EGAF AI-Regulated profile for high-risk AI systems

2026-08-30 · Assurance

The AI-Regulated profile is L2. It extends Core for high-risk AI systems. Core already carries NIST AI RMF 1.0 and NIST AI 600-1. This overlay adds the EU AI Act's high-risk obligations, ISO/IEC 42001, and the two agentic/LLM threat catalogues, and it pulls one L3 control down into L2.

Pre-release, founder-operated system; current capability labels describe the evidence boundary.

How the Evidence-Gated Assurance Framework works

The Evidence-Gated Assurance Framework (EGAF) is a spec-first assurance bar. It is in flight. A project claims conformance to EGAF L1, L2, or L3 at a given commit, with evidence attached. It is not a certification, an accreditation, or an authorization. Independent assessment of any organization remains a separate question.

Three parts make a run:

  1. A control catalog. Each control is either a machine check (an artifact an independent tool produced) or an evidence: attestation (a signed statement by a named principal). A named check with no backing artifact is DRIFT, never PROVEN. An empty-but-parsable file is not a verdict.
  2. A deterministic evaluator. Same inputs produce a byte-identical conformance report. The evaluator does not run tests or scanners. It reads artifacts other tools already produced and records their digests. An agent cannot manufacture PROVEN without producing the artifact the check names.
  3. A committed profile. .egaf/profile.yaml in the consumer repo is what CI evaluates. If that file is absent, the evaluator uses the bundled core profile, so L1 mandatory controls always apply. Profiles may only tighten a threshold; a loosening override is rejected at validation time.

Three levels. L1 Core is mandatory to claim any EGAF conformance: traceability, complexity and duplication caps, whole-codebase statement and branch coverage, mutation on changed files, verifier integrity, signed evidence, supply-chain hygiene. An EXCEPTION never satisfies an L1 control. L2 Regulated adds industry-profile controls (evidence packs, retention, model-risk records). L3 Safety-critical adds independence, structural coverage, tool-qualification inputs, and hazard linkage. Fail one applicable L1 control and the report's first line is L0.

Control status is derived from the seven check outcomes already used in Koinessa evidence bundles:

Check outcome Control status
PASS PROVEN
FAIL FAILED
SKIPPED-NOT-APPLICABLE with a recorded reason NOT-APPLICABLE
INFRASTRUCTURE-ERROR, TOOL-MISSING, EVIDENCE-INVALID, or ABSTAIN DRIFT
Missing evidence DRIFT

A level is achieved only when every applicable control at that level and below is PROVEN or NOT-APPLICABLE (L2/L3 may also count an unexpired EXCEPTION with an independent approver). The report always shows two honesty lists: Not proven (FAILED and DRIFT) and Accepted without proof (NOT-APPLICABLE and EXCEPTION). A clean badge over an empty honesty block is the failure the framework exists to prevent.

Frameworks are guidance. Controls are behavior. Evidence proves the behavior occurred. Organizational burdens — training, physical security, BAAs, a C3PAO assessment, DO-330 tool qualification, a SOC 2 Type II engagement — stay with the organization. The catalog cites clause identifiers and paraphrases; it does not reproduce normative text.

The catalog this series describes is a measured snapshot: 109 controls, vendored into Koinessa at pin 0.0.0-dev+98d08ef on 2026-08-30. The public specification repository is planned; until the operator creates its GitLab group, treat EGAF as a labeled pre-1.0 draft with publication pending. Koinessa's own first mainline conformance report is planned.

Read the Koinessa whitepaper for the surrounding Evidence-Gated Agentic Development thesis, and the verified-delivery report for what this system already publishes about itself.

The AI-Regulated profile (id ai-regulated, L2)

In flight. Source: vendor/egaf/catalog/profiles/ai-regulated.yaml. Extends core.

Added frameworks:

AGT-01 through AGT-09 are catalogued at L2, so they are already active once this profile declares level: L2. AGT-10 ("AI-produced safety artifacts independently verified") is catalogued at L3. This profile explicitly controls.includes AGT-10 so independent verification of AI-produced safety artifacts is in the L2 active set. That include is the load-bearing difference from a plain L2 overlay.

Applicability: agentic, services, and ui. Releases and deploys declared.

Retention is 12 months, the inherited Core floor. EU AI Act Art. 26(6) sets a separate six-month floor specifically for a high-risk AI system's automatically generated logs (the Art. 12 logging obligation), reconciled with applicable data-protection law. That is a narrower record class than this profile's assurance-evidence retention. Twelve months already exceeds the Art. 26(6) minimum for that class. ISO/IEC 42001 Clause 7 requires retention of documented information without stating a number.

What the organization must still do: the EU AI Act's provider and deployer obligations that are not pipeline-checkable (quality-management system, post-market monitoring, fundamental-rights impact assessment, notifying authorities). EGAF can require that model identity, tool-call arguments, and independent verification of safety artifacts are in the evidence input. It cannot register a system in an EU database.

What this page does not claim

EGAF conformance is a per-commit statement a project makes about itself with evidence attached. It does not confer a certification, a SLSA level, or an authorization to operate. Citing a framework in a profile is a crosswalk, not a state the adopter has reached. Planned items carry no dates.

This article is one of ten in a series. Each page keeps the same mechanics and changes the profile.

The rest of the engineering blog covers the verification loop that EGAF sits on. The Koinessa home page states the product position.