# Determinism: what is reproducible is the decision https://www.explicio.com/en/trust-determinism/ Last updated: 2026-11-02 > Explicio serves its models on hardware that is not bit-stable at temperature=0: twenty identical requests return seventeen different texts. Flips happen only at near-ties – above a margin of two nats nothing flips –, which is why Explicio does not promise bit-identical answers. What is stable are the decisions taken in a short separate call; Explicio measures the completeness of extraction at every release. TRUST CENTER · DETERMINISM # For us, reproducible means the decision, not the wording. The hardware we use to serve our models to EU customers does not return bit-identical answers at temperature=0. We measured where that flips and moved the decisions to where the noise cannot reach them. Twenty identical requests against an idle server, seventeen different texts. That is the finding of September 19, 2026. It is here because a customer would eventually have measured it themselves. Explicio serves its models on hardware that is not bit-stable at temperature=0: twenty identical requests return seventeen different texts. Flips happen only at near-ties – above a margin of two nats nothing flips –, which is why Explicio does not promise bit-identical answers. What is stable are the decisions taken in a short separate call; Explicio measures the completeness of extraction at every release. - 20 runs per cell - 5,478 positions above 2 nats, 0 flipped - dense model 60 of 60 bit-stable - **Measured**: September 19 and 20, 2026 - **Hardware**: NVIDIA GB10 · vLLM 0.24.0 (NGC 26.07) - **Model**: Qwen3.6-35B-A3B in NVFP4 - **Whitepaper**: [PDF, 14 pages, English](https://www.explicio.com/dokumente/explicio-determinismus-en.pdf) - **Whitepaper**: [PDF, 14 pages, German](https://www.explicio.com/dokumente/explicio-determinismus-de.pdf) LIMITS FIRST We do not promise a bit-identical answer for the same input. After the redesign the structural decisions were stable in every checked case (clarification, question, language, the edges used for evidence), and we measure the completeness of extraction at every release instead of promising it. Measured is one hardware platform, one container, one model family. FOUR NUMBERS ## What has been measured. Every number has a raw file. The tables come from scripts, not from hand. LADDER 17 of 20 distinct texts in 20 identical requests against an idle server, production recipe. No serving switch changes that: without prefix cache, without speculation, with a bf16 cache, all together – 15 to 18 of 20. FLIP THRESHOLD 0 of 5,478 positions with more than two nats between the best and the second-best token at which a run flipped. Flips happen only at near-ties. SOURCE 60 of 60 bit-identical runs of the dense sibling model in the same format and container. The noise sits in the NVFP4 kernel that mixes the experts: shown in situ in the first layer, and with fp8 weights on a different kernel the same model is bit-stable. AFTER THE REDESIGN 8 of 8 production cases with a stable decision across three runs each, since the decision is taken in a short call of its own. THREE LEVELS ## Three levels of reproducibility. Whoever says “reproducible” should say which level is meant. We say it for each. - BIT-IDENTICAL 01 The same character sequence What temperature=0 promises. Available on our hardware: with fp8 weights on a different kernel for 30 to 45 percent of throughput, or with the dense model at five to eight times the compute time. That is not how we run it today; the trade-off is in the whitepaper. - DECISION-IDENTICAL 02 The same decisions Clarify or not, question or knowledge, which edges. This is the level that counts for a knowledge-management system. After the redesign we measure it in every checked case. - MEANING-IDENTICAL 03 The same thing in other words Not measurable enough to build a gate on. So we do not build one on it. WHAT WE BUILT ## The decision goes where the noise cannot reach it. PATTERN Decide short, extract long One call of about forty output tokens decides: clarification, question, language. A second, long call extracts nodes and edges. A short output has few positions at which a tie can arise. CONTRACT A contract asks only for what exists Three of the four causes of the case sat in our own prompt contracts: an unlabeled date became a page number, a label became a parenthesis, a citation rule cited where there was nothing to cite. Fixed, with a test wall in front. RELEASE Completeness is measured, not gated The extraction still flutters; a gate that turned red on it would always be red on this kernel. The probe counts the deviation and writes it into every model release. RULE Two engineering rules A decision call must not sit on a contract boundary. And where a prompt gets the same thing wrong three times, the format is at fault, not the wording. READ ON ## Where it is written down. [WHITEPAPERReproducible Decisions on Hardware That Is Not Bit-Stable→14 pages: method, results, redesign, reading for audits and the EU AI Act, limitations. English and German.](https://www.explicio.com/dokumente/explicio-determinismus-en.pdf) MEASUREMENT REPORT The ladder, rung by rung The full report with every cell, the failures and the incident the measurement itself caused sits in the platform repository under docs/messungen/determinism (in German). TOOLS Probe, ladder, logprob probe, harness Four tools without dependencies, raw data per run, a public prompt set of twenty synthetic notes. A public repository for them is in preparation. QUESTIONS ## Whoever wants to recompute this gets the raw data. Auditors, customers with their own measurements, and universities that want to build the invariant kernel receive raw data, tools and the thesis assignment on request. [Request the raw data](https://www.explicio.com/en/contact/) [Back to the Trust Center](https://www.explicio.com/en/trust/)