01 / BRIEF
A wiki that maps Brazilian jiu-jitsu into one interconnected graph: techniques, positions, transitions and concepts, each page linked to the ones it flows from and into, the way positions actually flow on the mat.
Jiu-jitsu is not a list of moves, and every list-shaped resource fails the same way: it can tell you how to do a thing, but not where that thing sits. Which position it comes from, what it gives up, what answers it. A graph can hold that. Nobody wrote it page by page. It grows.
02 / HOW IT WORKS
The loop. A local model runs in a generation loop on a home GPU box, under a systemd timer that never stops. Each cycle it gets exactly one gap in the graph, a position with no escape written, a submission with no defense linked, and it writes one page.
The gate. The page does not count yet. A verifier the model does not control decides whether it holds: does it link to positions that exist, does it contradict a page already written, does it answer the gap it was given. A reject goes back with the error attached, not with a scolding. Then the loop asks the graph for the next gap and starts over.
Why it compounds. Each accepted page changes the shape of the graph, which changes what counts as a gap, which changes what gets written next. The corpus is not a pile of articles generated in a batch; it is the residue of a process that has been running for months and keeps finding its own next question.
The human. The pages get pruned by someone who actually trains. That is the split that works: the machine has infinite patience for coverage, the practitioner has the judgment for what is real on the mat and what only works in theory, or on a partner who is being polite.
03 / DECISIONS
One gap per cycle. Asking a model to write a wiki produces slop. Asking it to fill one specific hole, with the surrounding graph as context, produces a page worth keeping.
The verifier is not optional. Without something outside the model deciding, the loop is a machine for generating confident text at scale. The verifier is what turns volume into a corpus.
Local, not hosted. The loop runs for months. That is only sane on hardware that is already paid for, which also means the model has to be small and the harness has to be good. The constraint is the point.
The graph is the memory. No prompt carries the state of the project. The next task is derived from what exists, which means the process survives restarts, model swaps and my own absence.
04 / STATUS
1,200+ pages and growing. The loop has not stopped in months.