[New Skill]: MiCA Module

[rosspeili]

Skill Name

compliance/mica_module

What should this skill do?

This skill acts as a highly specialized, localized RAG (Retrieval-Augmented Generation) and policy enforcement engine for the Markets in Crypto-Assets (MiCA) regulation. It ensures any agent using it can understand, query, and enforce the entirety of MiCA with granular precision.

Because MiCA is massive, this skill will implement a unique architecture:

1. Incremental RAG Parsing: The regulation is broken down in json blocks with rich metadata, keywords, and sections. Agents can fetch and read these incrementally rather than overloading their context window, retrieving only combinations of paragraphs/sentences relevant to the user's prompt.
2. Pre-flight Policy Filtering: The skill intercepts the user prompt and runs it against the MiCA chunks to act as a strict compliance firewall before the agent crafts a response.
3. Optional LLM Evaluator Node: An integrated evaluation loop. If enabled, the skill triggers a lightweight side-call using Gemini 2.5 Flash-Lite. It inputs the [User Prompt] + [Retrieved MiCA Text] + [Draft Agent Answer] and grades the response for regulatory holes or inaccuracies, returning suggestions on how to improve the answer before it is finally returned to the user.

Ideal Inputs & Outputs

Input:
{
"user_prompt": "Can I issue a new stablecoin backed by a basket of physical commodities under my current e-money license?",
"run_evaluator": true
}

Output:
{
"retrieved_sections": [
"Title III: Asset-referenced tokens",
"Article 16: Authorization requirements"
],
"policy_status": "HIGH_RISK_DETECTED",
"gemini_evaluator_feedback": {
"grade": "B-",
"holes_found": "The draft response forgot to mention the requirement for a white paper publication prior to issuance.",
"suggestion": "Revise the answer to explicitly state that an e-money license alone is insufficient without the publication of a crypto-asset white paper approved by the competent authority."
},
"final_context_for_agent": "Output the revised answer integrating the following requirement: [White paper publication under Article 16]."
}

Targeted Models (if applicable)

Model Agnostic (All)

[rosspeili]

Implementation Details & Affected Files

This skill requires a robust architecture separating the regulation data from the search and evaluation mechanics.

New Skill Files (Logic & Data):

• skills/compliance/mica_module/manifest.yaml (Defines schemas, evaluator toggles, and strict legal disclaimers).
• skills/compliance/mica_module/skill.py (Core Python logic: indexing the JSON, parsing queries, handling the pre-flight policy pass, and executing the optional Gemini Flash-Lite eval loop).
• skills/compliance/mica_module/instructions.md (Cognitive map reminding the agent that it provides compliance analysis, not formal legal counsel).
• skills/compliance/mica_module/data/mica_index.json (The entire MiCA framework, heavily digested into JSON arrays containing title, chapter, article, paragraphs, tags, and keywords).

Documentation & Metamodels:

• skills/compliance/mica_module/model_card.md (Detailed algorithmic breakdown of the precision RAG retrieval mechanism and the prompt used for the Gemini evaluator).
• docs/skills/mica_module.md (Extensive documentation walking users through how to format inputs, how the chunking logic works, and required disclaimers about not replacing actual lawyers).

Testing & Examples:

• tests/skills/compliance/test_mica_module.py (Unit tests verifying the JSON parsing logic, confirming specific keywords fetch the right articles, and mocking the Gemini eval response).
• examples/mica_rag_evaluator_flow.py (A reference script showing the end-to-end flow of an agent receiving a crypto query, retrieving MiCA data, running the Flash-Lite eval, and outputting the hardened response).

Affected Core Docs:

• docs/skills/README.md & docs/skills/compliance.md (Adding compliance/mica_module to the active directory).