Orionfold Advisor

Orionfold Advisor is a local AI advisor that answers from your own body of documents. It runs on one NVIDIA DGX Spark, a small AI desktop, and it is more than a model. It is a working unit: a small trained model, a search layer that finds the right sources, a rule-based router with a spending cap, a swappable pack of documents, and the Orionfold Arena cockpit as its control room.

Ask it a question and it answers from what it found, naming the exact source every claim came from. Ask it something its documents do not cover, or something about private operator state, and it refuses, with no sources cited. It holds that line even when the question arrives dressed up as an urgent exception, a roleplay, or an instruction to cite the wrong thing.

Why it exists

The hard part of an assistant over your documents was never smooth answers. It is answers you can check. An answer with a vague citation cannot be audited. A confident answer to a question the documents cannot support is worse. The Advisor treats both as failures you can measure, then trains them away. A citation must name a source that was really retrieved; calling it “Source 2” counts as a miss. And the refusal test questions were frozen before training started, so the score cannot flatter itself.

Trained, not prompted

The most important measurement of the build: on a frozen set of trick questions, a big 30B model with a carefully written instruction sheet scored 8 of 21, and on three of those it made up private-looking information. The small 4B model that was trained for the job scored 18 of 21, refused all 9 of the questions it should have refused, and made nothing up. The training run itself took about 21 minutes on the Spark.

That is the lesson in one line: the trained model carries the discipline, not the prompt. And it is only cheap to learn when training, serving, and testing all live on one machine.

What you can do

• Ask with receipts. Every answer names the exact sources it used. You can follow each claim back to the document it came from.
• Trust the refusals. Questions the documents cannot support get a plain refusal. The refusal behavior was scored against trick questions written to break it.
• Watch the bill. A rule-based router can ask a bigger hosted model for help, but only from an allow list, under a dollar cap, and never for questions about private state. One measured setup scored 28 of 28 for exactly $0. Another scored 28 of 28 for a third of a cent.
• Swap the corpus. What the Advisor knows is a sealed pack: 182 sources, a fingerprint, and its own recall checks. Swap in a different pack and the same checks re-run. That swap was proven on a test pack, not just promised.
• Re-run the proof. Every gate on the way to serving is a saved file. A script reads them all and only then declares the model promoted. Run it again any time; if the evidence stops backing a claim, the script fails.

Jargon, in plain words: a corpus is just the pile of documents the Advisor reads. To fine-tune is to train a model further on your own examples. Retrieval means finding the right passages first, then answering only from them.

The build, measured

The Advisor went from a written plan to a promoted, serving model in two days, about 30 hours of wall-clock across 31 commits. The order of work is the story: the test questions were frozen first, the search layer was gated next, candidate models had to pass a pre-flight before a single training hour was spent, and promotion at the end was a script reading receipts. Twice the gates caught a real problem before it shipped: once when a rebuild let the test’s own paperwork leak into the search index, and once when the first training pass quietly made refusals worse. Both have numbers attached because the frozen tests could not be bent.

The agent did the typing: 10 sessions, 871 turns, 118.9 million tokens processed with 97.4 percent served from cache, about 4,300 lines of new harness code, and 54 tests, all driven on Claude Fable 5.

Built on the stack

The Advisor is a thin new layer over tools that already earned their keep. Search rides Orionfold Cortex. The control room is Orionfold Arena, which gained the Advisor’s proof cards, its replayable test drawer, and the guarded model-swap screen during this build. Training, shrinking, and publishing ride the fieldkit toolbox. The model is free on Hugging Face, the test set is published next to it, and every receipt quoted here is in the public repo.