Skip to content

Campaign friction

  • CLI MCP lifecycle wrappers absent in session; use supported CLI. No new wrapper implementation required for this campaign.
  • Managed image build rejected shared default build-log directory permissions before execution. Recovery: isolated campaign build home with private 0700 build logs. Durable invariant: campaign build logs are owner-only; retain failed and successful receipts. No shared-directory permissions changed.
  • Quality launcher rejects native preflight as control-evidence/v1; the initial relative path also did not resolve. Both invalid launches sent no model requests. Corrected launch declares requested_unverified and keeps native preflight as separate functional evidence. Future verified comparison must use the documented control-evidence index; no schema rewriting of raw preflight.

  • GLM thinking-disabled preflight failed structured JSON and exposed reasoning as visible output. Enabled preflight passed all six checks. Configuration limitation retained; capacity comparison uses thinking enabled and sufficient output headroom, with reasoning counts separate.

  • SWE preflight rejected absent model-host identity. Corrected spec explicitly declares the benchmark client co-resident with model host; all seven checks pass. Evidence declared SWE preflight.

  • Image inspection refuses mutable tags. Inspected exact immutable ID from managed build receipt; retained image-identity.json. No raw Docker fallback needed.

  • Native128word C1 scout completed0/8 in1113s. Every response hit10240tokens, repeated code beyond the requested128words, and exceeded validation capture. Canary started correctly with no foreign markers; capture truncation made isolation unprovable. All performance excluded. Initial suspicion of excessive reasoning alone was incomplete: retained evidence shows roughly9900visible chunks and317reasoning chunks for first request. Do not pool with shorter workloads. Use one16word/2048token scout before expanding, reusing established GLM APC workload.

  • /tmp tmpfs used30GiB including11 inactive Pi build directories fromSep20. No cwd/exe/fd handles found for those roots and no relative symlinks escaped them. Preserve those identified directories under private archived temporary-artifact store and leave original path symlinks; no deletion of contents or other temporary files. Preview and incremental receipt retained. This is storage preparation, not model performance evidence. Future substantial build staging belongs on the approved large-artifact store.

  • First MiMo startup passed containment but SGLang v0.5.20 rejected ambiguous legacy --cuda-graph-max-bs. Closed by retained versioned recipe v2 using --cuda-graph-max-bs-decode 1; runtime parsed it and reached model initialization. No performance evidence from this attempt.

  • Corrected MiMo attempt hit CUDA OOM constructing AudioProjection before weights loaded or KV allocated. Each GPU reported 93.84GiB PyTorch allocation, with a128MiB request against126.75MiB free. Failure is runtime/configuration-specific; it does not prove all MiMo recipes infeasible. Preserve log and status, unload candidate, restore exact GLM while researching a documented remedy.
  • Post-GLM-unload host admission was initially short of52GiB+12GiB reserve. Inactive temporary artifacts were preserved on disk with original-path symlinks. Forty roots verified. Read-only test/cache directories produced partial cross-filesystem move refusals; remaining regular files were hash-compared with archive copies and retained separately before restoring the original root path via symlink. Complete final mapping: tmp-preservation-final.json. No global permissions, unrelated services or model limits changed. Future large build staging should use /data.