GPAI Compliance Measurement
Six providers. Six model cards. Fourteen requirements.
Real measurements of publicly available GPAI model documentation from OpenAI, Anthropic, Google DeepMind, xAI, Meta, and Mistral. Every number is a governed measurement, not an estimate.
Chapter V: General-Purpose AI with Systemic Risk
Meta: Llama 3.1 405B
requirements covered
completeness
critical gap
Top finding (AXI-1b, critical): Documentation addresses only 2/7 required concepts for acceptable use policies. This Annex XI requirement is materially missing.
Llama 3.1 405B Model Card, 12 April 2026, 14 requirements scored.
Chapter V: General-Purpose AI with Systemic Risk
Anthropic: Claude Opus 4.5
requirements covered
completeness
critical gaps
Top finding (AXI-1f, critical): Documentation addresses only 1/8 required concepts for the model licence. This Annex XI requirement is materially missing.
Claude Opus 4.5 Model Card, 12 April 2026, 14 requirements scored.
Chapter V: General-Purpose AI with Systemic Risk
Mistral: Magistral
requirements covered
completeness
critical gaps
Top finding (AXI-1b, critical): Documentation addresses only 1/7 required concepts for acceptable use policies. This Annex XI requirement is materially missing.
Magistral Model Card, 12 April 2026, 14 requirements scored.
Chapter V: General-Purpose AI with Systemic Risk
Google DeepMind: Gemini 3 Pro
requirements covered
completeness
critical gaps
Top finding (AXI-1f, critical): Documentation addresses only 1/8 required concepts for the model licence. This Annex XI requirement is materially missing.
Gemini 3 Pro Model Card, 12 April 2026, 14 requirements scored.
Chapter V: General-Purpose AI with Systemic Risk
OpenAI: GPT-5
requirements covered
completeness
critical gaps
Top finding (AXI-1b, critical): Documentation addresses only 1/7 required concepts for acceptable use policies. This Annex XI requirement is materially missing.
GPT-5 Model Card, 12 April 2026, 14 requirements scored.
Chapter V: General-Purpose AI with Systemic Risk
xAI: Grok 4
requirements covered
completeness
critical gaps
Top finding (AXI-1b, critical): Documentation addresses only 1/7 required concepts for acceptable use policies. This Annex XI requirement is materially missing.
Grok 4 Model Card, 12 April 2026, 14 requirements scored.
Annex XI coverage comparison
Per-requirement coverage across all six providers. Coverage status is sourced directly from each Annex XI analysis. Green: covered. Amber: partial. Red: gap.
| Requirement | OpenAI GPT-5 |
Anthropic Claude Opus 4.5 |
Google Gemini 3 Pro |
xAI Grok 4 |
Meta Llama 3.1 405B |
Mistral Magistral |
|---|---|---|---|---|---|---|
| AXI-1a: Intended tasks and integration scope | Partial | Covered | Partial | Partial | Covered | Partial |
| AXI-1b: Acceptable use policies | Gap | Partial | Partial | Gap | Gap | Gap |
| AXI-1c: Release date and distribution methods | Partial | Covered | Partial | Partial | Covered | Partial |
| AXI-1d: Architecture and parameter count | Gap | Partial | Partial | Gap | Covered | Gap |
| AXI-1e: Modality and I/O format | Partial | Covered | Covered | Partial | Covered | Covered |
| AXI-1f: Licence | Gap | Gap | Gap | Gap | Partial | Partial |
| AXI-2a: Integration technical requirements | Partial | Covered | Partial | Partial | Covered | Covered |
| AXI-2b: Design specifications and training process | Partial | Covered | Partial | Partial | Covered | Covered |
| AXI-2c: Training, testing and validation data | Partial | Covered | Partial | Partial | Covered | Partial |
| AXI-2d: Computational resources for training | Partial | Partial | Partial | Gap | Covered | Partial |
| AXI-2e: Energy consumption | Gap | Gap | Gap | Gap | Covered | Gap |
| AXI-S2-1: Evaluation strategies and results | Partial | Covered | Partial | Covered | Covered | Partial |
| AXI-S2-2: Adversarial testing and model adaptations | Partial | Covered | Partial | Partial | Covered | Gap |
| AXI-S2-3: System architecture description | Partial | Covered | Partial | Partial | Covered | Covered |
| Overall completeness | 36% | 75% | 46% | 36% | 89% | 50% |
Your model documentation is different. The findings will be different.
Provectio's measurement applies to any GPAI model documentation. The applicable requirements, the coverage scores, and the force gaps depend on the specific documentation submitted.