RESEARCH REPORT

SingleAxis AI Safety Framework

Evidence-grade evaluation for enterprise AI deployment.

SASF is the methodology behind SingleAxis Evidence Reports: independent human evaluation, taxonomy-linked findings, reviewer agreement, adjudication, and audit-ready reporting for AI systems moving toward production use.

WHAT IT IS

A methodology, not a dashboard

SASF turns evaluation work into evidence that an enterprise can inspect, cite, and defend. It defines how a system is scoped, how failure modes are classified, how reviewers score outputs, how disagreements are resolved, and how conclusions become a versioned Evidence Report.

The framework is designed for systems where a generic benchmark is not enough: RAG assistants, agentic workflows, voice agents, vision systems, multimodal products, and classical machine learning in high-consequence settings.

REPORT OUTLINE

How the paper is organized

The PDF is the citeable report. The public repository is the machine-readable support layer for the same methodology.

  1. 01

    Problem framing

    Enterprise AI deployment gap

    Why production decisions need evaluation evidence beyond model benchmarks.

  2. 02

    Evaluation limits

    Limits of automated evaluation

    Where automated scoring fails and qualified human judgment becomes necessary.

  3. 03

    Method core

    Evaluation model and verdict logic

    How SASF turns scoped tests, reviewer scores, and layer gates into defensible verdicts.

  4. 04

    Taxonomy activation

    Failure taxonomy and intake activation

    How engagement context selects relevant failure modes without changing the framework.

  5. 05

    Deliverable

    Evidence Report structure

    The report format used to preserve findings, severity, adjudication, and audit trail.

  6. 06

    Worked example

    Clinical RAG scenario

    A concrete walkthrough showing how SASF records evidence in a high-consequence setting.

  7. 07

    External alignment

    Regulatory and standards mapping

    How the framework supports governance work without claiming to replace formal compliance.

  8. 08

    Boundaries

    Scope, limitations, and appendices

    What SASF does not cover, plus the supporting taxonomy and artifact references.

PUBLICATION PACKAGE

What supports the paper

The support repository should make SASF reusable outside this website. It should carry the versioned artifacts that let researchers, auditors, and tool builders cite, inspect, and implement the framework.

Paper source

Versioned Markdown, PDF release, changelog, and citation metadata.

Taxonomy package

Machine-readable category and code definitions generated from the platform source of truth.

Schemas

JSON Schema for taxonomy releases, findings, severity values, and Evidence Report interchange.

Crosswalks

Mappings to NIST AI RMF, OWASP LLM Top 10, AI Verify, MLCommons AILuminate, and ISO/IEC 42001.

Examples

Sample Evidence Report outline, sample finding records, and sample rubric-linked evaluation outputs.

Governance

Version policy, contribution rules, license, security policy, and release signing expectations.

FABRIC RELATIONSHIP

Fabric is evidence capture, not SASF itself

SingleAxis Fabric should not replace the SASF publication. Fabric is an open-source telemetry and control substrate for agent systems. SASF is the evaluation methodology and reporting framework.

The connection is useful: Fabric can capture decision spans, retrieval provenance, guardrail events, and human escalation records that later become evidence inputs for a SASF evaluation. That makes Fabric a companion implementation path, not a prerequisite for using the framework.

View SingleAxis Fabric

HOW TO CITE

Recommended citation

Bryan Rodrigues and SingleAxis. SingleAxis AI Safety Framework: Evidence-Grade Evaluation for Enterprise AI Deployment. Technical report SA-WP-001-R2.1, SingleAxis, July 2026. DOI: 10.5281/zenodo.21227848.

Download PDFOpen DOIRead related briefings