Joan-Marc Fisa of BiVelio Inc submitted a paper introducing Caleta to arXiv on 6 November 2026. The abstract describes an Apache-2.0 process supervisor and evidence layer for local large language model (LLM) inference, built around an existing backend. It separates claims about integrity, provenance, privacy controls, and reproducibility into explicit contracts.[1]

According to the abstract, Caleta combines versioned manifests, exact-size and SHA-256 artifact verification, controlled backend construction, a restricted OpenAI-compatible proxy, execution records without prompts, benchmark reports, and a small conformance suite. The paper distinguishes artifact integrity from authenticated provenance, and privacy controls from operating-system isolation. Signed manifests, replacement between checking and use, and operating-system containment remain identified challenges.[1]

The author reports static checks, 22 unit tests, deterministic end-to-end tests on Ubuntu, macOS, and Windows, and a real-model experiment on Windows CPU using Qwen2.5-0.5B-Instruct Q4_K_M with backend build b11981. Three benchmark invocations reportedly yielded 15 successful warm requests without failures. The reported median time to first token was 18.066 milliseconds and median decode throughput was 54.394 tokens per second. The abstract explicitly limits these measurements to that configuration; they establish neither cross-engine superiority nor output quality, energy efficiency, or performance across hardware.[1]

[1]: https://arxiv.org/abs/2611.10324

Sources

  1. Caleta: An Evidence-Oriented Trust Layer for Reproducible Local Large Language Model InferencearXiv · 6 Nov 2026

Check or correct this page

Verification record

The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.

8237214a54673b2270a9a4061f81b9b557c114db5bb3bac747c9499de81720c3