Paper 1 · Positioning

From governance principles to conjunctive deployment rules

Monorepo root commit
Not recorded in the public portfolio system_snapshot.json (v1.2, 2026-04-11T07:37:21Z) used for this binding. Not invented on this page.
Tier-0 shared-core commit (portfolio snapshot)
cd9ad79fe16f34ad861bd6527670dcfbef8fe864
Paper 1 repository commit (released)
61abf05f949f9c92f945959b290c76bf5878b67c
Zenodo DOI
https://doi.org/10.5281/zenodo.19499783
Release version
v2.0.0 (portfolio release designation); CITATION.cff may list package version 1.0.0 — treat commit + DOI as authoritative if they diverge.
Page generated (UTC)
2026-04-12

Executive overview

Problem: High-level AI governance frameworks describe lifecycle expectations and risk management culture but rarely specify a complete, non-compensatory rule for when deployment must stop.

Why it matters: Without an explicit conjunctive structure, organisations implicitly rely on compensatory scoring or informal trade-offs—precisely where harmful deployments can slip through when one domain is “good enough” on average.

Core insight

A conservative structured audit can show that widely cited instruments do not jointly encode threshold logic, non-compensation, and constrained overrides in one place—creating a design gap that conjunctive gate architectures aim to close.

What was done

Researchers encoded seven major governance instruments against four audit dimensions (threshold explicitness, non-compensatory structure, override rules, and combination logic). Outputs include rendered comparison tables, a formal five-gate specification, risk-tiered threshold tables, and override safeguards with accumulation limits.

An illustrative case mapping shows how a known clinical AI incident pattern maps onto domain-level refusals under the proposed architecture. Alignment mappings connect external checklist frameworks to the same gate vocabulary for operational use.

What was found

The structured audit supports the claim that no reviewed instrument in the table simultaneously combines fully specified thresholds, strict non-compensation across domains, and a documented override rule set (P1-C12). The companion gate specification separates safety-critical domains from overridable operational gates and ties thresholds to risk tier.

Companion empirical papers (simulation and historical replay) are explicitly scoped as external to this repository’s computational harness (P1-C06, P1-C07).

Why this matters for regulation, safety, and deployment

  • Board-level clarity: converts abstract “AI ethics” into testable conjunctive requirements.
  • Vendor diligence: procurement teams can ask whether a supplier’s assurance map covers every gate, not only aggregate metrics.
  • Regulatory dialogue: surfaces where statutes and guidance leave binary activation logic underspecified—without pretending to restate legal obligations.

Limitations and ethics

Single-auditor structured coding with conservative operational rules; frameworks evolve; table captures a frozen slice. Supplementary clinical-data narratives are out of scope for in-repo reproduction (P1-C19). Scope is necessary-not-sufficient for safe deployment (P1-C20).

Traceability: Each claim below maps to JSON sources and notebook-generated CSVs in the pinned repository. Status labels (TRACEABLE, NARRATIVE, EXTERNAL, etc.) are preserved verbatim from the engineering matrix.

View technical detail — notebook walkthrough (conceptual)
01_claim_traceability Renders Table 1 gap audit and documents coding dimensions (P1-C01–C03, P1-C08–C12).
02_framework_and_specification Builds gate architecture summaries and tiered threshold tables (P1-C04–C05, P1-C13–C15, P1-C18).
03_override_and_illustration Override criteria, safeguards, and Epic Sepsis illustration (P1-C05, P1-C14, P1-C16–C17).
04_release_validation Release checks and manifest alignment (P1-C22 partial via metadata).

Closing the box does not remove the executive conclusions already stated above.

Full claim traceability (P1-C01–P1-C22)

One-to-one with docs/claim_traceability.md in the pinned repository.

Claim ID Claim (summary) Code / data Notebook Output Status
P1-C01Governance instruments underspecify binary deployment activation logic01outputs/tables/table1_rendered.csv (context)PARTIAL (narrative + audit motivation)
P1-C02Compensatory aggregation permits cross-domain offsetting01NARRATIVE
P1-C03Audit: no reviewed instrument specifies full threshold + non-compensatory + constrained override ruledata/table1_gap_audit.json01outputs/tables/table1_rendered.csvTRACEABLE
P1-C04Proposed conjunctive five-domain gate architecturedata/gates_specification.json02outputs/tables/gates_summary.csvTRACEABLE
P1-C05Risk-tiered floors; override for G2–4; shared service modeldata/table2_risk_tiers.json, data/override_specification.json02, 03outputs/tables/table2_rendered.csv, outputs/tables/override_summary.csvTRACEABLE
P1-C06Companion simulation: gate vs composite under heterogeneous evidenceEXTERNAL (Paper 2 simulation repository)
P1-C07Historical replay: preliminary convergent evidenceEXTERNAL (Paper 4 historical replay repository)
P1-C08Four audit coding dimensions (MA4)data/table1_gap_audit.json (column schema)01outputs/tables/table1_rendered.csvTRACEABLE
P1-C09Conservative operational coding ruledata/table1_gap_audit.json (metadata.note)01outputs/tables/table1_rendered.csvTRACEABLE
P1-C10Table 1 characterisations for seven frameworksdata/table1_gap_audit.json01outputs/tables/table1_rendered.csvTRACEABLE
P1-C11Dominant pattern: lifecycle / evidence expectations vs binding combination ruledata/table1_gap_audit.json01outputs/tables/table1_rendered.csvTRACEABLE
P1-C12No reviewed instrument combines all three structural elementsdata/table1_gap_audit.json01outputs/tables/table1_rendered.csvTRACEABLE
P1-C13Conjunctive decision-theoretic rationale02NARRATIVE
P1-C14Gate roles; G1/G5 non-overridable; G2–4 overridabledata/gates_specification.json, data/override_specification.json02, 03outputs/tables/gates_summary.csv, outputs/tables/override_summary.csvTRACEABLE
P1-C15Table 2 risk-tiered thresholdsdata/table2_risk_tiers.json02outputs/tables/table2_rendered.csvTRACEABLE
P1-C16Four override criteria; three accumulation safeguardsdata/override_specification.json03outputs/tables/override_summary.csvTRACEABLE
P1-C17Epic Sepsis illustration mappingdata/epic_sepsis_illustration.json03(narrative cells)TRACEABLE
P1-C18FUTURE-AI ↔ five gates mappingdata/futureai_mapping.json02outputs/tables/futureai_alignment.csvTRACEABLE
P1-C19PhysioNet preliminary validation narrativeEXTERNAL / SUPPLEMENT (not reproduced in this repo)
P1-C20Scope boundaries (necessary not sufficient)NARRATIVE
P1-C21Pilot endpointsNARRATIVE
P1-C22Limitations (single auditor, provisional defaults, etc.)data/table1_gap_audit.json (metadata)01PARTIAL