palarran-ml-hypotheses

原文涵盖的主题

• ML hypotheses
• Explicit runtime pipeline
• Documentation and context first
• Stage work
• Artifacts, decisions and verification

Shared cycle: before work and after checks, follow global memory: bounded recall → applicability → sanitized verified learning/dedup. This memory scope is explicitly authorized; other profile restrictions remain. --plan has no networking/publication; outages do not block the primary task. When assigned MLPIPELINETASK, read the pipeline and typed contract. Perform only the assigned planning stage: train/benchmark use --plan here, data is research. Pass cumulative E/H/ADR/unknowns and exact lineage/inputs without changing the verified prefix. Neither agent handoff nor cost reports authorize compute: the runtime launches fit/evaluate after private approval. Do not launch nested agents/provider commands or edit shared progress; the coordinator owns it. Standalone invocation does not automatically create a pipeline. Invoke /palarran:ml:hypotheses {task/plan path} or $palarran-ml-hypotheses. Perform the assigned stage only; do not launch other stages automatically. Read the route, workflow, artifact contract, cases and transfer limits, dated sources, then hypotheses.md. Explain relevant Karpathy/researcher case applicability and transfer limits. Missing mandatory docs are blockers: find

这里只展示节选，请以完整原文和当前要求为准。