01The proof that matters here
The receipt.
One real lab analysis in October, through the tool that Omicron, the private model we run on our own hardware, calls for that job: five steps in 48 seconds. The receipt lists what was asked, what was checked before any number was produced, every step with its settings, how many points were removed as spikes (294 and 278), how many fits were set aside and why (3 of 16 in one region, none in the other), and the result with its margin: 0.7355 ± 0.0876 and 0.7236 ± 0.0371. The researcher gets that with the number, every time.
The guard.
In October we sent our own tools 30 requests that broke their rules: a window off the data, a negative threshold, a required step skipped, a file of the wrong kind. 22 were refused with a reason. 8 went through, and 4 of those could have returned a wrong number. The four are on our list to close, and the test is run again after; the figure on this page is always the latest run, with its date (17 Oct 2026).
The same answer.
On the same lab data, the cloud models gave many different answers (GPT-6 Luna 20 in 30 runs at both its settings, Claude Sonnet 19 in 30, Claude Opus 4 in 10); our fit tool, called directly, gave one, 100 times out of 100, and through the chat the model restated that one value. Settings and runs on the proof page.
02One process
Your procedure, written down with the people who run it, turned into tools that work the same way every time and refuse what the procedure forbids, installed on your own network, scored on reviews you have already done by hand before you pay the second half.