There is one detail that sets Distribird apart from many scientific tools built on LLMs: every model call runs locally. No cloud API receives descriptions of parameters or details of unpublished models. Only the generated search terms go out to public literature databases. For an application designed for Bayesian calibration of process-based models, this choice is not an accessory — it is the condition that allows a research group to use it without giving up control over its data.

Distribird is an agentic web application that automates the construction of informative prior distributions from the scientific literature. Given a parameter name, a physical description and domain context, a multi-agent pipeline searches the literature, extracts and weights reported values by domain relevance, and fits a probability distribution through AIC model selection. When no literature is available, the system falls back to sensible uninformative alternatives and explicitly reports both the evidence and the confidence level of every prior it produces.

The published evaluation covers 24 parameters across 10 scientific domains and compares three open-weight models — Qwen3.6 27B, Gemma 4 31B and Mistral Small 4 119B — with a single-prompt LLM baseline. On prior quality, the full pipeline matches the baseline. But the point is not point-estimate accuracy. Every prior is traced to the specific papers and values from which it was constructed. A built-in validity layer refuses to produce priors for out-of-scope requests; the single-prompt baseline returns confident but unfounded priors for those same cases in 11 of 30 model-parameter combinations. The difference is not marginal: it is the difference between a tool that can be used in a scientific context and one that produces plausible but unverifiable answers.

There is an underlying thesis here that matters more for those who follow on-premise architectures than for anyone chasing yet another benchmark. When a system must produce evidence-based decisions — like a Bayesian prior that conditions an entire process-based model — traceability, validity and data sovereignty count more than a marginal gain in point-estimate accuracy. This is not ideology; it is an operational consequence. A lab working on unpublished physical models cannot send sensitive descriptions to a third-party LLM provider. A reviewer or regulator cannot accept a prior without knowing which papers and values it derives from. A pipeline that runs locally on open-weight models answers both needs.

This choice has second-order implications for infrastructure design. Running open-weight models up to 119 billion parameters locally is not free: it requires adequate VRAM, often multiple GPUs and, in many contexts, quantization techniques to contain the footprint. The TCO of such an agentic pipeline is not measured only in cost per token, but in the ability to keep sensitive data inside the lab or enterprise perimeter, minimizing exfiltration. The fact that only search terms leave the perimeter is an architectural pattern that many self-hosted deployments are adopting beyond scientific research: not isolating the system, but minimizing the surface of data that leaves the boundary.

For those evaluating on-premise deployment of similar pipelines, AI-RADAR offers analytical frameworks on /llm-onpremise to map the trade-offs between VRAM, quantization and TCO without turning sovereignty into a slogan. The Distribird case shows that the frontier is not only between cloud and local, but between systems that declare the provenance of their decisions and systems that ignore it.