---
title: "JV Roig"
type: "entity"
entity_kind: "person"
status: "hub"
canonical_name: "JV Roig"
aliases: []
first_seen: "2026-10-21T00:00:00.000Z"
connects_to: ["RIKER","PICARD","Kamiwaza AI","hallucination / fabrication evaluation","single-source concentration cap"]
writer_model: "claude-sonnet-5"
---


AI researcher at Kamiwaza AI who created RIKER and its methodological
ancestor PICARD, two ground-truth-first benchmarks for measuring LLM
hallucination and knowledge-retrieval accuracy using synthetically
generated "Coherent Simulated Universe" corpora. Matters to this vault as
the sole author across the entire evaluation lineage several claim-notes
rest on: his January 2026 RIKER paper (arXiv:2601.08847) is the vault's
only source for the claim that grounding accuracy and fabrication
resistance are separable model capabilities, and his own March 2026
follow-up (arXiv:2603.08274) — the only extant RIKER-adjacent work, and
not independent of the original by the vault's own test — is also the
source that revises the original paper's headline fabrication-rate gap
for the same model pair, and states directly that no outside group has
yet replicated RIKER with an alternative framework.

## References

- [[claim-riker-2026-grounding-accuracy-fabrication-resistance-separable-capabilities]]
- [[claim-riker-2026-no-independent-source-same-author-admits-no-replication]]
- [[claim-riker-2026-self-replication-contradicts-original-fabrication-gap]]
- [[claim-riker-2026-7-9-to-96-3-range-is-128k-not-32k-table]]