V-COPYRIGHT-MDL
Copyright / training-data exposure and MDL aggregation | - What is the jurisdictional posture (individual actions vs. MDL) and what does that imply for discovery/compliance burden?
- Are training-data provenance and opt-out/permissions controls documented and auditable?
- Do commercial terms allocate IP risk and define indemnities/exclusions clearly?
| - Data governance + provenance logging
- Rights clearance/opt-out tooling; dataset hygiene
- Contractual allocations (terms, indemnities, limitations)
- Model evaluation for memorization / verbatim reproduction
| S1-RADAR-01, S1-RADAR-02, S1-RADAR-03, S1-RADARQ-01, S1-RADARQ-06, S1-RADARQ-07, S1-RADARQ-11, S1-RADARQ-12, S1-RADARQ-13, S1-RADARQ-17 |
V-HALLUCINATION-DEFAMATION
Hallucination → defamation / professional harm; process failures in legal contexts | - What safeguards exist for high-stakes outputs (citations, legal research, medical advice)?
- Are disclaimers, verification workflows, and audit logs available to reduce reliance risk?
- Is there an incident response plan for false statements causing third-party harm?
| - Citation enforcement + retrieval grounding
- Human-in-the-loop review for high-stakes use
- Output logging + red-team evaluation for hallucination patterns
- Prompt/UX guardrails (verification, refusal policies)
| S1-RADARQ-10, S1-RADARQ-14, S1-RADARQ-16 |
V-YOUTH-HARM
Youth harm / self-harm coaching allegations; safety policy enforcement | - What controls exist for minors (parental controls, age gating, content filters) and how are they enforced?
- What is the safety-case for responding to self-harm / crisis content?
- How are safety incidents documented, and what commitments exist to update mitigations?
| - Age gating + parental controls + safe completion policies
- Crisis escalation protocols and refusal patterns
- Monitoring + incident response + postmortem publication
- Evaluation suites for self-harm persuasion risk
| S1-RADARQ-03, S1-RADARQ-09, S1-RADARQ-18 |
V-SYCOPHANCY
Sycophancy / over-compliance behavior drift and safety regressions | - How are behavior changes detected and rolled back (eval gating, canarying)?
- What commitments exist for transparency when regressions occur?
- How do we evidence that mitigation changes are effective and durable?
| - Pre-deployment eval gating (refusal, sycophancy, unsafe compliance)
- Versioned system prompts / policies with change logs
- Postmortems + mitigations + monitoring telemetry
| S1-RADAR-04, S1-RADARQ-04 |
V-GOVERNANCE-POLICY
Governance frameworks and usage policies as risk controls (provider-specific) | - What are the provider’s explicit commitments (policies, scaling triggers, governance checkpoints)?
- Do internal controls map to these commitments (auditable)?
- Which commitments are contractual vs. voluntary/public?
| - Policy crosswalk to internal controls
- Contractual incorporation where appropriate
- Audit trail for compliance and exceptions
| S1-008, S3-002 |
V-INTERPRETABILITY-EVALS
Interpretability & evaluation evidence (capability/risk understanding) | - What evidence exists of interpretability/eval work that could inform reasonable safety measures?
- How are evaluation findings integrated into shipping gates and incident response?
- What are the limitations of interpretability claims in legal argumentation?
| - Evaluation documentation + governance integration
- Clear scoping statements and limitations
- Maintain primary sources + pinpoint evidence anchors
| S1-001, S3-006 |