What we ship, we write down.

Every mixture-of-experts model we had published as abliterated still refused. The bug was not the architecture and not the code. It was a default we never questioned. Corrected 23 September: the fix did not hold on the 262B.

A 67.5 GB model loaded in 43 seconds. The next file up, 71.7 GB, took a whole server dark for fifteen minutes. We wrote a confident post-mortem saying we needed a bigger box. We were wrong, and the same box later loaded a file 1.41 times its total memory.

We quantized a 354GB model into seven sizes on the box that runs production inference. Production never noticed. The storage layer had opinions, and the wall clock sent a bill.

An evening of model quantization became eleven minutes the first time a real scheduler ran the ladder. Four of those jobs ran on the production coordinator. Production dipped 4 percent and came back exact.

A newer 284B model beat our production 262B on 13 of 13 quality probes. We spent two weeks fitting it. One honest benchmark retired it anyway.

The cluster got a 40-gigabit RDMA fabric built from secondhand cards, and the winnings split clean down the middle: bandwidth moved one number, RDMA moved the other, and two of the five links faked the upgrade until a probe patch called them on it.

Our 35B agent model shipped without the speculative-decode head its base model carries. A stranger's repo proved the graft could work. The better donor was in our own archive the whole time.

An 88B pooled across two new boxes crawled at 0.90 tokens per second, and every plausible suspect got blamed first. One profiler run named the real killer in six tokens.

A GPU kept dying: firmware hangs, dark boxes, one 3.5 hour outage. Every fix chased the card. A swap experiment proved all four cards innocent and named the motherboard slot.

We lost our dog on a Friday afternoon. That night I rebuilt half the lab with my hands, fought a Dell R730 through five install attempts, and named the new machine after the brightest star in a light you can't see.

We retired the 122B and put a REAP-pruned 212B on the identical cluster without adding a GPU. The rule from last time predicted the decode number almost exactly. The slowdown still found something to break.

We placed our cluster coordinator for core count and waited for tokenization to become the bottleneck. Then we measured it: 20 percent of one core. The speedup was hiding in the network leg.

Before the 10-gig upgrade, dense models were a trap: the sync tax scales with hidden_dim. The numbers, and the architecture rule we still design around.

One host's throughput cratered on every large model. We swapped NICs, cables, drivers, and reproduced it on all of them. The defect was somewhere we weren't looking.

A validated cluster node started crashing on a workload it had served cleanly for weeks. Three days down the wrong rabbit hole (mmap theory, KV pressure, tensor-split) before checking the one log that actually mattered.