Public docs

AI Governance

Martenweave has a simple AI rule:

AI proposes.
Validators verify.
Humans approve.

This is the safety model.

What AI May Do

AI may help with:

• extracting evidence from notes, tickets, validation reports, and workshop summaries
• drafting PatchProposals
• drafting GitHub issues
• suggesting affected model objects
• explaining impact from structured model context
• preparing validation or review checklists
• summarizing handover context for AMS/support teams

What AI Must Not Do

AI must not:

• silently mutate canonical model files
• bypass deterministic validation
• mark a finding as validated without evidence
• invent SAP technical facts
• delete approved objects without explicit review
• treat raw datasets as model truth
• send raw client datasets to an external provider by default

Why This Matters

AI is useful at turning messy project evidence into draft structure.

AI is not a reliable source of model truth by itself.

Martenweave keeps AI inside a controlled change workflow:

1. Start from bounded context or evidence.
2. Create a structured PatchProposal.
3. Validate IDs, types, references, and domain rules.
4. Review impact and diff.
5. Approve or reject as a human action.
6. Apply only approved changes.
7. Record the change trail.

Good AI Outputs

Good outputs are reviewable:

• “This note appears to affect ATTR-CUST-SALES-CUSTOMER-GROUP.”
• “Create a draft issue for missing mapping evidence.”
• “Propose a value-list update, status review.”
• “Explain downstream impact for FEP-S4-KNVV-KDGRP.”

Bad outputs pretend to be truth:

• “I updated the model.”
• “This SAP mapping is correct” without evidence.
• “Delete these canonical files.”
• “Validation passed” without deterministic validation.

Martenweave is built for the first category.