E.17.EFP:11 - SoTA Alignment and Source-Scope Boundary
Source-use rule. A source supports only claims within the problem population and action it actually studies. The external sources below concern AI explanations, NLP/model interpretations, LLM-generated explanations, RAG outputs, or interactive XAI systems. They do not establish a universal architecture for ordinary human-authored engineering notes.
| Claim need | Exact source and actual scope | Local use | Boundary or rejected transfer |
|---|---|---|---|
| Keep claim-bearing episteme, source-to-target relation, publication form, and carrier distinct. | Current FPF C.2.1, A.6.3, and E.24.PUB. | Identify claims, exact concern and effective scheme, then test expression sufficiency for the declared use before EFP classification. | This is current internal ontology, not a conclusion imported from an architecture-description standard. |
| Explanations of AI-system results are purpose- and recipient-sensitive and must state knowledge limits. | Phillips et al. (2021), Four Principles of Explainable Artificial Intelligence, NISTIR 8312, DOI 10.6028/NIST.IR.8312; government-guidance lineage. | Adapt bounded reader use and explicit limits when an AI explanation is current. | Do not generalize this XAI guidance into mandatory fields or classes for every technical explanation, and do not present it as the whole current research line. |
| Plausibility and faithfulness of model interpretations are different evaluation questions. | Jacovi & Goldberg (2020), Towards Faithfully Interpretable NLP Systems, ACL DOI 10.18653/v1/2020.acl-main.386; research lineage. | For NLP/model interpretation, do not infer faithfulness from persuasive prose. | The paper studies interpretable NLP systems, not ordinary human engineering exposition; later work further distinguishes self-consistency and intervention-based evaluation. |
| Output-level consistency tests for LLM explanations are not automatically tests of faithfulness to model internals. | Parcalabescu & Frank (2024), On Measuring Faithfulness or Self-consistency of Natural Language Explanations, ACL DOI 10.18653/v1/2024.acl-long.329; later repair of an overclaim in the evaluation line. | Name the actual check as self-consistency when that is what it measures. | Apply only to generated/LLM explanation use; do not require it for human-authored notes. |
| Current LLM-explanation work tests faithfulness through model-behaviour intervention rather than surface plausibility alone. | Chuang et al. (2026), FaithLM: Towards Faithful Explanations for Large Language Models, EACL DOI 10.18653/v1/2026.eacl-long.177; current research line. | Use an intervention-shaped evaluation only when the current task actually asks whether an LLM explanation reflects model decision behaviour. | EFP’s source ClaimGraph comparison is not a FaithLM score and does not import model-internal faithfulness into ordinary engineering text. |
| Retrieval quality, answer faithfulness, and answer relevance are distinct RAG evaluation dimensions. | Es et al. (2023), RAGAS, arXiv:2309.15217; Saad-Falcon et al. (2023), ARES, arXiv:2311.09476; RAG-evaluation method lineage. | Keep retrieved context, source use, and claim recoverability separate for RAG-generated explanations. | These metrics do not define FPF ontology, do not exhaust current RAG evaluation, and do not apply without a retrieval pipeline. |
| Repeated queries, evolving models/data, responsiveness, and traceability create system-level demands for interactive XAI. | Labarta et al. (2026), X-SYS: A Reference Architecture for Interactive Explanation Systems, arXiv:2602.12748v3; current emerging preprint. | Use interaction-sensitive prompts only for an actual interactive explanation system. | Do not transfer a five-component XAI system architecture or its fields to a static human-authored note, and do not treat an emerging preprint as settled standard. |
| Decide whether ordinary human-authored engineering explanation needs EFP at all. | No external source in this set establishes EFP’s four-class architecture for that population. Local evidence is the two-case task replay in E.17.EFP:5.7. | Prefer a source locator plus one bounded/blocked-use sentence when that performs the task. Use EFP only when class ambiguity changes action. | Present this branch as provisional local design rationale, not current external SoTA. Reopen if exact technical-writing, discourse, or decision-record evidence changes the comparison. |
Source-grounded branch. The XAI/NLP/RAG sources justify caution about generated or model-facing explanations: fluency, plausibility, retrieved context, or an AI summary label does not establish claim preservation, evidence, or reliance. They support the focused identity and use check only when that population is current.
Local human-authored branch. For ordinary human explanation, the architecture is justified only by the concrete local problem and the E.17.EFP:5.7 replay. The default is non-use when a simpler source-linked boundary sentence is equally comprehensible, preserves the claims, costs less, and prevents the same overread.
Retained result. Keep only the identity-and-bounded-sufficiency screen, an explanation class when it changes the next action, the compact bounded/blocked use, and a reopen condition. Add reader-model, trace, provenance, evidence, RAG, self-consistency, or interactive-system details only when their exact source-scoped situation is present.