Research

Research

ZeroModel is not a judge, detector, or model. It is a deterministic artifactization layer over outputs those systems already produce.

The research program starts with one question:

Can deterministic Visual Policy Map artifacts improve inspection of AI traces compared with scalar dashboards, raw tables, or narrative JSON reports?

Priority lanes:

1. Agent trajectory and provenance — traces, tools, memory, evidence, intermediate claims, final answers.
2. RAG hallucination and grounding — claim/source/span support and citation-risk maps.
3. Training and continuous evaluation — tracker telemetry, checkpoint progress, regression safety, and evaluator drift.
4. LLM-as-judge reliability — disagreement, consistency, rubric failure, and human agreement surfaces.
5. AI observability — model/eval/infrastructure telemetry as one artifact surface.

Start with Agent Trajectories as Visual Artifacts.