For weeks this journal has been publishing benchmarks, a CPU port, a fine-tuning campaign, without ever properly introducing the system behind them. We are fixing that today: meet Morrigan, the second project of the Scarlet Wolf ecosystem.
Morrigan is our research assistant, and our answer to a question that obsesses us: how far can you push an AI with zero cloud, zero GPU, zero API key, on the machine you already own?
Concretely: an ordinary laptop, 16 GB of RAM, no graphics card. Everything runs on it, locally, text generation included. Not a single request leaves the machine. It is the most radical version of the Scarlet Wolf thesis: the AI that belongs to you, pushed all the way down to the hardware.
The project rests on three bets.
Morrigan does not generate with a transformer, the architecture behind every major model on the market. It generates with RWKV-7, a recurrent network of 2.9 billion parameters: linear cost, constant memory, and inference that remains dignified on a plain CPU.
This choice is not a dogma, it is a working hypothesis: if intelligence per watt is the metric that matters for local AI, then recurrent architectures deserve to be pushed hard. We push them, and we measure what it yields, in both directions: our fine-tuning campaign on an RWKV embedder ended with a verdict against our own thesis, and it is published with the same level of detail as the wins.
Morrigan's founding bet fits in one rule: zero structural hallucination. The assistant works in strict RAG over its federated documentary corpora: code documentation, man pages, system administration, networking, security, but also labor law, tax doctrine, RFCs, administrative procedures, nutrition, cooking. Twelve corpora, about 250,000 passages, each behind its own reliability threshold. If retrieval finds no context judged trustworthy, the generation model is never called: Morrigan answers "I don't know".
This is not an instruction in a prompt, it is the architecture. A prompt can be worked around; a model that is not called cannot invent.
And it is measurable. On our production-conditions benchmark, 22 out-of-corpus questions out of 24 are refused, where the market's reference embedder lets 7 through. "Knows how to say I don't know" is not a slogan, it is a column in our result tables.
Morrigan advances by benchmark: eight embedders compared on our real queries, external judges, decision gates written before seeing the numbers. And everything gets published, including what fails: the negative results, the training runs that regress, the bug that silently collapsed our embeddings. A result without its failures is an advertisement, not a result.
This discipline has already produced concrete things for the ecosystem: an open-source CPU port of an embedding model that only existed in a GPU version, a fix reported to the model's maintainer, and the embedder selected by the benchmark now also powers part of Gungnir's memory.
Morrigan once lost 38,000 Wikipedia passages. On purpose.
We had added an encyclopedic corpus to the federation, and we measured what the relevance gate did with it. The result: on encyclopedic text, a tangent trap ("1998 World Cup" retrieving the 2002 World Cup article) scores HIGHER than a genuine match ("Napoleon" retrieving the Napoleon article): 0.91 versus 0.85. No threshold can separate the two. The threshold that blocks the trap also kills the right answer.
On the technical corpora, the same gate separates cleanly. The problem was not the tuning, it was the corpus: encyclopedic text is saturated with near-neighbors that look exactly like answers.
So we deleted the corpus. Morrigan lost small talk about emperors and football, and got back the only property that matters in a work tool: when it answers, it is grounded; when it cannot ground, it refuses. Coverage is a vanity metric. Trustworthiness is the product.
The story has a second act, and it says the same thing. Wikipedia has since begun its return, this time curated by theme (history, geography, science) instead of randomly sampled. First curation attempt: articles sorted by popularity. The evaluation's verdict: rejected as well. Sorting by page views dropped the canonical articles (the Canberra article fell below the cut, and "what is the capital of Australia?" answered Sydney). The curation was redone by sorting on the centrality of the encyclopedic graph, the incoming links, and the three corpora will stay in quarantine until the evaluation goes green. Nothing enters the federation without passing the judge. That is the whole difference between a corpus you endure and a corpus you choose.
Two projects, one thesis, two fronts:
Morrigan's thesis is public, its benchmarks are public, its CPU port is open source. Its internal architecture is not: that is a deliberate choice, the lab protects its recipe while it is being built.
Partly. The benchmarks, the method and the CPU port of the embedding model are public (Apache-2.0, on Codeberg). The core of the system remains private at this stage: the thesis is public, not the recipe.
An off-the-shelf laptop: 16 GB of RAM, an i5 processor, no graphics card. It is the project's official test bench, precisely because it is the machine everyone already owns.
No. Morrigan is a research project. The ecosystem's product is Gungnir, the sovereign assistant in production; Morrigan feeds it with everything it discovers.
Cost: an RNN has linear inference and constant memory, which matters enormously on CPU. It is a research hypothesis, not a religion: when measurement proves the thesis wrong, we publish it, and the tool in production remains the best measured candidate.
Morrigan is part of the Scarlet Wolf ecosystem: sovereign AI tools, built to run on your machines, under your control. The full thesis is here, and the technical detail, failures included, is in the journal.
