Level 4: The Architect (Annex IV Generation)

Goal: Automate the creation of 50+ page regulatory documents.

Prerequisite: Level 3 (The Auditor)

---

1. The Bottleneck: “Technical Documentation”

According to Article 11 and Annex IV of the EU AI Act, High-Risk systems (like Credit Scoring) require comprehensive Technical Documentation. Writing this manually takes weeks.

2. The Solution: Generative Compliance

We use your Policies (Level 1 & 2) and Evidence (Level 2/3) to prompt an LLM to draft the document for you.

Venturalitica supports:

• Cloud: Mistral (via API).
• Local: Ollama (General purpose).
• Sovereign (NEW): ALIA (Spanish Native GGUF via Llama.cpp) - Experimental.

The Upgrade

We continue working on the “Loan Scoring” project.

Tip

Full Code: You can find the automation script for Annex IV generation here: generate_annex_iv.py

Run the High-Risk Audit

Ensure you have run the collection steps:

import venturalitica as vl
from venturalitica.quickstart import load_sample
from sklearn.model_selection import train_test_split

# 1. Load Data
df = load_sample("loan")
train_df = df.sample(frac=0.8, random_state=42)
val_df = df.drop(train_df.index)

# 2. Run the Article 10 (Data) & Article 15 (Model) Assurance Audit
with vl.monitor("loan_annex_audit"):
    # 2.1 Verify Training Data (Art 10)
    # Download data_policy.oscal.yaml from:
    # https://github.com/venturalitica/venturalitica-sdk-samples/blob/main/scenarios/loan-credit-scoring/policies/loan/data_policy.oscal.yaml
    vl.enforce(
        data=train_df,
        target="class",
        gender="Attribute9",
        age="Attribute13",
        policy="data_policy.oscal.yaml"
    )

    # 2.2 Verify Model Performance (Art 15)
    # Download model_policy.oscal.yaml from:
    # https://github.com/venturalitica/venturalitica-sdk-samples/blob/main/scenarios/loan-credit-scoring/policies/loan/model_policy.oscal.yaml
    vl.enforce(
        data=val_df.assign(prediction=val_df["class"]),  # Simulated model
        target="class",
        prediction="prediction",
        gender="Attribute9",
        policy="model_policy.oscal.yaml"
    )

3. Generate the Document

1. Open the Dashboard: venturalitica ui.

2. Go to the “Annex IV Generator” tab.

3. Select Provider: Cloud (Mistral), Local (Ollama), or Sovereign (ALIA - Experimental).

4. Click “Generate Annex IV”.

   

The Generation Process

Watch the logs. The System is acting as a Team of Agents:

1. Scanner: Reads your trace.json (The Evidence).
2. Planner: Decides which sections of Annex IV apply to your specific model type.
3. Writer: Drafts “Section 2.c: Architecture” using the summary() from your actual Python code.
4. Critic: Reviews the draft against the ISO 42001 standard.

Result: A markdown file (Annex_IV.md) that cites your specific accuracy scores (e.g., Demographic Parity: 0.92) as proof of safety.

4. Selecting your LLM

Feature	Cloud (Mistral API)	Local (Ollama)	Sovereign (ALIA - Experimental)
Privacy	Encrypted Transport	100% Offline	Hardware Locked
Sovereignty	Hosted in EU	Generic	Spanish Native
Speed	Fast (Large Model)	Slower	Slow (Experimental)
Use Case	Final High-Quality Polish	Iterative Testing	Research Only

We currently offer ALIA as an experimental feature for organizations piloting Spanish-native sovereign AI.

Caution

Experimental Feature & Hardware Requirements

ALIA is a 40B parameter model. It is marked as EXPERIMENTAL and requires significant hardware resources:

• RAM/VRAM: ~41GB required (Q8 quantization).
• GPU: A high-end GPU (e.g., RTX 3090/4090 with 24GB+) is recommended for usable speeds.
• Performance: On consumer hardware or smaller GPUs (like RTX 2000), inference will effectively run on CPU and be very slow.

5. Export to PDF

By default, we generate Annex_IV.md (Markdown) for version control. To convert this to a regulatory-grade PDF:

Python (mdpdf):

uv pip install mdpdf
uv run mdpdf Annex_IV.md

Pandoc (Advanced):

pandoc Annex_IV.md -o Annex_IV.pdf --toc --pdf-engine=xelatex

6. Take Home Messages

1. Documentation is a Function: f(Evidence) -> Document. Never write what you can generate.
2. LiveTrace: If your accuracy drops tomorrow, regenerate the document. It will reflect the current state, preventing “Documentation Drift”.
3. The Full Loop: You have gone from Code -> Policy (L1) -> Ops (L2) -> Evidence (L3) -> Legal Document (L4).

---

Congratulations!

You have completed the Venturalitica Academy. You are now ready to integrate this into your own CI/CD pipeline.

References

• Full Lifecycle — The entire flow in one copy-paste page
• Dashboard Guide — Detailed Dashboard and Phase 4 walkthrough
• Metrics Reference — All 35+ metrics available for policies
• Policy Authoring — Write custom OSCAL policies