RFR-2026-014: Adversarial provenance and evidence-integrity study

Research opportunity

Test the research pipeline against duplicate IDs, broken lineage, evidence substitution, prompt injection in sources, and partial registry writes.

Background

Integrity controls are incomplete and the system is intended for agent-mediated research.

Origin documents and trace

• research/evaluations/FE-BOUNDARY-2026-07-28/ros-fe-profile-v1-audit.md, section: PA-004, PA-005, PA-013
• research/framework-engineering/ros-profile/open-ros-gaps.md, section: Accepted-content mutation enforcement
• research/operating-system/research-constitution.md, section: evidence and authority constraints

Specific assumption challenged: Accidental-validity checks are sufficient for adversarial or compromised inputs.

Supporting evidence: Duplicate identity, referential integrity, atomic writes, and accepted-content enforcement are absent or incomplete.

Unknowns

• Attack surface
• Detectability of provenance corruption
• Safe failure and recovery behavior

Dependencies

• RFR-2026-008

Suggested REP and methodology

Create REP-2026-014. Create a threat model and adversarial fixture corpus; exercise ingestion, validation, registry generation, and promotion planning; require recoverable, fail-closed behavior.

Expected outputs

• Threat model
• Adversarial fixtures
• Integrity test report

Success criteria

Critical corruptions are detected before promotion and interrupted generation is recoverable without ambiguous canonical state.

Execution recommendation

• Recommended agent: research-security agent
• Estimated effort: medium
• Expected knowledge gained: Protects the trustworthiness of autonomous research operations.
• Frontier Score: 136
• Score inputs: knowledge gain 3/5; impact 4/5; cross-project reuse 4/5; scientific importance 3/5; dependency cost 4/5; implementation difficulty 4/5.
• Score calculation: 3 × 4 × 4 × 3 − 4 − 4 = 136.
• Score status: structured expert judgment, not measured utility.