A2I — Algorithmic Auditability Index

• In industrial AI systems, traceability is usually handled as after-the-fact paperwork instead of an intrinsic property of the architecture.
• Regulatory frameworks exist — the EU AI Act, the NIST AI RMF — but there is no quantitative instrument to measure whether a decision pipeline is actually ready for an audit.
• A model’s technical maturity (accuracy, efficiency) does not guarantee its auditability, creating legal and safety risk in critical contexts.

Compliance and audit functions, and the engineering teams they assess: organizations running AI in production that must demonstrate — to regulators, clients, or themselves — that their decision pipelines can be reconstructed, contested, and trusted.

• Methodological research and development — I contributed to designing and defining the A2I index, a multidimensional instrument for quantifying traceability in AI pipelines.
• Empirical analysis and validation — case studies across four real industrial domains: predictive maintenance, computer-vision quality inspection, energy balancing, and supply chain.
• Auditing and assessment — I acted as a rater, collecting technical evidence (logs, MLOps artifacts, policies) and applying the standardized scoring rubric.

• Six dimensions structure the framework: data provenance (D1), model lineage (D2), decision-path reconstructability (D3), escalation history (D4), override traceability (D5), and contestability protocols (D6).
• A five-level scoring rubric — from “Absent” to “Architecturally embedded” — based on observable, verifiable evidence.
• Methodological validation via Cohen’s weighted kappa (κw = 0.748) and the intraclass correlation coefficient for inter-rater reliability.
• Member-checking: factual accuracy reviewed with senior technical contacts at the industrial partners involved.

• Geometric aggregation: the final index uses a weighted geometric mean, deliberately punishing systems where even one dimension is absent — traceability is a chain that breaks at its weakest link.
• Pipeline, not model: the unit of analysis is the whole flow — input, model, human action, feedback — rather than algorithmic explainability alone.
• Technology-agnostic by design: the index applies regardless of AI technique, and was successfully used on Transformers, CNNs, LSTMs, and gradient boosting.

• Comparative benchmarking across the four cases: the energy-balancing system scored highest (A2I = 72.4), the supply-chain system lowest (A2I = 24.4).
• Contestability (D6) emerged as the weakest, least standardized dimension in every industrial sector analyzed.
• Embedding traceability at design time (audit by design) produced scores 23–65% higher than compliance documentation reconstructed after the fact.

• Sophistication is not auditability. Technically complex systems (e.g. Transformer-based) are not automatically better prepared for audit; simple systems can be more transparent if well designed.
• Retrofitting resistance is real: adding traceability and contestation channels after a system is in production is extremely difficult and costly.
• The variance in inter-rater reliability on contestability shows exactly where regulation like the EU AI Act needs clearer operational guidance.