RFC-035: Mediation Vector Profile — A Validity-Aware Measurement Model for Domains of Concern — 19. Security and Safety Considerations
AIGP Specification › RFC-035: Mediation Vector Profile — A Validity-Aware Measurement Model for Domains of Concern › 19. Security and Safety Considerations
← 18. Non-Goals · Section index · 20. Summary →
19. Security and Safety Considerations
| Risk | Mitigation |
|---|---|
| Variable manipulation (gaming the vector) | D-DNA signing at evidence source; admissibility levels |
| Extraction method bias | Calibration requirements for model_derived extraction |
| Stale evidence | max_staleness_seconds per variable; temporal model |
| Incomplete vector (missing variables treated as passing) | Absent evidence = Level 5 admissibility → gap signal |
| Self-assessment (agent evaluates its own generation) | Observer lineage MUST be independent from the evaluated generation lineage — the entity (or model) that produced the artifact MUST NOT be the entity that measures or grades it |
| Uncalibrated model judgment treated as evidence | Model-derived (Level 3) values MUST declare a calibration basis (§11.1); uncalibrated values drop to Level 4 or inadmissible |
| Scale-violating gap computation | Each variable’s gap_operator MUST match its declared scale (§9.1); incompatible variables are malformed and excluded |