Article 9 -- Risk management system
Praxara ships a tenant-scoped Risk Register covering both platform risks (vendor model drift, hallucination, data leakage) and customer-defined risks. Each entry is graded on likelihood x impact, mitigations, residual risk, owner and review cadence. Entries flow into the Compliance Pack Risk Register artefact.
Article 11 -- Technical documentation (model cards)
- One model card per (model, version) pair
- Required fields: intended use, out-of-scope uses, training data summary, evaluation, bias assessment, metrics, human oversight notes, limitations, EU AI Act risk class
- Guided 4-step wizard at
/admin/ai-governancefor non-governance staff - Segregation of duties: HIGH and UNACCEPTABLE-class cards cannot be approved by their author
- Unapproved cards do not appear in Compliance Pack exports
Article 13 -- Transparency
- Every AI output carries a "Draft only -- requires qualified human review" label, contextualised per surface (narrative / MedDRA / duplicate / literature / signal)
- Reviewer canvas always renders the source-page citation alongside extracted fields
- The user-facing Help page exposes a registry of every LLM call ever made on behalf of the tenant, by skill
Article 14 -- Human oversight
- No AI output flows to a regulatory submission, Argus case closure, or PSUR sign-off without an e-signed human approval (21 CFR Part 11 controls)
- Reviewers can edit, reject, or reclassify every extraction before approval; rejection requires a controlled-vocabulary reason captured on the audit chain
- Skill drift detection (worker scaffolded) flags accuracy regressions on golden-case sets; affected skills can be paused tenant-wide
Cost controls (operational)
Per-tenant LLM monthly spend caps with soft + hard thresholds. Calls exceeding the hard cap return HTTP 402 and are audited as cap-blocked. Soft-cap notifications email tenant ADMIN. Visible at /settings/usage.
GxP Annex 22 alignment (draft)
Praxara is designed for GxP validation -- the architecture, audit chain, e-signature controls and validation-ready artefacts produced by the Compliance Pack are structured so customer QA teams can validate AI workflows in their own quality framework. Praxara itself is not a regulator-validated product; we provide the evidence trail that customers' own QA teams use.
Provenance + change control
- Every AI call logged with provider, model, taskKey, prompt-version-hash, tokens, cost, latency, audit-row-id
- Frozen prompt versioning (PromptVersion model) -- production prompts are not edited in place; new versions go through review
- Immutable audit chain on AuditLog (SHA-256 chained checksums); tamper-evidence verifier endpoint