What problem does it solve?
Catches hallucinated, chimeric, and otherwise broken references in .bib files. Runs Claude and Codex (GPT-5.4) agents in parallel using two genuinely independent verification methods, randomly assigns references to Codex so there are no fixed agent pairs, and stages an adversarial challenge round wherever they disagree. Produces an evidence-backed scorecard so no fake slips through silently. Trigger with "/kris path/to/refs.bib" or "/kris path/to/paper.tex".
Core Features & Use Cases
- Phase-driven orchestration: Phase 1 Locate/parse, Phase 2 Parallel verification, Phase 3 Reconcile, Phase 4 Adversarial challenges, Phase 5 Score/report.
- Two independent verification methods: Claude uses a deterministic API-cascade with Crossref/OpenAlex/Semantic Scholar/arXiv, while Codex applies a separate, signal-diverse check, with an adversarial round to resolve disagreements.
- Evidence-backed verdicts: Each reference yields per-source results, red flags, and a final verdict, with an annotated BibTeX output for downstream review.
- Use Case: Researchers running /kris on a references.bib to obtain a comprehensive REPORT.md and a KRIS-annotated bibliography for submission-ready documentation.
Quick Start
Audit the references.bib with /kris to generate a full REPORT.md and an annotated_bib.