Positioning — gaps in instruments and conjunctive gate design
Structured audit of how major governance instruments treat thresholds and compensation; proposes five-domain gate architecture, risk-tiered floors, and constrained overrides.
Ethical Alpha Audit™
Research layer: reproducible governance artefacts, independent QA, and version-pinned public releases.
Programme overview
Research programme: making AI deployment governance operational
Healthcare and high-stakes AI systems are governed by standards, risk frameworks, and procurement rules. Those instruments often stop short of a fully specified decision rule: what must be true, simultaneously, before a system is cleared to operate—and what happens when evidence is uneven across safety, bias, calibration, traceability, and evidentiary depth.
The Ethical Alpha Audit research programme develops and stress-tests non-compensatory (conjunctive) governance designs: parallel requirements that cannot offset each other. Work proceeds in a tiered architecture—shared reproducibility core, a primary empirical replay strand, and companion papers on positioning, simulation, threshold documentation, and large-scale sensitivity analysis—each with frozen code, manifests, and claim-to-artefact traceability.
- Monorepo root commit
- The public portfolio
system_snapshot.jsonused for programme binding does not record a single Git SHA for one monorepo root. This site does not invent one. - Tier-0 shared-core commit (portfolio snapshot)
cd9ad79fe16f34ad861bd6527670dcfbef8fe864- Portfolio snapshot
system_snapshot.jsonversion 1.2 · referenced state timestamp 2026-04-11T07:37:21Z (UTC)- Release designation
- v2.0.0 (portfolio release per snapshot update log)
- Zenodo DOI (portfolio)
- See each paper page for its minted DOI; package
CITATION.cffversion strings may read 1.0.0 while the portfolio uses a unified release label—treat commit + DOI landing pages as authoritative if they diverge. - Page generated (UTC)
- 2026-04-12
Why this work matters
- AI safety and patient risk: weak gates let unsafe or mis-calibrated systems deploy when aggregate scores look acceptable.
- Governance in practice: institutions need rules that can be replayed, audited, and explained—not only process checklists.
- Real-world failures: historical cases supply disciplined test material for whether declared rules would have fired before harm.
- Auditability: substantive claims map to outputs produced by automated reproduction pipelines and independent QA sessions.
Five programme papers
Select a paper for narrative overview, layered technical walkthrough, limitations, and full claim traceability.
Threshold justification — Monte Carlo simulation study
Large-scale simulation comparing non-compensatory gates to weighted composites and permissive baselines; quantifies unsafe deployment rates under heterogeneous evidence.
Framework comparison — Threshold Justification Stack (TJS)
Six structural failure modes in threshold practice; introduces a six-layer documentation stack and compares expectations to other governance instruments (interpretive analysis).
Historical replay — documented failures and controls
Applies a frozen five-gate engine to public-domain failure cases and regulatory control devices; ablation, compensation comparison, and robustness checks.
Sensitivity study — optimisation and global sensitivity
Preregistered Monte Carlo governance world-model; Pareto frontier over detection, throughput, harm, and friction; Sobol indices and lifecycle dynamics.
Limitations and ethics
This hub condenses programme narrative, navigation, and shared binding context. It is a summary layer, not a substitute for each paper’s full page. Every paper includes its own limitations, ethics framing, and scope notes; read those sections on the paper before drawing conclusions.
Interpretive and evidential boundaries—what the work claims, under which assumptions, and what it does not establish—are stated at the paper level. For any quantitative or artefact-backed reading, treat the repository commit and Zenodo DOI shown on that paper’s page as the authoritative binding, not this overview.
Label: Scientific rigour — claim extraction, structured traceability matrices, automated tests, and hash-validated outputs where declared in each repository. Label: Auditability — paper repository commits are registered in the portfolio snapshot alongside the Tier-0 shared-core pin. Label: Applicability — artefacts are designed for governance, risk, and clinical safety leadership, not abstract scoring exercises alone.
How the programme fits together
The shared reproducibility core supplies engines, validation patterns, and cross-paper consistency. Paper 4 (historical replay) is the tier-1 empirical anchor; Papers 1–3 and 5 extend positioning, simulation, documentation frameworks, and sensitivity and optimisation analysis.
Downstream website and client engagements should consume only released phases and preserve version binding (commit, DOI, generation timestamp) on every public page that asserts quantitative or artefact-backed results.