From Bayesian Knowledge Tracing to Verifiable Educational AI: A Design-Science Case Study in Cryptographic Attestation, Teacher Oversight, and EU AI Act Evidence Readiness

Version 0.3.2 (23 May 2026) is an author hand-pass over v0.3.1. The case-study scope, BKT replayability argument, evidence event model, EU AI Act mapping, threat model, and benchmark numbers are unchanged. v0.3.2 tightens the abstract, introduction, EU-AI-Act discussion, threat-model commentary, and conclusion in the author’s voice. The architecture figure is now embedded as a vector PDF in place of the v0.3.1 mermaid source block, and the references are rendered inline from the project bibliography (CSL APA-7) instead of referring out to references-v0.2.bib. Standing claim boundary preserved: cryptographic attestation cannot prove pedagogical validity, fairness, privacy compliance, or legal conformity. It can provide an integrity and replayability layer that makes selected governance claims independently checkable. Version 0.3.1 (18 May 2026) is a minor revision of v0.3 (17 May 2026, archived under the same concept DOI). v0.3.1 preserves the case context, the BKT model description, the evidence event model, the architecture diagram, the EU AI Act evidence mapping, the threat model, the evaluation artifacts and benchmarks (trace size 5,924 bytes; mean replay verification 22.89 ms; mean tamper test 50.84 ms), and the references list of v0.3. It applies the following cleanup so that the public preprint no longer carries internal editorial scaffolding: Rephrases §8.1 (Teacher and auditor review protocol) from authoring-stage language (“before submission to FAccT, AIES, LAK, EDM, or AIED, the paper should be paired with…”) to future-validation language describing what the protocol does, with no specific venue list. Rephrases §10 (Next work) to describe three planned derivative angles — fairness-and-accountability, learning-analytics, security-and-evidence — without naming the specific venues those derivatives might target. This revision affects positioning only; the substantive contribution of v0.3 stands. The paper remains a synthetic case study, not a claim of classroom efficacy, real-student evidence, production signatures, qualified timestamps, or legal compliance. v0.3 artifact-backed working draft (17 May 2026) of From Bayesian Knowledge Tracing to Verifiable Educational AI. This design-science working paper uses MATx / matx-hack and EATF as a case study for representing adaptive educational AI events as canonical, hash-chained, replayable evidence. The paper focuses on Bayesian Knowledge Tracing state updates, task recommendations, teacher-facing explanations, and teacher approval or override events. Version v0.3 adds a self-contained architecture figure, an explicit signature/RFC 3161 timestamp integration path, executable replay and tamper-test results, package-size and verification-latency measurements, and a short teacher/auditor review protocol. The artifact package remains synthetic and hash-chain-only: it contains no real student data and makes no claim of production cryptographic assurance. The central claim is narrow: cryptographic attestation cannot prove pedagogical validity, fairness, privacy compliance, or legal conformity, but it can provide an integrity and replayability layer that makes selected governance claims independently checkable.