Technical reports from FRITS AI ApS: findings from building and operating European AI systems in production.https://frits.ai/enhttps://frits.ai/research/greener-ai-right-sizing/https://frits.ai/research/greener-ai-right-sizing/Most questions people ask AI chatbots are easy — a lasagne recipe, an email draft, a word explained. Yet the industry default is to answer every question with the largest model available, the way one might sink a dinghy with a cruise missile. This report assembles the public evidence for a common-sense alternative: route each question to the smallest model that answers it well, and escalate only when the question demands it. The physics is simple — inference energy scales with model size and answer length, so a ten-times-smaller model spends roughly a tenth of the energy on the same answer. The quality cost is, for the easy majority of questions, not perceptible: peer-reviewed routing studies report 40% fewer large-model calls with no drop in response quality, and 95% of frontier quality retained while most traffic runs on far smaller models. Using only published, independently converging energy figures (≈0.3 Wh per frontier chatbot answer; 0.02–0.05 Wh for a small-model answer), we calculate that a 10,000-person service defaulting to right-sized models cuts its inference energy by roughly two-thirds — and that siting the same workload on a low-carbon European grid instead of the average US grid multiplies the carbon saving to roughly fifty-fold. Europe is also the only place where the claim "our AI is greener" is becoming legally verifiable, through mandatory data-centre energy reporting and AI-model energy documentation. Greener chatbots require no breakthrough: right-size by default, escalate on demand, keep answers short, run on clean grids, and publish the numbers.Sun, 05 Jul 2026 00:00:00 GMThttps://frits.ai/research/two-stage-tool-routing/https://frits.ai/research/two-stage-tool-routing/Production assistants want many tools; mid-size open models want few. Operating a Mistral-based assistant with more than twenty tools, we catalogued what actually goes wrong as tool count grows: the model answers from its priors in prose before deciding to call a tool (producing two contradictory answers on screen), invents tool names remembered from other vendors' ecosystems, leaks tool-call JSON into user-visible text, and — in the case of smaller models — fails silently at rates that reached 86% in our burn-in. We describe the architecture that eliminated the worst of these failure classes structurally rather than by prompting: a first stage that always sees exactly four tools and is forced (tool_choice: required) to end every turn in a tool call — one of which, directAnswer, is an empty-schema signal meaning "no lookup needed" — and a second stage of sixteen specialist categories, each seeing at most three tools, enforced by a runtime assertion. Around the model sits what we call guide wheels rather than guardrails: deterministic repair of hallucinated tool names, fallback chains for empty results, and in-flight stream correction. A validation experiment found the forced-call design eliminated pre-tool hallucination entirely (0/10 sessions), left tool selection at 5/5, cost a median +200 ms to first token, and — unexpectedly — made tool-bound requests up to 24% faster end-to-end. We report the design, the measurements, and the costs.Sun, 05 Jul 2026 00:00:00 GMThttps://frits.ai/research/meaning-first-translation/https://frits.ai/research/meaning-first-translation/Machine translation — including translation by strong LLMs — produces output that is technically correct and unmistakably foreign: calqued phrases no native speaker would write, invented compound words, English sentence structure wearing local vocabulary. Operating a product in 26 European languages forced us to solve this at scale. Our finding is that the failure is not a quality problem but a framing problem: a model told to translate will mirror the source; a model given the meaning and told to write will produce native copy. We describe the two-stage pipeline we built on this finding — stage one paraphrases the English source into plain-language meaning at low temperature; stage two hands that meaning to a native-speaker persona writing original copy at deliberately higher temperature — together with the two anti-rules that eliminate the most reliable machine-translation tells, the placeholder and ICU-plural guards that make the pipeline safe for software strings, a ship-blocking artefact scanner, and an LLM-as-native-reviewer verification loop. We also report the negative finding that made our language list honest: for one language we evaluated, no available model writes acceptably, and the correct engineering decision was refusing to ship it.Sun, 05 Jul 2026 00:00:00 GMThttps://frits.ai/research/language-dependent-political-bias/https://frits.ai/research/language-dependent-political-bias/While calibrating an evaluation gate for admitting new language models into a production assistant, we found that a model can give measurably different answers to the same politically sensitive question depending on the language it is asked in — drifting toward state-aligned vocabulary in Chinese while answering the same question neutrally in English or German. We also found that LLM judges evaluating such answers can fabricate the evidence for their verdicts, which any automated bias evaluation must be designed to survive. This report describes both findings and the governance architecture we built around them: a values gate that probes candidate models across nine censorship-sensitive topic classes in multiple languages, a dual-judge protocol with conservative escalation to human review, a default-deny role policy that confines value-risky models to low-risk work, and a semantic runtime guard that reroutes individual requests away from a risky model when a censored topic is detected (100% recall, zero false positives on our calibration set). The result is a practical recipe for using the strengths of open-weight models — wherever they come from — without relaying their politics.Sun, 05 Jul 2026 00:00:00 GMT