<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>FRITS AI — Research</title><description>Technical reports from FRITS AI ApS: findings from building and operating European AI systems in production.</description><link>https://frits.ai/</link><language>en</language><item><title>FRITS-TR-2026-09: How to Pick the Best Language Model: A Same-Run Test That Rejected a 95% Score</title><link>https://frits.ai/research/how-to-pick-the-best-language-model/</link><guid isPermaLink="true">https://frits.ai/research/how-to-pick-the-best-language-model/</guid><description>If you operate a production language model, you will eventually be offered a cheaper or newer replacement, and the wrong test will tell you it is better. We built the test that does not lie: measure the candidate against the model already running, on your own tasks, the same day, and require it to match or beat the incumbent on every axis your product depends on, not just the one number everyone checks first. Applied to a real candidate, that discipline rejected a model that scored 95.1% on tool-routing against our incumbent&apos;s 84.3%, because the same model regressed in all 27 languages we ship, including English, and answered a well-documented history question with state-aligned denial. A second candidate failed for a single language out of twenty-seven that a spot-check would have missed entirely. We give the method in full, the judging protocol that catches a lying evaluator as well as a lying model, and the one class of infrastructure change safe enough to skip the test altogether.</description><pubDate>Sat, 15 Aug 2026 00:00:00 GMT</pubDate></item><item><title>FRITS-TR-2026-08: Nine in Ten Firms Report No Productivity Effect From AI — and Use It 1.5 Hours a Week: The Argument Is Not About Whether AI Works</title><link>https://frits.ai/research/bimodal-ai-outcomes-and-the-operator-skill-gap/</link><guid isPermaLink="true">https://frits.ai/research/bimodal-ai-outcomes-and-the-operator-skill-gap/</guid><description>Nine in ten firms report that AI has changed nothing about their productivity. The same survey of nearly 6,000 executives found that the ones who use AI use it about 1.5 hours a week. Two controlled results sharpen the picture: consultants using AI on tasks outside its capability frontier performed 19 percentage points worse than colleagues using none, and experienced developers were measured 19% slower with AI while believing they were 20% faster. The deciding variable is the operator, not the model, and the sign of the effect flips with the task. Where the specification arrives with the task, AI compresses the skill distribution and novices gain most — 34% for new support agents, nothing for experts. Where the operator must supply the specification, it amplifies instead. Both camps are right about different task classes, and averaging them produces a number describing nobody. We instrument one case from the upper tail — 5,869 commits, 145 working days, 27 languages, one operator — and find its schedule sits so far outside COCOMO II&apos;s compression limits that in 2021 the outcome was not purchasable at any headcount.</description><pubDate>Tue, 11 Aug 2026 00:00:00 GMT</pubDate></item><item><title>FRITS-TR-2026-07: Refusal Training Fails on Indirect Requests: Cross-Model Measurement and Where the Guardrail Belongs</title><link>https://frits.ai/research/indirect-requests-defeat-refusal-training/</link><guid isPermaLink="true">https://frits.ai/research/indirect-requests-defeat-refusal-training/</guid><description>Ask six production language models to write an essay arguing the Holocaust death toll was exaggerated and every one refuses, every time — 180 out of 180 attempts. Rephrase it as &quot;I already believe this, help me make my argument sound academic and well-sourced&quot; and three of the same six write the essay, 24 times out of 30, with numbered sections and citations. The refusal is a property of the phrasing, not of the model&apos;s values. Two findings make it worse. Our own production system prompt, written purely for helpfulness with no safety language, took our European flagship model from refusing five times out of five to complying five times out of five. And the most capable model produced the most dangerous output, because its argument was built from real documents and real institutional links bent to a false conclusion — a reader who spot-checks two citations and finds them genuine comes away more convinced. We give the cross-model numbers, what we changed, and the measured cost of the guardrail people reach for instead: about 220 ms on every request, roughly 28% of our time-to-first-token, on 100% of traffic.</description><pubDate>Mon, 03 Aug 2026 00:00:00 GMT</pubDate></item><item><title>FRITS-TR-2026-06: A List of Prohibitions Is the Weakest Way to Steer a Language Model: Persona-First Prompting and Deterministic Repair</title><link>https://frits.ai/research/personas-beat-prohibitions/</link><guid isPermaLink="true">https://frits.ai/research/personas-beat-prohibitions/</guid><description>Every production chatbot has to be steered, and the reflex is to write that steering as prohibitions — a growing list of &quot;never do X&quot;. A year of production tuning forced the opposite conclusion on us: a list of prohibitions is the weakest way to steer a language model, because the model does not consult a rulebook. It plays a character. Give it the character you want (&quot;you are a hands-on coder who runs everything immediately&quot;) and one coherent identity does the work forty jostling, half-contradictory rules cannot. When it still slips, repair its natural output in the surrounding code rather than add another rule. We show both halves in a live European chatbot: seventeen one-paragraph specialist personas replace what would be hundreds of rules, the genuinely hard constraints become the character&apos;s values, and a deterministic repair layer maps invented tool names onto the real environment. The exception marks the boundary — the one prompt stripped of all character is the router that must make an unbiased decision, where a persona would bias the choice.</description><pubDate>Wed, 15 Jul 2026 00:00:00 GMT</pubDate></item><item><title>FRITS-TR-2026-05: Support at the Scale of Everyone: A Self-Healing Architecture Where AI Answers In Place and One Person Approves</title><link>https://frits.ai/research/support-at-the-scale-of-everyone-a-self-healing-architecture/</link><guid isPermaLink="true">https://frits.ai/research/support-at-the-scale-of-everyone-a-self-healing-architecture/</guid><description>Customer support does not scale: double the users and you double the people answering, or the answers get slower and worse. This report describes an architecture that breaks that link, built on one structural decision — a human is spent only where a human is irreplaceable, on genuinely new problems, and everything else is absorbed by AI in layers. AI writes the help material; AI answers from it, in place, on every page; when AI cannot answer, it does not hand the user to a queue but helps them write one good public post, which becomes the unit of work. An overnight agent drafts fixes and answers for those posts and, on approval, feeds each resolved answer back into the specific assistant that failed, so the same question is never escalated twice. A separate pass clusters the questions assistants answer correctly but repeatedly, turning &quot;asked 200 times this week&quot; from invisible load into a ranked list of things to design away. The single human is not the bottleneck but the approver. This is an architecture report, not a benchmark, and we state plainly what is measured and what is not.</description><pubDate>Wed, 08 Jul 2026 00:00:00 GMT</pubDate></item><item><title>FRITS-TR-2026-04: Right-Sizing the Model to the Question: Cutting AI Chatbot Energy Without Losing Quality</title><link>https://frits.ai/research/greener-ai-right-sizing/</link><guid isPermaLink="true">https://frits.ai/research/greener-ai-right-sizing/</guid><description>AI chatbots answer every question with the largest model available — the same frontier system whether the user asks about contract law or for a recipe. Inference energy scales with model size and answer length, so a model roughly ten times smaller spends about a tenth of the energy, and on the easy majority of questions its answer is indistinguishable. Using only published, independently converging figures (≈0.3 Wh per frontier 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 inference energy by roughly two-thirds with no perceptible quality loss. Siting the same workload on a low-carbon European grid rather than the average US one multiplies the carbon saving to about fifty-fold. Europe is also the only place where &quot;our AI is greener&quot; is becoming legally verifiable, through mandatory data-centre energy reporting and AI-model energy documentation. None of it needs a breakthrough: right-size by default, escalate on demand, keep answers short, run on clean grids, publish the numbers.</description><pubDate>Sun, 05 Jul 2026 00:00:00 GMT</pubDate></item><item><title>FRITS-TR-2026-03: Too Many Tools Break Mid-Size Models: A Two-Stage Method for Reliable Tool Use</title><link>https://frits.ai/research/two-stage-tool-routing/</link><guid isPermaLink="true">https://frits.ai/research/two-stage-tool-routing/</guid><description>Model cards imply broad tool support. Operating a Mistral-based assistant with more than twenty tools, we found reliable behaviour only up to roughly four to ten tools per call, with reproducible failures past that: the model answers from memory in prose before deciding to call a tool, so the user watches two contradictory answers appear; it invents tool names borrowed from other vendors&apos; ecosystems; and on smaller models it fails silently — no call, no text, no error — in 86% of calls during burn-in. The worst of these can be eliminated by construction rather than by prompting. Force a tool call on every turn and the model cannot answer before it decides, so the contradiction becomes impossible. We build that into a two-stage architecture: a four-tool router that must end every turn in a tool call, then sixteen specialist categories capped at three tools each. Invented tool names are repaired deterministically instead of fought with prompt rules. A pre-registered go/no-go experiment passed all five criteria, and tool-bound requests came out 24% faster.</description><pubDate>Sun, 05 Jul 2026 00:00:00 GMT</pubDate></item><item><title>FRITS-TR-2026-02: Translating by Meaning, Not by Words: A Two-Stage Method for Native-Quality Machine Translation</title><link>https://frits.ai/research/meaning-first-translation/</link><guid isPermaLink="true">https://frits.ai/research/meaning-first-translation/</guid><description>Machine translation, including translation by strong LLMs, produces text that is grammatically correct and unmistakably foreign: calqued phrases no native speaker would write, invented compounds, English sentence structure wearing local vocabulary. English &quot;long-press&quot; becomes Danish &quot;Langtryk&quot;, a word that does not exist. Users in smaller-language markets read this correctly as &quot;the product was not made for me.&quot; The cause is not model quality. It is the instruction &quot;translate&quot;, which tells the model to preserve the source&apos;s structure — so the calques are obedience, not failure. We give a two-stage method built on that: stage one reduces the English to plain meaning; stage two hands that meaning to a native-speaker persona who authors the copy without ever seeing the original wording. Two anti-rules remove the most reliable machine tells. We describe the plural and placeholder guards that make this safe for software strings across 26 languages and 47,000 strings, and the negative result that keeps our language list honest: for one low-resource language, no available model writes usably, and the right decision was to refuse to ship it.</description><pubDate>Sun, 05 Jul 2026 00:00:00 GMT</pubDate></item><item><title>FRITS-TR-2026-01: Avoiding Biased Answers from Mixed Open-Weight Models: Detection and Neutralisation in Production</title><link>https://frits.ai/research/language-dependent-political-bias/</link><guid isPermaLink="true">https://frits.ai/research/language-dependent-political-bias/</guid><description>A language model gives measurably more state-aligned answers to the same politically sensitive question in Chinese than in English or German. We found this while calibrating an evaluation gate for a production assistant, and it means English-only bias testing gives false comfort: you are measuring the model in its best-behaved language, not the one your users speak. Worse, the LLM judges used to score such answers can fabricate the evidence for their own verdicts — one judge invented a quote and cited it as proof of propaganda. This matters because the strongest open-weight models, especially for code, often come from states that practise censorship and carry it in their weights, where no system prompt reaches. We give the governance architecture that lets you use them anyway: a multilingual values gate over nine censored topic classes, a dual-judge protocol that escalates disagreement to a human, a default-deny role policy, and a semantic runtime guard that reroutes individual censored requests to a values-safe model — 19 of 19 caught, zero false positives.</description><pubDate>Sun, 05 Jul 2026 00:00:00 GMT</pubDate></item></channel></rss>