HAIEC publishes the Instruction Stack Audit Framework for AI accountability

In December 2025, HAIEC’s research foundation expanded with the Instruction Stack Audit Framework (ISAF), a technical preprint examining how AI behavior depends on more than a visible prompt or final output. The work traces multiple abstraction layers that influence system behavior and the evidence that reviewers need to interrogate them.

The original framework describes nine instruction layers and a 127-checkpoint audit protocol. It was released as an openly archived preprint with a persistent DOI, helping researchers and practitioners cite and examine the methodology. The framework is an author-issued technical work, not a claim of journal peer review or regulatory certification.

The ISAF research informed the ISAF Logger implementation and later research into delegated authority and consequence-aware assurance. ISAF v2 extends the conceptual scope; its archival publication remains distinct from the published v1 preprint.

In a modern AI deployment, the final user message is only one of several sources influencing the model. System instructions, developer messages, tool specifications, retrieved content and surrounding application logic can all constrain or redirect behavior. An accountability model that treats the final prompt as the entire cause of an outcome is incomplete.

The original Instruction Stack Audit Framework introduced a structured examination of instruction layers. It sought to give assurance practitioners a way to ask where an instruction entered, how it was transformed or prioritized, and what evidence should survive for subsequent review.

The author-issued December 2025 preprint described nine instruction layers and 127 checkpoints, with a persistent archival DOI. This creates a citable reference that can be discussed, challenged and extended without relying on ephemeral descriptions or marketing content.

The protocol is an audit methodology, not a claim that applying a checklist automatically proves a system safe. It emphasizes the kinds of records, provenance and decision logic an assessor should seek when establishing what occurred inside a configured system.

The initial research influenced HAIEC’s approach to recording instruction and execution context. Subsequent work has focused increasingly on the interaction between instruction lineage and the separate authority needed to cause effects in connected systems.

ISAF v2 expands that line of inquiry toward consequences, permissions, delegation and operational evidence. The original framework remains a distinct historical contribution, with its own DOI and publication status.