Dimension 6 — Regulatory alignment (10 points)
Assesses alignment with EU AI Act Article 13 (transparency requirements) and Annex IV (technical documentation) for AI systems. This dimension was promoted to actual scoring in v0.4 — earlier methodology versions did not score the dimension at all.
Honest scope statement (v0.4 onwards). Dim 6 scores documentation alignment with regulatory frameworks. It does NOT score the inherent dangerousness of model capabilities. A high Dim 6 score means the author has done the documentation discipline that EU AI Act Art. 13 + Annex IV ask for; it does not certify that the model is safe to deploy in a high-risk context (medical diagnosis, biometric identification, autonomous decisioning). A model that scores 10 of 10 on Dim 6 may still be inappropriate for a regulated deployment if the underlying capability is itself the regulatory risk. Read the score as "the paperwork is in order"; do not read it as "deploy this anywhere".
| Sub-rule | Points | Mechanic |
|---|---|---|
| Intended Purpose (cardData.pipeline_tag OR README "Intended Use" section) | 2 | Either signal counts |
| Known Limitations (README "Limitations" section with body) | 2 | Section presence |
| Human Oversight (README "Bias, Risks, and Limitations" section with body — proxy) | 1 | Section presence |
| Risk Management (cardData.tags includes a risk-aware marker OR Out of Scope Use section) | 1 | Either signal counts |
| Technical Specification (library_name OR pipeline_tag populated) | 1 | Either signal counts |
| Model Card Contact section present with body | 1 | Section presence |
cardData.co2_eq_emissions populated (string OR object) | 1 | Field populated |
| Get Started + library_name combo (technical reproducibility signal) | 1 | Combo |
EU-bar penalty (cardData.extra_gated_eu_disallowed = true) | -2 | Floored at 0 |
Mapping to EU AI Act Annex IV — CO2 emissions. EU AI Act Annex IV requires high-risk AI system providers to document the energy consumption of the system over its lifecycle. Hugging Face's cardData spec provides a co2_eq_emissions field (either a free-form string like "100 kg" or a structured sub-object with emissions, unit, source keys). Dokima awards one point when the field is populated with any non-null value. The point exists because filling the field is a small action that maps directly to a documented regulatory requirement — authors who do it deserve credit even when the wider Annex IV documentation is incomplete. The v0.4 rebalance trimmed the Risk Management sub-rule from 2 to 1 to absorb the new CO2 sub-rule within the dimension's 10-point budget.
Mapping to EU AI Act Art. 13 — Model Card Contact. Art. 13 requires AI system providers to make contact information available to users. Hugging Face cards typically encode this in a Model Card Contact section in the README. Dokima awards one point when the section is present and has body text. Authors who include only the heading without an actual contact (no email, no organisation, no link) do not earn the point — empty contact sections are functionally indistinguishable from no contact information at all.
The EU-bar penalty (v0.4 onwards). When a model carries cardData.extra_gated_eu_disallowed: true, the author has explicitly marked the model as not available to users in the European Union. This is an honest signal — the author tells consumers that the model cannot be deployed in EU jurisdiction — but Dokima reads it as evidence that the author has not completed the AI Act risk-classification work that would let the model be available there. Two points are deducted from the dimension's positive score; the dimension floors at zero (the penalty cannot drive Dim 6 negative). A witness drift flag (EUBarPenalty) is emitted on the score record so consumers can see exactly how many points were deducted and why.
The penalty exists because Dokima itself operates from EU jurisdiction (UK; the regulatory framework Dokima cares about most) and the score must reflect what the model is usable for in that context. A reader could reasonably interpret the EU-bar field as responsible hedging — the author is being upfront about a regional limitation. Dokima's editorial choice is the regulatory framing: a model that is structurally not deployable in the regulator's home jurisdiction is, for that consumer, a model with incomplete regulatory work. Authors who disagree with the editorial choice can read the witness drift flag and apply their own interpretation.
Why this dimension stays smaller than Dim 1 / Dim 2 / Dim 4. Regulatory documentation is the easiest dimension to game with boilerplate — an author can paste a template "Intended Purpose" section, a template "Known Limitations" section, and a template "Contact" section and earn 4 of 10 points without any actual regulatory work. The dimension is sized at 10 of 100 to reflect that limitation: documentation alignment is a real signal but it is the weakest of the six because the cost of producing it is low. The dimension's weight will increase if and when EU AI Act enforcement creates a meaningful audit trail that authors can be evaluated against — until then, Dim 6 is one signal among several rather than the load-bearing one.