Output Type Before Quality: A Standards-Derived XAI Admissibility Rubric for Autonomous-Driving Safety

TEXT ANALYSIS: Oracle Protocol v5.0

Paper: "Output Type Before Quality: A Standards-Derived XAI Admissibility Rubric for Autonomous-Driving Safety" — arXiv cs.AI (Submitted June 3, 2026)

THE DISSECTION

This paper performs a structural audit of the explainability tooling ecosystem in autonomous driving safety assurance. It identifies a systematic category error: safety standards (AMLAS, ISO 26262, ISO 21448, ISO/PAS 8800) demand causal, directed, root-cause evidence — mechanisms, not correlations — while the dominant XAI methods (SHAP and kin) produce ranked feature lists, which are structurally incapable of satisfying those evidentiary requirements regardless of implementation fidelity. The paper derives 19 testable criteria across 7 lifecycle stages and structurally scores six method classes against them. The result: causal XAI is not merely preferable but structurally required at hazard identification, incident investigation, and data management stages — with +62%, +50%, and +50% rubric gap closures respectively. A single VLA proof-of-concept on ~80K driving clips supports the classification. The conclusion reframes XAI selection as a standards-compliance problem, not a method-popularity problem.

THE CORE FALLACY

The paper operates within an implicit assumption that the gap is an engineering problem: admissible methods exist, fidelity validation is "the open assurance challenge," and with better implementations the XAI ecosystem can achieve compliance. This is incrementalist malpractice. The actual structural reality under DT lens:

The gap is not caused by SHAP being poorly implemented. SHAP returns a ranked feature list because that is its mathematical ontology. You cannot "implement effort" your way from a marginal attribution to a directed causal mechanism. The paper correctly diagnoses the mismatch but then retreats to procedural optimism: "an admissible method's specific output content may still be wrong, and validating that fidelity... is the open assurance challenge." This implies the category is redeemable. It is not. Marginal attribution methods are categorically incapable of producing directed causal chains because they are built on associational foundations. The paper acknowledges this ("no implementation effort can convert into a directed chain") but then treats it as a fidelity problem rather than a category death sentence for correlational XAI in safety-critical domains.

HIDDEN ASSUMPTIONS

1. That compliance is achievable at scale. The paper assumes safety standards will be enforced with sufficient rigor to force method selection based on structural compliance rather than convenience, familiarity, or litigation avoidance. This assumes the institutional lag defense holds — that standards bodies and regulators will converge on the paper's framework before mass autonomous vehicle deployment. Unlikely. The incentive structure strongly favors retrospective explainability theater over prospective causal rigor.

2. That causal XAI methods are currently viable enough to deploy. The paper scores causal XAI as structurally required but does not rigorously address the implementation maturity gap. Causal discovery from observational driving data is computationally expensive, assumption-heavy, and fragile to unmeasured confounders. The rubric identifies what is necessary but does not adequately grapple with whether it is currently sufficient and deployable.

3. That the lifecycle stages under analysis are the primary risk vector. The entire analysis is framed inside the existing autonomous driving development pipeline. It assumes the production system will be sufficiently constrained by standards compliance that XAI method selection drives safety outcomes. Under DT logic, the more pressing concern is that the production system itself — the AI driving model — is rapidly moving toward sparse, black-box, transformer-based architectures where causal explanation is not merely difficult but conceptually incoherent. The paper optimizes for explainability within the current paradigm without interrogating whether that paradigm survives.

4. That "admissibility" is a stable target. The standards invoked (ISO 26262, ISO 21448) were written for human-driven development processes. They are being retrofitted onto ML systems retroactively. The evidentiary criteria derived from them inherit this structural mismatch at the design level.

SOCIAL FUNCTION

Classification: Prestige Signaling + Transition Management

This paper performs a valuable forensic function — it correctly identifies that the dominant XAI paradigm is categorically wrong for safety-critical applications. That is a genuine contribution. However, its social function is to manage the transition rather than declare the existing paradigm dead. It offers a rubric — a procedural framework for rational method selection — which is precisely the kind of institutional infrastructure that allows incumbents (regulators, OEMs, the SHAP ecosystem) to absorb the critique without fundamentally changing behavior. The paper says causal XAI is required; it does not say correlational XAI is disqualified. "Admissible" is framed as necessary but not sufficient — which is technically accurate but operationally translates to "keep using SHAP, just add causal methods where convenient."

The paper is also prestige signaling within the academic XAI research community: it takes the intellectual high ground by aligning with safety standards while remaining publishable because it does not threaten existing method families — it just scores them. A paper declaring correlational XAI permanently inadequate for autonomous driving would face massive resistance. A rubric that politely ranks them is fundable.

THE VERDICT

The paper is a partial truth with high diagnostic accuracy and low prescription courage. The evidence-type gap is real. The causal XAI requirement is real. The structural scoring methodology is rigorous within its assumptions. But the paper treats a categorical failure (correlational methods cannot produce causal evidence under any implementation) as an implementation gap (fidelity validation is the open challenge). This is intellectually defensible as a narrow technical claim but systemically misleading: it implies the gap is closeable through effort when it is, in fact, a function of mathematical ontology.

Under DT lens, this paper is a lag indicator. It documents that the existing XAI ecosystem is structurally mismatched to safety standards at the moment those standards become binding on deployed systems. The lag it identifies is real. The question the paper does not ask: what happens when the autonomous vehicle market deploys at scale before causal XAI matures to production readiness? The answer, under DT logic, is that explainability theater substitutes for genuine safety evidence, incidents accumulate, regulatory capture or regulatory panic follows, and the honest causal methods arrive too late to prevent the failure mode. The paper is an accurate autopsy of the current method ecosystem. It is not, however, an intervention that changes the outcome.