The First Machine We Didn't Own 21 September 2026 · 15 min · The Cortex Fleet #13 A friend brought a MacBook over for an evening. Eight minutes after enrolling it was serving claims; within a day it had found a bug that had been in our worker since August and was invisible on every machine we owned. T… Fourteen Cases, Three Defects: Letting the Workloads Tune the Scheduler 26 August 2026 · 7 min We wrote the expected route for every kind of work the cluster does before sending a single request. The first run was 6 hits, 5 misses, 3 gaps. Every miss was a scheduler defect nobody had noticed for weeks — including … The Probe That Measured Itself 23 August 2026 · 9 min · The Cortex Fleet #12 A model emitted one token 23,195 times and every health check said fine. Auditing why led to a capability-measurement system that was mostly measuring its own test harness — struct defaults serialized as findings, probes… Small Models, One Semaphore: NPU Use-Cases, a Concurrency Rewrite, and Flux.2's Ghost Dependency 1 August 2026 · 10 min · The Cortex Fleet #7 The NPU bring-up from a few posts back finally got real workloads: a chat router and speech-to-text, live. Getting there meant reverting a concurrency design after review, which surfaced a genuine scheduler bug behind fl… The Cortex Grows Senses: ROCm for LLMs, Multimodal In and Out, and Four New Agents 11 July 2026 · 11 min · The Cortex Fleet #6 Six weeks since the flux2-dev ROCm debugging saga, the fix turned out to generalize far past image generation. A tour of what shipped: ROCm for LLM serving, the broker's first multimodal-input feature, four new agent rol… Why 512×512 Works But 1024×1024 Doesn't — Debugging Flux2-dev on a Ryzen AI Max 395+ 1 June 2026 · 9 min · The Cortex Fleet #5 A two-day debugging session into why Flux2-dev generates 512×512 images fine but silently fails at 1024×1024 — Vulkan heap limits, GTT pools, IOMMU, and a GPU memory advisory call that makes buffers CPU-inaccessible at e… Scheduling GPU Workloads at Home: A Capacity Model That Scales 31 May 2026 · 8 min A rigorous capacity model for home GPU clusters — one that works on a single node today and extends naturally when new hardware arrives. Built around three state vectors, six diagnostic metrics, and the actual implementa… Getting Wan2.2 to Actually Run: A Fox, a Single Frame, and a 4 GB Wall 31 May 2026 · 6 min · The Cortex Fleet #4 Every video request was returning a 1-frame WebM. Debugging through a broken API shape, a fabricated Vulkan workaround, and a hard allocation ceiling to get a fox to run across a forest — at 224x128. Fleet Audit: Retiring Four Models and Justifying the Rest 27 May 2026 · 9 min · The Cortex Fleet #3 After a probe run that broke half the cluster, I used the wreckage as an opportunity. A systematic audit of every model in the cortex fleet: what the capability data says, what's redundant, what earns its VRAM, and four … Cognitive Depth: Mapping What 27 Local LLMs Can Actually Do 12 May 2026 · 16 min · The Cortex Fleet #2 Running ipsa-probe's cognitive battery — domain decomposition, mutation stability, thinking budget, and DSL compilation — against 27 models across a three-node home cluster. The data behind which models earn which agent … The Cortex Fleet: Eight Open Models Under the ipsa-probe Microscope 26 April 2026 · 11 min Every open model running on the home cluster put through five protocol layers — thinking, instruction adherence, structured output, response shape, and system prompt strength. The numbers behind which model goes where.