Since 1 November 2026, two white papers have been available for download on www.explicio.com, each in German and English. Both are based on measurements on real hardware and state their method and their limits.
Reproducible decisions on hardware that is not bit-stable
Large language models do not always return the same result, even at temperature=0: in the measurement, twenty identical calls on an idle server produced 17 different texts. The white paper shows where the differences come from, when a decision flips and how Explicio sets up business-critical decisions so that they stay stable. As of 20 September 2026.
Small, calibrated, verifiable: a workshop report on Explicio pair4
Before Explicio searches, it makes up-front decisions: what does the message want, is a colloquial yes really a yes, does it need the conversation so far? The white paper describes how Explicio develops its own multilingual decision models for this and runs them on the CPU of the Box, trained without customer data. On the release test sets, pair4 identifies the intent in 95.5 % of cases and reads yes and no correctly in 94.7 %. As of 6 October 2026.
Figures, raw data and test scripts for both white papers are available on the Explicio Campus page and in the Trust Center.