What reproducibility means for an evidence corpus
Put the same question to the same corpus twice and you must get the same answer — the same sources, the same figures, in the same order. That is the whole property, and it is rarer than it sounds.
A generative system re-writes its answer on every run. Most of the time the variations are cosmetic; sometimes a source appears in one run and not the next, or a figure drifts. For casual use that is a nuisance. For clinical evidence it is disqualifying, because an answer that varies between runs cannot be audited: nobody can check yesterday's answer against today's, attach it to a decision record, or find out where a wrong number came from. A reproducible answer can be wrong — but a wrong answer that reproduces can be found, traced and fixed, and the fix can be proven to hold. An answer that shifts under your hands can be none of those things.
So the corpus behind this site is built for retrieval rather than re-generation: the stable core of an answer is stored, and what varies is only the line that fits it to the patient in front of you.
Stable before better
The working rule, quoted as it is written: "Reproducibility is the objective; 'better' is undefined without it. Judge a system's output first on whether it is stable — the same input yields the same answer — and only then on whether it is good." The reason is not aesthetic. Pursuing "better" ahead of "stable" optimises what cannot be measured: if two runs of the same question return different answers, there is no fixed thing to evaluate, so any claimed improvement is a comparison between two moving targets. Stability is what makes quality a measurable property at all.
Source: the working constitution's own standing rule, 2026.
The demonstration
Below is a live, self-contained demonstration of the property — no external requests, nothing stored, nothing sent. Pick a gray-zone scenario, ask, then run it twenty times: the attributed evidence core returns byte-identical on every run, while only the patient fine-tuning line moves. It illustrates the design behaviour; as the demonstration itself states, it is an illustration of the property, not the measurement — the corresponding test on a deployed system is run separately, and no deployed version is claimed on this page.
Source: the corpus programme's own reproducibility demonstration, approved for this site in 2026.
How this discipline runs through everything else built here — the guards, the planted mutations, the byte-for-byte comparisons — is described on How we build with Claude.