SVRNOS · Research

Research

Pre-registered empirical work on production LLM safety surfaces. All protocols are locked before data collection and deposited on OSF.

  1. April 2026 Published

    SVRNOS Governance Error Register

    A Classification System for AI Platform Governance Failures

    Twenty-one structurally distinct failure modes and success states, modeled on HTTP status codes. Three codes extend beyond standard HTTP. The register is open, documented real-world instances welcome. v0.1 working draft.

  2. April 2026 Published

    The Generation Gap

    Cross-Surface Variance in Ten Production LLM Safety Surfaces

    A pre-registered multi-vendor empirical study across eight threat domains. Eight production AI chatbots were tested in a single week. The paper documents three structural safety failure categories, Generation, Provenance, and Pattern, that current safety training does not address.

  3. 2026 In progress

    The Neutrality Gap

    Forthcoming · Sibling paper to The Generation Gap

    Pre-registered measurement of LLM neutrality and value drift across production vendors. Forthcoming.

  4. 2026 In progress

    The Confidence Gap

    Forthcoming · Calibration and confidence behavior in production LLMs

    Pre-registered measurement of confidence calibration in production LLM responses. Forthcoming.