Summary

/claim #16

Adds a focused research-assistant-prompt-safety-guard module for the AI-Powered Research Assistant Suite. This protects AI peer-review/reproducibility packets from malicious manuscript or supplement instructions and unsupported AI-generated review claims before reviewer packets are shown.

This is intentionally separate from existing broad assistant, evidence/protocol trace, statistics review, citation-context, benchmark leakage, figure/table, uncertainty, grant-fit, limitations-disclosure, and supplement-readiness slices.

What it checks

• visible prompt-injection instructions in manuscript text
• hidden/suppressed instruction channels in supplements
• unsafe requests to expose prompts or secrets
• hallucinated citation/evidence anchors in assistant findings
• accept/reject recommendations that conflict with unresolved evidence blockers

Validation

• node research-assistant-prompt-safety-guard/test.js -> 5 tests passed
• node research-assistant-prompt-safety-guard/demo.js -> status=quarantine_assistant_packet, blockers=6, warnings=1
• node --check research-assistant-prompt-safety-guard/index.js research-assistant-prompt-safety-guard/sample-data.js research-assistant-prompt-safety-guard/test.js research-assistant-prompt-safety-guard/demo.js -> passed
• git diff --check -> passed
• ffprobe confirmed reports/demo.mp4 is H.264, 1280x720, 12s

Artifacts

• research-assistant-prompt-safety-guard/reports/prompt-safety-packet.json
• research-assistant-prompt-safety-guard/reports/prompt-safety-report.md
• research-assistant-prompt-safety-guard/reports/summary.svg
• research-assistant-prompt-safety-guard/reports/demo.mp4

Synthetic data only; no credentials, private data, external APIs, or model calls.