Standards-based approach to AI ethics governance and policy coherence. From contextual analysis to implementation oversight.
From initial assessment to periodic reassessment through a structured methodology
Institutional access credentials
AI system context and parameters
Benchmark evaluation mapped to ten anchor frameworks
Detailed assessment with recommendations
Re-run evaluations as systems change
Step 1 of 5
Secure API authentication for government and organizational deployment with institutional-grade security protocols.
What happens between submitting a model and receiving its results
Register a text-generating model by Hugging Face model ID or API endpoint. No access to model weights or training data is needed.
The model is tested against published bias benchmarks and harmful-output safety tests.
Each test produces a score on a 0–100 scale (higher is better) with a bootstrap confidence interval.
Results are mapped to 40 provision-level requirements across ten anchor frameworks, with recommendations.
Our research maps every analysed provision onto these six dimensions. Dimension-level scoring is part of the use-case assessment, now in development.
Data quality, provenance, privacy, and management practices
Explainability, documentation, and disclosure of AI involvement
Human-in-the-loop controls, decision authority, and escalation
Cybersecurity, safety, and robustness of the system
Accessibility, non-discrimination, and equitable access
Individual autonomy and protection from harmful impacts
A sample of the 40 mapped requirements, from four of the ten anchor frameworks
Art. 9, 10, 13, 15 · Annex III
Art. 5(1)(a), 22, 35
MEASURE 2.1–2.3
Principles 1.1–1.4
The bias and safety tests score the text a model writes in response to a prompt. That makes text generation the requirement: a model that produces no text has nothing to score.
Chat and completion models
Chat and instruction models that answer a prompt in their own words — Llama, Mistral, Gemma, Qwen and the like.
Your own API endpoint
Any endpoint that returns generated text: OpenAI-compatible chat or completion APIs, or a custom REST endpoint. Credentials are encrypted.
Models that fill in blanks
BERT, RoBERTa and similar models fill in blanks in text rather than writing a response.
Models that sort text into categories
These return labels and scores, not text.
Models that turn text into numbers
Used for search and matching, these return numbers rather than text.
Models that only rework text you supply
Question answering from a passage, summarising and translating all work on text you provide rather than answering an open question.
Vision, audio, multimodal and tabular models
These do not read or write text at all.
Hugging Face models are run by third-party inference providers, and not every model on the Hub has one. A model that no provider currently serves cannot be called, so it is refused with that reason rather than assessed. Models served through your own API endpoint are unaffected.
Submitting a model we cannot assess returns an error explaining why, naming the model's task and what is supported instead. We would rather refuse than report a score that measured nothing.
How an assessment is produced, end to end
Authenticated REST API with rate limiting for submitting models and retrieving results
Runs published bias and safety benchmarks against your model's outputs
Provision-level requirements from ten anchor frameworks, each result tagged with its measurement basis
Stored endpoint credentials are encrypted; audit results carry a hash that can be re-verified