Instruction lineage
Understand why it chose
Account for technical, policy, model and context influences that shape an agent’s action.
ISAF v2.0 · Instruction Stack Audit Framework
From what shaped the AI to what its actions can cause.
ISAF v2 traces the path from instructions and system context to agent capabilities, consequential choices and observable effects. It separates policy, permission, delegated authority, implementation and runtime evidence, so conclusions stay grounded in what can be established.
Follow the consequence. Compare the authority. Know the evidence.
ISAF v2.0 is an author-issued research framework. The archival publication and external peer review are separate milestones.
INSTRUCTION TO CONSEQUENCEISAF v2 accountability path
One action, several distinct accountability questions
OriginsModel, system context, policies and retrieved information
Instruction and intentAgent and toolAuthority and choicePotential consequence
Evidence keeps each claim in focus
Reachable action paths, approved operating rules, effective access, delegated choice and observed activity each contribute a distinct answer.
Readable path first. Evidence detail on demand.
Why consequence-level accountability matters
A valid tool call is not the whole answer.
An AI agent can use an approved tool and a valid credential while a consequential decision still raises questions about what shaped it, whose authority governed the choice and which effects are supported by evidence. ISAF v2 brings these questions into one research model.
Action path
Find where it can lead
Trace reachable tools, handlers and indirect paths to potential effects beyond a single API call.
Evidence-backed decision
Separate capability and authority
Compare what was asked, permitted, granted, implementable and observed, including authority over consequential choices.
Explore the seven-part research model
One framework for the full accountability question.
ISAF v2 connects seven research constructs. Choose one to see the question it answers, what it contributes and why it matters for consequential AI decisions.
01 | The Instruction Stack
What shaped this AI decision?
Trace the layers that influence model and agent behavior, from technical substrate and training to system instructions and retrieved context.
Explore HAIEC methodologyExample
A support agent’s refund recommendation can be affected by its system policy, retrieved transaction record and tool configuration, not only by the latest user message.
What this contributes
A foundation for reconstructing the decision’s technical and instruction provenance.
Preserving the original Instruction Stack
Historical layers remain part of the research
ISAF retains historical technical layers 0 through 8: physical substrate, logic, instruction set architecture, programming and compilation, operating environment, network distribution, machine learning frameworks, training data and objective functions. ISAF v2 adds modern runtime influences such as prompts, retrieved context, memory and tool information without changing the original layer identities.
Explore the ISAF Logger for training lineageFive independent evidence planes
Separate what was requested, permitted, possible and observed.
The five evidence planes are independent, not a sequence in which every source must be present to identify a useful technical risk. Select one to understand what it can establish.
Requested
What did the task ask the agent to accomplish?
A verified task or request can establish intended operation and scope for the decision under review.
The distinct question
The requested task does not by itself establish approved policy, effective permission or delegated authority over a sensitive choice.
Useful next step
Connect the request to the relevant consequence path and operating policy.
Delegated Action Integrity asks whether the agent was entrusted with a consequential choice. Delegation is a distinct authority relationship across these evidence planes, not another evidence plane or an extra required upload to see source risks. Explore DAI · See HAIEC assurance
Consequences, accountability and proof
Let established findings stand. Investigate the specific remaining question.
Source analysis can identify a consequential action route before all runtime evidence is connected. Additional source, identity or authority context helps determine what that path means for a business decision. ISAF v2 explains how to connect those facts without allowing one source to speak for a different claim.
A refund action as an example
From a visible tool to a consequential choice
- Instruction
- Refund an eligible transaction
- Capability
- Agent may call a refund handler
- Policy
- Refund to original method, when applicable
- Effective grant
- Credential access assessed separately
- Delegated choice
- Who may select a new recipient?
- Observed effect
- Only recorded events support an execution claim
First useful finding: a potentially consequential refund path exists in the inspected source. A recipient authority question can be investigated without erasing that finding.
A scientific framework with scoped validation
Research should show its method, evidence and limits.
ISAF v2 provides a structured research program for testing consequential AI assurance, including normative checkpoint specifications, frozen reference evaluations and applied evidence studies. Those activities are distinguishable from a blanket production guarantee.
78Research checkpoint catalog, separate from HAIEC product control profiles
20/20DCD-Bench v0.2 internal frozen reference-suite outcomes
34/34Hardened v0.3 qualification on its separately scoped internal suite
What the evidence program examines
Deterministic replay
Frozen inputs and explicit evaluation rules allow controlled comparisons and reproducible reported outcomes.
Adversarial examination
Hard-negative testing can expose gaps, including failures the initial reference suite did not reveal.
Enterprise applied evidence
Scoped telecom control cases show how evidence, rules and human response are documented without claiming universal prevention.
Research manuscript and publication
ISAF v2.0: From Instruction to Consequence
The author-issued ISAF v2 research paper provides the full evidence-bound model for AI agent governance, authorization, delegation and consequential action. Its October 16 print edition follows the October 9 research manuscript.
Author-issued manuscript · Version 2.0
From Instruction to Consequence
An Evidence-Bound Framework for AI Agent Governance, Authorization, Delegation and Consequential Action
Subodh KC · HAIEC Research · 9 October 2026
The publication archive and independent peer review are separate from the author-issued manuscript. The website research collection will link the downloadable edition when available for public distribution.
Research framework
ISAF v2 model, accountability geometry and supporting evidence requirements.
Author-issued manuscript
Final paper and print-ready editorial edition prepared in October 2026.
Archival publication
Formal archive citation and DOI to be linked following publication.
Research and product roles
One research foundation. Different ways to use it.
ISAF v2 Framework
The accountability model
Understand the instructions, action paths, delegated choices and boundaries that shape a defensible conclusion.
Explore the frameworkHAIEC Assurance
The applied evaluation product
Examine source, policy, access and qualified observations through the currently available product evaluation and reporting surfaces.
See the assurance platformISAF Logger
The training-lineage utility
Generate selected framework, data and objective records for machine learning pipelines. The Logger is not the complete ISAF v2 research framework.
Explore the LoggerFrequently asked questions
Understanding the ISAF v2 framework.
What is ISAF v2?
ISAF v2 is the Instruction Stack Audit Framework’s evidence-bound research model for understanding what shaped AI behavior, how an agent action can cause effects, which authority governs the action and which claims the available evidence supports.
How is ISAF v2 different from the ISAF Logger?
The framework describes consequence-level accountability across instruction lineage, capability, delegated authority, evidence and time. The ISAF Logger is an earlier Python utility focused on generating selected training-lineage records. They play different roles.
What are the five evidence planes?
Requested, Policy Authorized, Effectively Granted, Code Capable and Observed. Each answers an independent question and contributes facts only within its supported scope.
Why is delegated authority different from permission?
A credential may permit an API call. Delegated authority concerns who entrusted an agent to make a consequential business choice, such as selecting a payment recipient. An approved rule can sometimes determine the choice without additional discretion.
Can a technical risk be found before runtime logs are connected?
Yes. Supported code and source analysis can identify an action path and potential consequence. Runtime or target evidence is required for a claim that something actually occurred, not for every earlier source-level finding.
What is the Proof Frontier?
It is the exact boundary between what the current evidence supports and the next conclusion that would require a different or stronger source. Established findings remain useful even when another question is open.
What is the Evaluation Epoch?
It anchors the comparison to a compatible system, policy, grant, source and runtime state. This helps avoid mixing facts from unrelated or outdated configurations into one conclusion.
Is ISAF v2 an official regulatory certification?
No. ISAF v2 is an authored research framework. It can inform assurance work and standards mappings, but it does not itself issue regulatory certifications or guarantee safe production behavior.
Is the manuscript available on Zenodo?
The author-issued version 2.0 manuscript and print-ready edition have been prepared. Zenodo archival publication remains forthcoming. The official archive link will be added when a record and DOI are available.
How does HAIEC relate to ISAF v2?
HAIEC is a reference implementation of many evidence-bound concepts, including action-path analysis, distinct evidence planes, delegated authority comparison and deterministic evaluation. Product availability and evidence coverage remain specific to the connected system and supported implementation.
High assurance in every consequence
Move from understanding AI behavior to assessing consequential action.
ISAF v2 provides the accountability framework. HAIEC provides applied assurance workflows for the supported sources, systems and decisions that matter to your enterprise.
Apply the framework
One consequential system. One important assurance question.
Explore action reachability, compare authority and use the relevant evidence to support a concrete decision. Start with the facts your connected sources can establish.
Explore HAIEC Assurance