ATRS-HS46 Audit Evidence Pack v0.1: Strict Event-Level Reconstructability Scoring, Source Register, and Three Deep-Audit Cases

This evidence pack supports a structured audit of 46 high-stakes public records from the UK Algorithmic Transparency Recording Standard (ATRS). The pack contains a SHA256-verified source pack, metadata, a strict event-level PLAN reconstructability scoring table, a recomputation script and output, method notes, and three theoretically sampled deep-audit cases. The audit examines whether public ATRS disclosure records preserve the minimum decision-event information required to reconstruct and contest specific AI-mediated public decisions. The high-stakes subset was selected from a 131-record GOV.UK ATRS corpus using a direct-public-consequence criterion, covering welfare and benefits administration, social care and child protection, health and medical decision support, criminal justice and enforcement, fraud detection, identity verification, immigration routing, and transport regulation. Headline strict event-level metrics reproduced from the scoring table and script are: N=46; mean strict PLAN score = 6.87/12; F1 actor identity = 46/46 records identify institutional responsibility only; F5 uncertainty absent = 29/46 records; F6 event-level oversight trace present = 4/46 records; F6 oversight mentioned at any level = 42/46 records. The evidence pack also includes three deep-audit cases: DWP Universal Credit Advances Model, Newcastle City Council Magic Notes, and Care Quality Commission Risk Categorisation. These cases illustrate three reconstructability failure modes: deliberate withholding of decision-point uncertainty, workflow transparency without event-level actor traceability, and strong process-level disclosure that remains bounded by schema granularity. Scope limitation: this is a structured single-rater strict event-level reconstructability audit of public disclosure records. It does not evaluate the accuracy, legality, fairness, safety, operational effectiveness, or policy merits of the underlying algorithmic tools. Inter-rater validation is not claimed in this version. The purpose of the pack is to preserve the source boundary, scoring table, recomputation pathway, and deep-audit evidence trail for inspection and future research.Author: Hon Bor So, Independent Researcher. ORCID: 0009-0008-2768-7494.