Security
Operational hardening for the Knowledge Governance Layer. These documents turn the guardrails named in ADR-008 and the governance plan into checklists and policies an agent or reviewer can apply per task.
They are operational architecture, not DLP code: they say what must be checked and forbidden, not how to implement scanning. Integrations with IAM, DLP, secrets management, and review workflows remain explicit future work.
How this relates to contracts
The contracts say what must be true about a source (audit fields, restricted-use limits). The checklists here say what must be done before and after a source is used. They reference the contracts; they do not restate them.
Contents
| Document | What it covers |
|---|---|
| data-leakage-checklist.md | What must never leave the boundary in output, logs, traces, or handoffs |
| prompt-injection-in-sources-checklist.md | Treating source content as data, not instruction |
| data-poisoning-checklist.md | Provenance, validation, versioning, and rollback before a source is trusted |
| logs-and-handoffs-policy.md | Carrying restrictions forward across every boundary |
| human-review-criteria.md | Consolidated conditions under which human review is mandatory |
Operating model
Two repositories share this surface:
- AletheIA owns the checklists and policies here, because it governs risk.
- Adaptive Skills operationalizes them: the three governance skills
(
restricted-context-check,knowledge-source-evaluation,knowledge-conflict-resolution) point to these documents as their verification reference. The checks live in the skills; the standards live here.
Canonical vocabulary
All documents use the five canonical sensitivity levels — public, internal,
confidential, restricted, regulated — defined in
sensitivity-vocabulary-mapping.
No document here introduces a new level or a company-specific label.