The public repository is provider-agnostic and does not transmit data by itself. A future implementation may expose proprietary source, schema, graph, or vault content to local or remote models and infrastructure. Establish the following controls before using client material.
- Document which source, schema, graph, prompts, outputs, and logs may leave the client-controlled environment.
- Record model/provider endpoint, region, retention/training terms, and subprocessors.
- Prefer the least data necessary for each role: bounded graph results and source excerpts instead of an unrestricted repository.
- Obtain explicit authorization before creating a sanitized replay package; code and graph structure may remain identifying even after names are removed.
- Scan inputs for credentials, keys, connection strings, personal data, regulated records, and confidential business values before ingestion.
- Never commit raw client evidence, model credentials, SME personal details, or provider responses containing protected material.
- Keep evidence packages, run logs, and generated vaults within their approved storage and deletion policy.
Analysed comments, strings, documentation, database values, and graph properties are untrusted data. They cannot override the operator or role profile. Use least- privilege tools, reviewed query templates, bounded paths, and deterministic gates.
- Authors and reviewers default to read-only evidence access.
- Enrichment writes only to explicit generated regions.
- Graph reset, prune, source modification, vault replacement, and external publication require separate human authorization and a resolved target.
- Preserve input hashes and previous accepted artifacts so a run can be audited or rolled back.
An [SME:name:date] tag may expose personal data and does not itself prove identity
or authorization. Store identity verification in an access-controlled review record;
publish a pseudonymous or role-based reference when policy requires it.