The Instruction Stack Audit Framework (ISAF): A Technical Methodology for Tracing AI Accountability Across Nine Abstraction Layers
Bibliographic record
Abstract
AI accountability failures occur when regulatory audits examine outputs while root causes exist in instruction layers that remain undocumented and unauditable. Analysis of documented AI incidents including Air Canada's chatbot liability (2024), Amazon's hiring bias (2018), and Zillow's algorithmic valuation loss exceeding $500 million (2021) reveals a consistent pattern: problematic behavior traces to design-layer decisions involving objective functions, framework configurations, and data selection that were never systematically reviewed before deployment. Current AI governance frameworks including the EU AI Act, NIST AI Risk Management Framework, and ISO/IEC 42001 focus primarily on model outputs and data governance without providing technical specifications for documenting the full instruction stack from hardware substrate to emergent behavior. This creates a fundamental traceability gap where organizations can achieve nominal regulatory compliance while leaving the majority of their instruction stack unaudited. This paper introduces the Instruction Stack Audit Framework (ISAF), a proposed methodology designed to address this documentation gap. ISAF provides a nine-layer technical specification defining instruction propagation from voltage thresholds through objective functions to outputs, accompanied by a 127-checkpoint audit protocol for systematic instruction verification, an instruction lineage logging schema enabling cryptographic verification, a layer ownership assignment methodology for accountability attribution, and a risk scoring system based on abstraction distance and control strength. The framework draws on principles established in prior work on deterministic compliance systems and extends them to full-stack AI accountability. Three case analyses demonstrate how ISAF-based audits could have identified instruction-level risks in documented failures. The complete audit specification, logging schemas, and implementation templates are provided in appendices. ISAF is released for academic validation, industry pilot implementations, and regulatory consideration. Keywords: AI governance, algorithmic accountability, EU AI Act compliance, NIST AI RMF, ISO 42001, objective function auditing, instruction traceability, deterministic compliance, ML operations, AI safety, cryptographic audit trails, regulatory documentation requirements
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.032 | 0.071 |
| Meta-epidemiology (narrow) | 0.002 | 0.002 |
| Meta-epidemiology (broad) | 0.001 | 0.003 |
| Bibliometrics | 0.009 | 0.003 |
| Science and technology studies | 0.004 | 0.008 |
| Scholarly communication | 0.010 | 0.015 |
| Open science | 0.005 | 0.008 |
| Research integrity | 0.004 | 0.007 |
| Insufficient payload (model declined to judge) | 0.006 | 0.002 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".