Machine-Readable Behavioural Compliance Evidence for AI Systems: A Specification Profiling Framework

Fuente: Zenodo
Salvato in:
Dettagli Bibliografici
Autore principale: Caprazli, Kafkas M.
Natura: Recurso digital
Lingua:inglese
Pubblicazione: Zenodo 2026
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866901163555684352
author Caprazli, Kafkas M.
author_facet Caprazli, Kafkas M.
contents Existing methods for evaluating AI behaviour conflate personality measurement with specification compliance. This paper presents the Specification Profiling Framework (SPF), a specification-verification method that produces machine-readable evidence of whether an AI system's observable output conforms to an explicit behavioural specification. SPF evaluates systems across eight behavioural constraints using a two-turn protocol that isolates specification effects from baseline behaviour. Methodology validation with four commercial AI systems reveals significant per-system variation: compliance ranges from 0/8 to 6/8 constraints. A specification reversal anomaly (D8 DomainStrictness) demonstrates that multi-dimensional separated assessment surfaces structural failures invisible to scalar scoring. All evidence artefacts are structured (JSON), reproducible, and mapped to EU AI Act conformity assessment requirements (Annex A).
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_18984072
institution Zenodo
language eng
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle Machine-Readable Behavioural Compliance Evidence for AI Systems: A Specification Profiling Framework
Caprazli, Kafkas M.
AI conformity assessment
behavioural specification
EU AI Act
specification compliance
LLM evaluation
machine-readable evidence
CEN/CENELEC
AI standardisation
specification profiling
DGD framework
Existing methods for evaluating AI behaviour conflate personality measurement with specification compliance. This paper presents the Specification Profiling Framework (SPF), a specification-verification method that produces machine-readable evidence of whether an AI system's observable output conforms to an explicit behavioural specification. SPF evaluates systems across eight behavioural constraints using a two-turn protocol that isolates specification effects from baseline behaviour. Methodology validation with four commercial AI systems reveals significant per-system variation: compliance ranges from 0/8 to 6/8 constraints. A specification reversal anomaly (D8 DomainStrictness) demonstrates that multi-dimensional separated assessment surfaces structural failures invisible to scalar scoring. All evidence artefacts are structured (JSON), reproducible, and mapped to EU AI Act conformity assessment requirements (Annex A).
title Machine-Readable Behavioural Compliance Evidence for AI Systems: A Specification Profiling Framework
topic AI conformity assessment
behavioural specification
EU AI Act
specification compliance
LLM evaluation
machine-readable evidence
CEN/CENELEC
AI standardisation
specification profiling
DGD framework
url https://doi.org/10.5281/zenodo.18984072