuses.Source: §1.3 · p. 3
  • In the executed wing-spar workflow, AI-generated CAD entered through deterministic admission, established tools were retained, and human judgement addressed evidence that did not resolve a criterion.Favourable numerical values did not yield a pass without required convergence evidence.Source: §1.3 · p. 3

2 Background and Related Work

The reviewed literature provides complementary approaches to generation, assurance, verification, risk control, oversight and tool use, but addresses different objects and questions. The paper identifies their integration for individual engineering workflow units as an unaddressed gap in the sources reviewed.

2.1 AI-generated engineering artefacts

  • AI-generated CAD and engineering artefacts are evaluated for geometric, topological or executability properties, but those measures do not establish workflow admission rules or what happens when an artefact is rejected.Examples include DeepCAD (Wu et al., 2021), CAD-MLLM (Xu et al., 2024) and FreeCAD-script generation.Source: §2.1 · p. 3

2.2 Assurance, governance and autonomy-level approaches

  • Aviation guidance, organizational risk frameworks and assurance cases address applications, organizations, products or machine-learning constituents, while Bloomfield and Rushby (2025) describe architectural compensation for weakly assured components.The paper does not claim a correspondence between those approaches and its participation-and-assurance regimes.Source: §2.2 · pp. 3–4

2.3 Deterministic verification and formalization fidelity

  • A deterministic checker can assess an artefact against a formalized property, but checker soundness does not establish that the formalization captures the informal requirement.Tools and traceability practices support formalization fidelity without proving it.Source: §2.3 · p. 4

2.4 Statistical acceptance and selective risk

  • Selective prediction and risk-controlling calibration provide foundations for statistical admission, whose guarantees depend on sampling assumptions, generator configuration and sample size.The paper’s construction jointly bounds selective risk and requires minimum coverage at stated confidence, unlike the reviewed approaches that maximize or report coverage.Source: §2.4 · p. 4

2.5 Human oversight, established tools and model adequacy

  • Human acceptance provides authority but not correctness, while retaining a deterministic tool path provides repeatability without establishing validation, qualification or model adequacy.Automation bias and complacency can affect both naive and expert users, and training alone does not prevent automation-bias errors.Source: §2.5 · p. 5

2.6 Synthesis and research gap

  • The sources reviewed offer mature mechanisms but no identified scheme integrating their distinct limits and evidence for each workflow unit; the paper formalizes that integration without treating dependability as a scalar.The paper positions its framework as an integration of existing mechanisms rather than a replacement for them.Source: §2.6 · p. 5

3 Formal Framework

The framework assigns each workflow unit an AI-participation arrangement and the assurance needed to justify it, distinguishing participation boundaries and guarantee kinds rather than using a common scalar scale. It separates regime classification and evidence-backed deployment readiness from the guarantee claimed when ready, distinguishes verifier soundness from formalization fidelity, and gives conditional unit-level criteria for comparing gated AI-assisted work with an incumbent process.

3 Formal Framework

  • Types A–E distinguish deterministic admission, population-calibrated admission, authorized human assistance or adjudication, retained deterministic paths, and justified non-participation; their guarantees are not ordered on a common scale.AI-produced deliverables require the regime’s prescribed acceptance mechanism; under C[assist], an authorized person produces the deliverable with AI limited to labelled advice.Source: §3 · p. 5, The record · p. 7

3.1 Activities, Workflow Units, Requirements and Formalizations

  • Workflow units, not necessarily whole activities, are classified; their requirements may be noncomputable or admit expert disagreement, and an approved formalization captures only a declared scope of the requirement.Guarantees are conditional on inputs meeting the unit’s valid-input assumptions; formalization fidelity is distinct from verifier soundness.Source: §3.1 · pp. 5–6

Gate, fallback, and released output

  • A gate admits or rejects an AI-produced artefact and rejection routes execution to a declared fallback, but does not ensure that the fallback produces a deliverable or that it is correct.Fallback availability and quality must be evidenced separately where they matter.Source: Gate, fallback, and released output · p. 6

The record

  • The versioned assurance record specifies the requirement, any approved formalization, regime, mechanism, applicable fallback, obligations, guarantee, and readiness; AI may participate only under a READY record permitting it.Readiness requires obligations to be supported by adequate evidence; proof obligations are discharged, while empirical, organizational, and engineering obligations are satisfied with stated evidence.Source: Evidence obligations · pp. 6–7, The record · p. 7

3.3 Type A — Deterministically Gated Units

  • Type A admits AI-generated artefacts only through a deterministic verifier that is sound relative to an approved formalization, so its guarantee does not establish fidelity to the underlying requirement.Completeness and availability are not guaranteed; generation–verification asymmetry may justify AI participation, but is not part of the guarantee.Source: §3.3 · p. 9

3.4 Type B — Statistically Gated Units

  • Type B uses calibration against trusted reference evaluations to admit artefacts with a statistical guarantee for a declared population and fixed generator version, not per-instance correctness.Its guarantee bounds selective error while requiring positive coverage; population or generator changes require recalibration, and one-off units cannot qualify.Source: §3.4 · pp. 9–10

3.5 Type C — Human Judgement and Adjudication

  • Type C requires an identified, authorized person to release the deliverable or decision, either by producing it with AI advice or explicitly accepting an AI-produced draft.Its guarantee is procedural: authorized release and identifiable AI contributions, not correctness of the released work.Source: §3.5 · p. 11

3.6 Types D and E — Retained Deterministic Paths and Non-Participation

  • Type D retains an established deterministic tool path with validation evidence and a stated validity domain when AI adds no justified value; Types D and E exclude AI-produced artefacts from the deliverable.Type D is retention of a deterministic path, not verified AI.Source: §3.6 · pp. 11–12
  • Type D retains an established deterministic tool path, whose correctness, validation evidence, and any actual qualification do not establish engineering-model adequacy.Mesh adequacy, idealization, and boundary conditions require their own workflow-unit records.Source: §3.6 · p. 12
  • Type E is justified non-participation: no AI-produced artefact contributes to the deliverable, whereas useful AI advice followed by a human-produced deliverable is C[assist].The auditable obligation is to record that no AI artefact contributed.Source: §3.6 · p. 12

3.7 Relations Among the Participation and Assurance Regimes

  • The guarantee regimes are not levels on one scale: A is a per-instance logical implication, B a population-level statistical statement, C a release procedure, and D/E participation boundaries.For a shared approved formalization, a sound Type A gate has zero selective acceptance error when it admits artefacts, but this does not guarantee coverage because it may reject most artefacts.Source: §3.7 · p. 12

3.8 Formalization Fidelity

  • Requirement-to-formalization fidelity and formalization-to-verifier soundness are distinct gaps; approval of a formalization does not prove it faithfully represents the requirement.Informal requirements rely on engineering evidence and authorized interpretation; where requirements are already formal, fidelity may collapse or be supported by refinement or equivalence evidence.Source: §3.8 · p. 13
  • Formalization is itself a classified workflow task, with human judgement needed to accept fidelity for natural-language requirements and justified deterministic checks separable as their own units.A change to an approved formalization creates a new version, requiring reevaluation of every dependent gate; approval alone does not establish γ = 0.Source: §3.8 · pp. 13–14

3.9 Gated-Fallback Comparison with an Incumbent Process

  • A gated process’s comparison with an incumbent depends on admitted-artefact error (η), formalization-fidelity residual (γ), and the incumbent’s error on admitted instances, qM(E).The comparison is for a single workflow unit under a declared recurring population, positive admission probability, fallback routing for rejected instances, and fallback-performance equivalence; it does not automatically apply to one-off decisions.Source: §3.9 · pp. 14–15
  • For a sound Type A gate within its valid population, η = 0; under ideal fidelity γ = 0, the comparison supports non-degradation without estimating incumbent error or generator reliability.The improvement size still depends on admission probability and qM(E), and approval of a formalization does not establish perfect fidelity.Source: §3.9 · p. 15
  • Type B comparison is conditional on calibration evidence bounding η and establishing positive coverage, while Types C, D, and E receive no comparison of this form.For Type B, the incumbent’s error on the gate-admitted subpopulation must also be evidenced; the framework does not presume fallback equivalence for human adjudication.Source: §3.9 · p. 15

4 Staged Classification and Readiness Procedure

The procedure decomposes engineering activities into workflow units, assigns each a participation and assurance regime, and separately records deployment readiness. It also sets composite-release and coverage rules while limiting assurances to the unit level.

4.1 Participation and Assurance Predicates

  • Participation predicates assess whether AI can contribute, whether an incumbent path exists, and whether AI adds justified value; assurance predicates assess formalization, verifier soundness, and calibration.Mechanism-availability predicates require an existing mechanism supported by evidence, not one that could be developed.Source: §4.1 · pp. 15–16

4.2 Staged Classification Procedure

  • Stage 1 assigns Type D, E, or C[assist] where applicable; Stage 2 selects Type A, B, or C[adjudicate] according to available mechanisms.Classification is distinct from readiness, and a target regime attainable in principle carries no guarantee.Source: §4.2 · pp. 17–18

4.3 Composite Activities and Coverage

  • A composite activity releases each unit through its READY prescribed mechanism or fallback, or through the existing non-AI process when NOT READY, and must evidence requirement coverage.Coverage is a fidelity claim about decomposition, supported by engineering evidence and approval rather than generally proved mathematically.Source: §4.3 · p. 18
  • Composite records retain distinct unit guarantees rather than a scalar overall guarantee; narrowing review to the residual can leave requirement portions unchecked.The probability that all units are correct is not generally the minimum of their individual correctness probabilities.Source: §4.3 · p. 18

4.4 Reclassification and the Workflow Boundary

  • A unit’s regime may change with requirements, mechanisms, permissions, AI capabilities, or population shifts, and its record’s version determines the applicable assignment and readiness.A material change in generator version or population requires recalibration of a Type B guarantee instance.Source: §4.4 · p. 18
  • Records support local claims only for inputs within declared valid domains; workflow dependability, error propagation, and engineered-system safety are not derived.Composition theory for checking whether upstream guarantees support downstream input assumptions remains outside scope.Source: §4.4 · p. 19

5 Case Study: AI-Assisted Structural Analysis of a Wing-Spar Segment

The executed wing-spar workflow combined deterministically gated AI-generated CAD, retained deterministic computation, and explicit human judgement across units with different assurance needs. Its outcomes were governed by predeclared evidence requirements, not favourable numerical values alone.

5.1 Case-study objective and workflow

  • AI-generated CAD entered through a deterministic gate, established deterministic tools handled downstream computation, and human judgement addressed residual adequacy and interpretation.The case study used a tapered, stiffened, cantilevered wing-spar shell model evaluated under positive and negative tip loading.Source: §5.1 · p. 19, Figure 5.1 · p. 20

5.2 Instantiating the classification

  • The executed set comprised 17 units: one Type A, six Type D, four C[adjudicate], and six C[assist], with no Type B or E.Type B was not instantiated because the one-off study lacked a recurring deployment population.Source: §5.2 · p. 21

5.3 AI generation with deterministic admission

  • The AI-generated CAD program passed all 23 geometric and validity checks and was admitted unmodified on its first attempt; the fallback was not invoked.Admission establishes conformance to the approved geometric formalization, conditional on verifier soundness, not every engineering property.Source: §5.3 · p. 22

5.4 Deterministic downstream execution and reproducibility

  • A 61-entry hash manifest pinned specifications, criteria, material data, analysis plans, tools, and solver binary, but solver analyses were run only once.The evidence supports repeatability of the implemented process, not physical validity, engineering-model adequacy, or formal tool qualification.Source: §5.4 · pp. 22–23

5.5 When a favourable numerical result is not sufficient for PASS

  • The flange project metric changed from 182.6 MPa to 246.5 MPa as the maximum moved to an unanticipated tip load-introduction corner.A post-verdict interior-flange diagnostic showed about 0.2% change, but was not a verdict input and did not establish compliance.Source: §5.5 · pp. 23, 25
  • The four stress clauses were NOT_CONVERGED and received ADJUDICATE despite a flange stress ratio of 0.526, because the predeclared rule required converged metrics.The flange project metric changed by 25.9%; stiffener stress changed by 16.4%.Source: Figure 5.2 · p. 24, §5.5 · p. 25
  • The approved stress formalization omitted the tip load-introduction region revealed in execution, so any refinement must be fixed before a future rerun.No criterion was changed retrospectively; the gap surfaced as adjudication.Source: §5.5 · p. 25

5.6 Final verdicts and engineering adjudication

  • Buckling passed its predeclared rule, while all 17 READY units ended with 12 PASS, 3 ADJUDICATE, and 2 Not applicable outcomes.The smallest positive buckling factors were approximately 4.835 and 12.086, above the threshold of 1.5.Source: §5.6 · pp. 25–26

6 Discussion

The discussion argues that the framework’s structure is transferable in principle, while its evidence supports only a single author-led structural-engineering case. It distinguishes unit-level assurance from certification and identifies further evidence and development needed.

6.1 Transferability beyond the case study

  • Cross-domain applicability is argued from the framework’s domain-independent structure, not empirical validation; each application needs fresh decomposition, formalization, evidence assessment and validity-domain decisions.The executed case instantiated Types A and D, C[adjudicate] and C[assist], but neither Type B nor Type E.Source: §6.1 · pp. 26–27

Framework-level limitations

  • Framework allocations depend on engineering and organizational judgements, formalization fidelity is not proved by approval, and each regime’s guarantee has distinct conditions.Readiness also involves judging evidence adequacy, while unit-level records establish no workflow-level dependability or error-propagation bound.Source: Framework-level limitations · p. 27

Case-study limitations

  • The case was author-led, and its results are interpretable only within idealized modelling choices, including perfect-geometry linear buckling and project criteria not equivalent to certification requirements.The model also used a mid-surface shell representation and a linearized rigid tip-load relation without established finite-rotation validity.Source: Case-study limitations · p. 28
  • The case does not establish execution isolation, cross-platform reproducibility, physical model validity, or productivity improvement against a manual baseline.Solver analyses were not independently repeated, and no comparable manual baseline exists.Source: Case-study limitations · p. 28

6.3 Dependability, certification and accountability

  • The framework supports unit-level assurance and governance by making mechanisms and evidence inspectable, but it does not establish certification, regulatory compliance, system safety or legal duties.It identifies where participation, acceptance or execution decisions occur and where associated evidence is held; it is not a theory of legal liability.Source: §6.3 · p. 28

6.4 Future work

  • Future assurance work includes stronger formalization traceability, possible proof-based Type A soundness evidence, execution isolation and independent reproducibility checks.The paper also calls for recurring deployments to exercise Type B, independent engineering cases, and empirical study of Type C review issues and efficiency against a predefined baseline.Source: §6.4 · p. 29
  • The wing-spar stress rule could be refined only as a new formalization version fixed before a new run, while workflow composition theory remains open.A composition theory would need to address rejected outputs and fallbacks, iterative revisions, and dependencies among assurance records; reusable assets would still require scope, version and validity-domain checks.Source: §6.4 · pp. 29–30

7 Conclusion

The paper presents a formal framework for assigning AI participation and assurance to individual workflow units, with distinct regimes and guarantees. In an executed wing-spar workflow, deterministic admission and retained tools were combined with human judgement; insufficient convergence evidence kept favourable stress values from yielding PASS.

7 Conclusion

  • The framework records each unit’s requirement, applicable formalization, regime, mechanism, fallback, evidence obligations, guarantee and readiness, while keeping formalization fidelity distinct from verifier soundness.Its staged procedure separates classification, target regime and deployment readiness, and states recurring-population conditions for comparison with an incumbent process.Source: §7 · p. 30
  • In the executed wing-spar case, AI-generated CAD entered analysis through deterministic admission, established tools were retained, and human judgement handled residual or unresolved evidence.Favourable stress values did not yield PASS without the required convergence evidence, and READY assurance records did not imply PASS.Source: §7 · p. 30
  • The conclusions are bounded to local, unit-level assurance in a single author-led case; independent engineering applications are needed to validate the framework beyond it.The case does not claim workflow-level dependability, certification, structural safety or productivity.Source: §7 · p. 30

Appendix A. Proof and derivational details for Proposition 1

The appendix directs readers to Section 3.9 for the proof’s notation, setting, assumptions and defined quantities.

Appendix A. Proof and derivational details for Proposition 1

  • The proof’s notation, setting (S1)–(S4), fallback-performance equivalence and quantities dgate, dM, η, γ, qM(E) and ω are defined in Section 3.9.The appendix passage provides a cross-reference rather than restating those definitions.Source: Appendix A · p. 32

A.1 Proof of Proposition 1

The derivation uses the stated assumptions and fallback-performance equivalence to rewrite the conditional probability of failure in terms of the proposition’s quantities.

A.1 Proof of Proposition 1

  • Fallback-performance equivalence makes the second terms in the derivation coincide.This substitution follows the use of the law of total probability under (S4).Source: §A.1 · p. 33
  • Partitioning ¬R according to ˆS gives P(¬R | E) = γ + (η − ω).The proof identifies η as the sum of the two conditional probabilities for ¬ˆS, one with R and one with ¬R.Source: §A.1 · p. 33
  • The derivation obtains P(R(i, o) | E) = 1 − η − γ + ω.The proof states that this follows by substituting the partitioned probability and the fallback-equivalence terms.Source: §A.1 · p. 33

A.2 Remark on the over-restriction residual

The over-restriction residual ω is needed for the exact decomposition (4), but not for (3) or its derived sufficient conditions for non-degradation.

A.2 Remark on the over-restriction residual

  • ω is required for exact decomposition (4), but not for (3) or the sufficient conditions for non-degradation derived from it.Source: §A.2 · p. 33

A.3 Corollary 1 (Type A): further detail

With γ = 0, gain is strictly positive exactly when the incumbent errs on admitted instances with positive probability, and its size is limited by admission probability.

A.3 Corollary 1 (Type A): further detail

  • With γ = 0, gain is positive exactly when the incumbent errs with positive probability on admitted instances, and its size is limited by admission probability.Practical benefit therefore depends on how often the generator produces artefacts the gate admits.Source: §A.3 · p. 33

A.4 Corollary 2 (Type B): further detail

When α + γ exceeds qM(E), the coverage bound cannot provide a lower bound in that direction.

A.4 Corollary 2 (Type B): further detail

  • If α + γ > qM(E), the bracket is negative, so the coverage bound yields no lower bound in that direction.Source: §A.4 · p. 33

A.5 Corollary 3 (Types C, D, E): derivational detail

For Types D and E and C[assist], no AI artefact is admitted; C[adjudicate] often fails the bivalence condition and lacks framework-bounded human-gate error or presumed fallback equivalence.

A.5 Corollary 3 (Types C, D, E): derivational detail

  • Types D and E use no AI and C[assist] is advisory, so none admits an AI artefact and P(E) = 0, failing (S3).Source: §A.5 · p. 33
  • C[adjudicate] often fails (S1) because R is not bivalent; even when bivalent, no record clause bounds η or guarantees fallback-performance equivalence after the adjudicator sees AI material.Any comparison rests entirely on empirical human-factors evidence, outside the framework’s guarantees.Source: §A.5 · p. 33

Supplementary Material

The supplementary material describes an author-led application of the framework to an executed mixed engineering workflow, with engineering judgements retained by the study lead.

Supplementary Material

  • The case study applies the framework to an executed mixed engineering workflow, routing insufficient evidence to ADJUDICATE rather than PASS.It is not a certification, airworthiness or independent assurance exercise; its scope limits are those stated in Section 6.2.Source: Supplementary Material · p. 34
  • Tan retained responsibility for the engineering judgements and approvals required by the study, with human review where the assigned regime required it.Deterministic steps relied on their defined mechanisms and recorded tool outputs as evidence.Source: Supplementary Material · p. 34
  • Reported study values come from frozen execution records, except for the explicitly identified later re-hash; no engineering analysis was rerun for the Supplement.The later re-hash is dated 2 October 2026.Source: Supplementary Material · p. 34

S1 Executed workflow and study setup

This section describes the executed 17-unit wing-spar workflow and its retrospective evidence review: AI-generated CAD was deterministically gated, established tools handled downstream computation, and engineers retained defined judgement and adjudication roles. Insufficient convergence evidence led affected clauses to be adjudicated rather than passed, while the original case-study record lacked evidence for its assigned obligations and was not ready.

S1.2 Workflow decomposition and regime allocation

  • The workflow comprised 17 units, all READY before execution, combining AI CAD generation, deterministic processing and human judgement without assigning a regime to the workflow as a whole.The controlled executed-set count is distinct from the separate 22-item retrospective audit.Source: S1.2 Workflow decomposition and regime… · p. 34

S1.3 Architecture of the executed mixed workflow

  • AI produced the CAD program and advisory or draft material, while the study author retained judgement and adjudication; deterministic computation ran automatically without manual repetition.The AI-generated CAD entered the computational chain only through a deterministic geometric gate with bounded fallback.Source: S1.3 Architecture of the executed… · pp. 34–35
  • Four stress clauses and the Q-5 cross-check were dispositioned ADJUDICATE because unresolved evidence was escalated rather than converted into PASS.Favourable stress magnitudes alone did not suffice where predeclared mesh-convergence evidence was insufficient.Source: S1.3 Architecture of the executed… · p. 35

S1.5 Frozen study setup and reproducibility

  • The study froze a 61-entry hash manifest for specifications, criteria, analysis plans, tools and environment locks, but solver analyses ran only once and were not independently repeated.The contemporaneous report recorded a 61/61 post-run match without a preserved standalone machine-readable check; a later re-hash reproduced it. The evidence supports process repeatability, not physical validity or formal tool qualification.Source: S1.5 Frozen study setup and… · pp. 35–36

S3.1 Approved geometric formalization

  • The 17-clause geometric formalization covered specified properties only, with residual adequacy kept as a separate adjudication unit; engineering evidence supported gate soundness, not formal proof.The gate's evidence included a 32-case test suite with stated expectations met, 19 logic-level tests and author review; the study author approved the formalization before generation.Source: S3.1 Approved geometric formalization · p. 38, S3.3 Verifier evidence · p. 39

S3.5 Outcome

S4.2 Solver input decks

  • The deterministic downstream chain checked every deck before solver execution: 21/21 decks passed an independent checker and regenerated byte-identically, and 21/21 solver jobs completed without recorded errors or warnings.The decks covered 12 linear static cases, six linear buckling cases, two doubled-load buckling checks and one NLGEOM run.Source: S4.2 Solver input decks · p. 40, S4.3 Solver · p. 40

S5.3 Results

S5.3 Results / S5.4 Clause verdicts

  • The web stress metric converged, but the stiffener metric did not; Q-5 was predeclared for adjudication because finite-rotation validity of the linearized tip-load relation was unestablished.At ultimate load, δ/L was 0.0131, above the 0.01 escalation threshold; the doubled-load check raised no anomaly, and each fine-mesh buckling run returned 20 positive factors.Source: S5.3 Results · p. 42, S5.4 Clause verdicts · p. 43

S5.5 Final dispositions

  • Favourable stress ratios (at most 0.526) did not yield PASS without convergence evidence; Q-1± and Q-2± were adjudicated, Q-3± passed, and Q-5 remained adjudicated.No threshold or criterion was changed; Q-4 was reported, and Q-5 was supported but retained its predeclared ADJUDICATE disposition.Source: S5.5 Final dispositions · p. 43

S5.6 Unit outcomes: READY is not PASS

  • Of 17 READY units, 12 were labelled PASS, three ADJUDICATE, two not applicable by pre-run decision, and none FAIL.T7b passed because it executed its approved criteria as specified, even though some dispositions it produced were ADJUDICATE.Source: S5.6 Unit outcomes: READY is not PASS · p. 44

S6.1 Engineering judgements made by the study author

  • The study author accepted the geometry and three-level analysis with limitations, kept Q-5 adjudicated, and deferred model refinement to a new versioned iteration.The demonstration was accepted as a research demonstration, not certification; no criterion was changed.Source: S6.1 Engineering judgements made by… · p. 44

S6.2 Post-hoc diagnostics (not verdict inputs)

  • Post-hoc diagnostics were computed after verdicts and changed no verdict or criterion; they suggested flange non-convergence was confined to end regions, while the stiffener remained unresolved.The interior flange values were 179.4 MPa and 179.7 MPa (0.2% difference); the stiffener values were 41.9 MPa and 48.2 MPa (13.0% difference). These diagnostics did not establish compliance.Source: S6.2 Post-hoc diagnostics (not verdict… · p. 45

S6.3 Finite-rotation behaviour of the tip link

  • At positive ultimate load, the tip rotation was approximately 0.0202 rad, and the NLGEOM moment deficit matched θ2/2, consistent with the link’s unestablished finite-rotation validity domain.The deficit was 2.01 × 10−4 versus θ2/2 = 2.05 × 10−4.Source: S6.3 Finite-rotation behaviour of the… · p. 45

S6.4 Formalization limitation observed in execution

  • The executed stress rule represented the root-clamp region but not the tip load-introduction region revealed by the run; any refinement was reserved for a future version fixed before rerun.The refinement was not adopted retrospectively.Source: S6.4 Formalization limitation observed… · p. 45

S7 Retrospective framework audit of the original case-study record

  • A retrospective audit identified 22 original-record items, with 0/22 ready; checks presented as formal or statistical verification were classified as unrecorded human adjudication on surviving evidence.The audit’s 22 items differ from the controlled execution’s 17 workflow units and must not be pooled or compared as rates; NOT READY reflected unmet evidence obligations, not a finding that the work was unsafe or wrong.Source: S7 Retrospective framework audit of… · pp. 45–46

Related papers

Loading 2610.04084v1…