Skip to content

The public journal / Jul → Oct 2026

Keep the journey in the record.

The configuration is a snapshot. The journal is how we got here: controlled experiments, useful failures, model changes, and decisions that did not survive the next test.

Jul → Oct 2026 / Learning in public

The setup changed. The lessons stayed.

All journal notes ↗

Chapter 01 / Jul 11, 2026

+57.4% code / −17.0% creative

Six-position MTP changed two workloads in opposite directions. The optimization became a per-workload decision.

Read the decision ↗

19 OF 19 NOTES / NEWEST FIRST

  1. Configuration

    DeepSeek on two GB10 hosts, with the coding gate still intact

    Normal DeepSeek now serves four shared slots through TensorFold. The matched GLM panel tied on coding; capacity, cache replay, memory and maintenance have separate evidence.

    6 MIN READ

  2. Configuration

    Two GB10 nodes, eight slots, and the limits behind the settings

    A fresh inspection of both workers, October runtime probes, cache capacity, startup cost, and the Studio alias that changed models.

    5 MIN READ

  3. Measurement

    Publishing the benchmark methods and model history

    The lab’s test prompts, scoring rules, promotion criteria and model variants, including a complete Aider run and an incomplete Flash archive.

    3 MIN READ

  4. Configuration

    Follow the config, not the old dashboard labels

    Direct inspection corrected hardware identities and model placement: dual GB10 coding, Studio agents, and workstation utilities.

    2 MIN READ

  5. Configuration

    Snapshot, 1 October 2026

    Public-safe facts from the 13:51:36 CDT lab pull. Dual-Spark TP2 coder-flash resident; live generation idle.

    1 MIN READ

  6. Workflow

    The agentic repository improvement loop

    A bounded pass, independent review, verified integration, and a durable record.

    1 MIN READ

  7. Field notes

    The notebook is public. The machines are not.

    First public note. What this site is, and what it will not publish.

    1 MIN READ

  8. Evaluation

    The Studio gets its own measured roles

    A 48 GB M5 Max evaluation separated fast agent, coder, and writer roles, and exposed proxy and lifecycle bugs.

    2 MIN READ

  9. Experiment

    Production rebaseline + Mistral Small 4

    Keep the split. Lightning owns latency, Qwen owns code, Mistral stays a quality watch — not a default.

    2 MIN READ

  10. Experiment

    Nemotron-3.5-Lightning gate

    Speed win, not a coder. Challenger kept as on-demand lightning. Qwen stays primary.

    1 MIN READ

  11. Experiment

    Muse Glimmer dual-role reject

    Meta’s open 30B rejected for both coder and always-on agent. Tools were real; decode was not.

    1 MIN READ

  12. Experiment

    SWE-bench leaderboard review

    Three models ranked above the coder. None won primary. Ornith took specialist; coder-laguna retired.

    2 MIN READ

  13. Experiment

    Coder revert to Coder-Next

    Same-node A/B reversed the July 23 Laguna promotion. Qwen restored; Laguna kept on demand.

    2 MIN READ

  14. Experiment

    Laguna SM121 runtime

    Serving tuned as far as it would go on GB10. Eager + .79 allocator shipped; DFlash K=7 rejected. Reversed July 26.

    1 MIN READ

  15. Experiment

    Image gen + Laguna / BTL-3 refresh

    Mage-Flow Turbo 4B won the local 1024² envelope. The coder refresh produced a promotion that July 26 reversed.

    1 MIN READ

  16. Experiment

    Three-lane coder and the July 21 gates

    Three-lane dispatcher deployed; 2×256K qualified; Laguna S rejected; keep UD-Q4_K_XL; vision split to Nemotron Omni.

    2 MIN READ

  17. Experiment

    llama.cpp b10069 and the 256K coder bakeoff

    Newer llama.cpp promoted on both Sparks. No challenger beat Coder-Next on framework code.

    1 MIN READ

  18. Experiment

    Multimodal creative agent → Gemma-4-26B QAT

    Four uncensored multimodal candidates. Gemma-4-26B QAT kept as always-on agent. Goetia and Fable rejected.

    1 MIN READ

  19. Experiment

    Speculative decoding helped code and slowed writing

    A quiet-box Qwen3.6 comparison showed why a drafting setting should be qualified per workload.

    1 MIN READ