01The proof that matters here
We gave the same lab measurement and the same plain-English request to our tool and to the AI tools people already pay for, many times over, and counted how many different answers came back. The spread to compare against: the fit itself carries an uncertainty of about ±0.03 on this exponent.
| Runs | Different answers | Spread of the exponent | |
|---|---|---|---|
| Our fit toolcalled directly (no language model in the loop) | 100 | 1 | The identical result, 100 of 100. On a second researcher's file: 100 of 100 after a fix on 18 Oct; before it, 98 of 100 (a second request arriving between two steps made the service drop the first request's cleaning step). |
| Omicron with our toolthrough the chat (the product as used) | 100 | The tool's one value in 93 of 94 completed runs (the model wrote it to four decimals in 72 and rounded it in 21); 1 run gave another value while our own fault test was hitting the same service | 6 runs stopped before the fit and gave no number |
| Claude Opus 5.5run as an agent with a shell, paid plan | 10 | 4 | 0.988–1.003 |
| Claude Sonnet 5.5run as an agent with a shell, paid plan | 30 | 19 | 0.97–1.03 |
| GPT-6 Lunahigh reasoning, through the vendor's command-line tool | 30 | 20 | 0.69–0.93 |
| GPT-6 Lunavendor default (reasoning off), through the vendor's command-line tool | 30 | 20 (in 29 usable runs; one gave no result) | 0.73–1.21 |
The better the model, the tighter the band; none of them gives the same answer every time, and that is the point: in our system the method lives in the tool, not in the model. Two honest notes. Our tool's exponent on this file sits about 0.03 from the value in our own preprint, which used a different fit window; the choice of window moves any analyst's answer by that much, and our tool fixes one choice and keeps it. For the same reason our two rows print different exponents (1.066 called directly, with the published method's window; 1.005 through the chat, with the window the tool proposes from the data): each is repeatable, and the window is printed with the result. And the comparison is between agents with a shell on the vendors' paid plans and our chat; the exact settings and the added result-line instruction are printed with every run on the proof page. The outside tools were given the data as a labelled table; ours read the instrument's own file.
Harder questions on a lab file.
We also wrote ten harder questions on an invented lab campaign file (a mean that needs a correction factor from one table applied to counts from another, a furnace time summed over the samples that pass a filter, a date counted across a calendar), and asked each tool ten times with the whole file. Omicron 100 of 100; Claude Sonnet 5.5 100; Claude Opus 5.5 100; GPT-6 Sol at its high setting 100; GPT-6 Luna 96 at its high setting and 37 at its default (all without web search). On a third file, a supply contract with tiered prices, Omicron scored 98 of 100 (one VAT rounding step, twice), the same as Sonnet; Opus and Sol 100. The point is not the ranking: on hard reading-and-arithmetic questions the private model on our hardware matches the paid cloud tools, so the choice between them is about where the file goes, not about the answers.
In use at a research centre in Crete.
In May 2026, a batch of spectra that took about two weeks to analyse by hand was done in minutes, and the result closely matched the researcher's own (our preprint, one lab). Between 20 Apr and 18 Oct 2026, 16 people outside our core team typed 8,716 messages into the platform; in any one week, 4 to 9 of them used it (counted 19 Oct 2026).
02One process
Your instrument's method, written down with the people who use it, built into a tool that runs the same way each time it is called, installed next to the instrument's own computer, and scored on measurements you have already analysed by hand before you pay the second half.