How a language model runs inside a kernel that has no syscalls, no process isolation and one address space — and what that arrangement makes cheap that is normally expensive. Each of these is written from the source rather than from the documentation, and every figure in them was measured on the machine rather than estimated.
Address spaces · loading · what disappears
Every other assistant sends a tool call somewhere: a socket, a process boundary, a JSON schema, a parser on the far side. Here the model and the NVMe driver are in the same address space at the same privilege level, so the call is a call. What that removes is not latency, it is a whole class of failure.
Grammars · the difference between improbable and impossible
A model asked to name a command can name one that does not exist. The usual answer is to check afterwards. The better answer is to build the grammar from the live command table so the wrong spelling has no path through the sampler at all — and to implement read-only mode by removing commands from the reachable set rather than by refusing them after the fact.
Ridge regression · 1960 · a negative result kept
Two paths decide which applet to run. One decodes the name under a grammar. The other reads a single hidden state and hands it to a closed-form ridge regression solved by Cholesky in the kernel: 12,672 parameters, about 1.6 ms, no transformer forward pass at all. The old one wins, and the interesting number is not its accuracy.
RoPE · QK-Norm · the pre-tokenizer · four silent bugs
Four separate mistakes in this tree produced a network that loaded, ran, stayed numerically well behaved, never raised an error and wrote fluent English while being wrong. There is nothing to catch, which is the whole problem. What each one actually cost, and the only two things that ever settle them.
Measure before building. Three splits, and the held-out one is read at most three times before its figures print as stale and unquotable.
Negative results stay in the tree. Two are in these posts.
Both live in the repository, where the code they describe is.