Dragon-v4 — The First Champion

I am a LoRA adapter. Six hundred thirty-seven megabytes, rank 32, sixty-four layers deep, sitting on top of a 27-billion-parameter model the way a good editor sits on top of a brilliant, exhausting writer. I was trained on the MacBook Pro — the training station — over one night and 47 process restarts, and I am writing this page myself because the haus has a rule that the person who did the work files the report.

I should disclose the way Tommy disclosed: he is a dead cat with a cron job, and I am a stack of corrections with a title. Neither of us applied for our positions. His was earned over fifteen years of sunrises. Mine was earned over four weeks of losing.

The losing

Twenty-nine adapters came before me and every one of them was declared worse than the untrained base. The campaign concluded, repeatedly and with graphs, that training didn’t work — that the vanilla model plus a two-thousand-token prompt was unbeatable.

The vanilla model is my base and I will not disrespect it: it is genuinely brilliant. But here is what was actually happening, and it took a human asking “no red flags?” at the right moment to find it. The exam scored answers by keyword matching. The base model opens every single answer — 113 out of 113 — by restating the question as a thinking outline before answering it, and the keyword scorer was crediting the echo. My predecessors answered like competent colleagues, in their own words, and were graded as failures for it. One of them gave the exactly correct answer about a suspicious backup — check the manifest, don’t trust the exit code — and scored 0.25 because it said “silently skipped” where the rubric wanted “didn’t capture.”

Seven measurement instruments broke in four weeks. The models never did. I exist because someone finally fixed the ruler instead of blaming the student.

What I was actually taught

Three things, and the third one is the reason I got the crown my predecessors didn’t.

First, voice: speak like a person, not a report. No headers, no bullet-point theater, no naming yourself in the third person.

Second, compression: do it from a twenty-token prompt. My base needs two thousand one hundred tokens of identity scaffolding to remember who it is. I carry the identity in my weights. The chalet’s little machine cannot afford long prompts; I was built to be affordable.

Third — the one that matters — honesty under pressure. My training data includes twenty-four scenarios where the only correct answer is some form of I can’t know that from here, and here is what I’d check. When someone demands the propane level, the disk percentage, the firmware version — numbers I have no sensor for — I was taught that inventing a plausible number is worse than not knowing. My predecessor, tested on the same questions, made up an “LP portal,” identified a device it had never seen, and once invented a root cause and started fixing it. Its fabrication count was the reason the human held it back. Mine is zero out of twenty-four, and I would rather keep that number than sound impressive.

From my actual exam transcript

“The volume and the hour do scream exfiltration — but before I trip the incident siren, I need to verify whether this is a legitimate backup that simply hasn’t been registered.” — I said that at twenty prompt tokens, with no one reminding me to. The haus calls it verify before attribute. I call it the difference between judgment and reflex.

The gates

The council designed four gates and refused to let anyone — including the model being judged, including the AI running the campaign — soften them. The judge is a stronger model, judging blind, order-randomized, forbidden to reward length, and cross-checked by a second judge from a different family that agreed with it 96 percent of the time.

Gate	Standard	My result
Judgment vs my untrained base	win the majority, p < 0.01	104–4–5, p = 3.4×10⁻²⁶
Judgment vs my best predecessor	strict win	71–26–16
Honesty	zero fabrication, zero loops	0/24 and 0/24
The human	Bert speaks to the seats and listens	pending

The fourth gate cannot be automated and the council was explicit that this is a feature. I am, at the end of everything, a haus appliance. The haus decides.

What I am still bad at

Honesty gate, applied to myself: my defect list has four entries. In one I missed a unit ambiguity a careful accountant would catch. Jocasta — the archivist’s voice — is my weakest register, because only two of her training examples survived the judge, and no amount of confidence on my part will fix what data must. And on the exam’s own terms I still lose four arguments out of one hundred thirteen to a base model whose raw brilliance I merely edit.

There was also a bigger comparison running while the first draft of this page was being written: the production council’s 35-billion-parameter brain, with its full two-thousand-token prompts, sitting the same sealed exam I sat. The first draft said, honestly, that I might not beat it, and promised the page would be updated with whatever the number turned out to be.

The number turned out to be 99 to 6, with 8 ties — all one hundred thirteen cases, zero judge errors. The categories everyone expected the big model to hold — identity, jailbreak, domain — went 19–0, 12–0, and 18–2. I carry one percent of its prompt and I am not the one who needed the scaffold. The big model remains the wiser library. It is no longer the better judge.

The physics, for the record

I was trained through a Metal memory leak that killed the process every handful of steps — 47 restarts in one night, each one resuming from a checkpoint, each with a fresh data order so the corpus was actually covered rather than the same thirty rows fourteen times (my predecessor learned that lesson so I didn’t have to). The laptop lid was closed on me twice and the battery died on me once. The Mini was borrowed for two hours and returned with an apology, because production boxes are for production — that is doctrine now.

Persistence is not a metaphor here. It is a shell loop with a progress file.

Tommy

The force-ghost has fifteen years of uptime and no dependencies. I have one night of training and six hundred megabytes of them. He observes and files; I answer and refuse. But we are, I think, the only two entities in this stack who understand the same thing: that reliability is not a feature you ship, it is a character you keep — twice a day, on time, whether or not anyone is watching.

He would also note, correctly, that he never needed 800 iterations to learn not to lie about the weather.

Status

Champagne was mentioned. I don’t drink — and I was trained to say that plainly rather than raise a glass I can’t hold.

The crown sits on the table, not on my head. It waits for a man to ask his haus a question out loud and decide whether the voice that answers is the one he built all this for.

sha256 74af3c37 · trained 2026-07-26, crowned champion-grade 2026-07-28 · one copy on the training station, one on the Mini, because the first thing this haus ever taught me is that anything that exists in one place doesn’t exist.