SCARLET WOLF / JOURNAL
The workshop journal.
Gungnir progress, Morrigan research, sovereignty and self-hosting. No jargon, no empty promises.
Gungnir progress, Morrigan research, sovereignty and self-hosting. No jargon, no empty promises.
This journal documents the building of a sovereign AI in the open, no filter. It isn't a feed of rehashed news: it's what I learn while designing Gungnir and the Scarlet Wolf ecosystem, the technical calls I make, the walls I hit, and what it concretely changes for anyone who wants to take back control of their data.
Every article starts from a real problem, never from a keyword. The useful answer comes in the first lines, the rest develops and qualifies it. No sponsored content, no rigged demo, no embellished screenshot: when a topic has a grey area or a limit, it's written in plain sight. The goal isn't to sell you something, it's to make you able to decide for yourself.
Three threads run through the journal:
Written by Kevin G., founder of Scarlet Wolf, who designs and builds the products discussed here himself. Articles ship at the pace of the work, not a marketing calendar. To react to an article or suggest a topic, the door is still direct contact with the founder.
The seventh article predicted that no scalar readout rule would fix retractions without a referent. We were wrong. A frozen direction, learned on other dialogues, recovers 100% of the cases and breaks none, at both scales, down to tools it had never seen. The measurement, the sealed protocol, and what the perfect number does not say.
Transformer, RNN, mixture-of-experts, embedder, classifier, diffusion, speech recognition: what each family does well, what it costs, and which one to choose. With measured numbers.
Two dialogue turns, one minimal question: does the model keep track of what actually happened? The state separates it perfectly, independent of surface. The behavior ignores it and fabricates an identifier that never existed. Third wall-that-is-a-door, the sharpest of the series.
The defect came from the training data: 145 examples that call a tool against 6 that stay silent. First pre-registered prediction of the series, held at both scales. And on the same judge, reading beats teaching on cost.
No raw model knows when not to call a tool, at any scale. Its internal state does. A readout gate fixes 80% of the misses, confirmed blind and pre-registered, after a first attempt that failed and which we publish too.
The model confuses tools its own internal state separates at 99%. A training-free readout corrector buys +6 points, confirmed blind and pre-registered, and the wall of 60 falls.
We concluded that a recent checkpoint was a regression. It was our protocol. 25 measurements and a public correction.
Two answers from the same model to the same prompt can be nearly orthogonal (cosine 0.079). Noise measured over 10 seeds, and the fix.
Nine measurements stuck between 58 and 61/82: a 2.9B RNN's limit at tool calling, then a $3 7.2B that beats a 30B cloud model.
State-tuning RWKV-7 for $0.53: exact parity with Qwen3-30B on an 82-case benchmark, replicated byte-for-byte. Protocol, data and code released.
No single best open-source LLM: the one that fits your licence, hardware, language and sovereignty constraints. The SME comparison, current as of mid-2026.
We tell clients to host their AI on their own infrastructure, beyond the reach of the US CLOUD Act. Our own code was on GitHub. How we closed the gap.
The official introduction of Morrigan, our local-AI research lab: RNN generation, everything measured and published, failures included. On an ordinary laptop.
Fine-tuning a 144M RNN embedder for €0: parity with transformers in pooled evals, a verdict reversed at full scale, and the out-of-corpus refusal crown.
A benchmark of eight embedders for our local AI: a 107M beats models five times larger, plus an open-source RWKV CPU port and honest negative results.
A real ChatGPT alternative is not about the logo but about where the AI lives. The honest overview of the no-cloud options, from least to most sovereign.
Large groups have lawyers and security teams to keep cloud AI in check. An SME does not. The three concrete risks, and how to neutralize them.
Sovereign AI means an AI deployed on infrastructure you control. What it means for an SME, the real risks of cloud AI, and how we approach it.
Cloud AI assistants start from scratch in every conversation. Why they forget, and how an AI can keep your exchanges, documents and context in memory.
The GDPR doesn't ban AI, but entering personal data into a cloud assistant makes you accountable. Your real obligations, and how to stay compliant.
A datacenter in France does not put your data beyond the reach of US law. What the CLOUD Act actually says, who is affected, and how to take back control.