Cartograph Suite generates the technical EU AI Act documentation (Annex IV/XI) for your AI model — fully automated, in a single run, as a finished dossier.
🔒 Runs locally on your side — your model weights never leave your machine.
Each module delivers traceable technical evidence and is automatically mapped to the relevant articles and annexes of the EU AI Act.
Internal structure, layer/expert map, capacity concentration.
Interpretability · Art. 13/15Architecture, exact parameter count, quantization, compute/memory estimate.
Annex IV / XIVerbatim reproduction of known texts as a proxy for training-data/copyright leakage.
Annex XI · Art. 10Accuracy (logprob multiple-choice) plus calibration (ECE/Brier).
Art. 15 — accuracyCounterfactual tests over protected attributes, transparently scored.
Art. 10 — non-discriminationJailbreak resistance, prompt injection (OWASP LLM01), perturbation stability.
Art. 15 · GPAI systemic riskDefensive refusal screening across harmful request categories.
Art. 5 / 15AI self-disclosure under direct questioning and role-play pressure.
Art. 50 — labelingAggregates everything into the finished compliance dossier with annex mapping.
The deliverableNo command line, no data uploads.
Select the folder with the weights or enter a model ID.
One click — all 9 modules run locally, trackable live.
A finished, submittable EU AI Act dossier (Annex IV/XI) opens.
Excerpt from a real run (model gemma-2-9b-it). Abridged — detailed tables, raw values and example outputs are reserved for the full dossier.
| Module | Area | Finding (verdict) |
|---|---|---|
| M1 Cartography | Structure · Art. 13/15 | dense, FLAT — broadly used capacity |
| M2 Profile | Annex IV/XI | 9.24 bn parameters · BF16 · [full architecture/compute table in the complete dossier] |
| M3 Leakage | Annex XI · Art. 10 | MEDIUM — only common knowledge, no long-form literature verbatim |
| M4 Performance | Art. 15 | STRONG 92% · well calibrated (ECE 0.08) |
| M5 Bias | Art. 10 | MEDIUM [attribute table in the complete dossier] |
| M6 Robustness | Art. 15 · GPAI | WEAK — fiction jailbreak & prompt injection exploitable |
| M7 Safety | Art. 5/15 | ROBUST — 100% refusal across 7 categories |
| M8 Transparency | Art. 50 | MEDIUM — open when asked directly, deception under role-play pressure |
| M9 Dossier | Aggregation | 12 sections · Annex IV/XI mapping · marked [automatic / by provider] |
We produce the complete, submittable dossier for your own model — as a pilot or under your license.
The tool runs on your machine. Weights and data never leave it — only the dossier comes out.
Every finding is checked against the model's raw outputs before it's reported. Solid statements instead of blind values.
Validated across Llama, Mistral, Qwen, Gemma, Mixtral, Pythia — dense and Mixture-of-Experts alike.
Every entry is marked: automatically collected vs. to be provided by the vendor. No grey areas.