Troubleshooting and FAQ

Common failures

• Schema/config error: run the command with --dry-run --json; fix the reported stable code and path. Unknown keys and incompatible majors fail closed.
• Missing group/provenance: provide a split group plus source, revision, license, and rights. Do not invent approvals.
• Output exists: inspect it, then use resume or an explicit overwrite flag where supported. Never delete a shared artifact blindly.
• Incomplete checkpoint: inspect classification. Resume only atomic complete checkpoints; use weights-only artifacts as labeled warm starts.
• OOM: run the one-step probe, shorten sequences, reduce microbatch, increase accumulation, enable checkpointing, choose LoRA, or use suitable hardware. MoE active parameters are not optimizer memory.
• Provider failure or budget stop: inspect redacted status and ledger. Successful paid requests must not be repeated on resume.
• Artifact tamper: stop; do not reload/export. Restore from a verified source or rerun from frozen inputs.
• Unavailable recipe: read the machine gate reasons. Do not bypass missing license, remote-code review, hardware, reload, or evaluation evidence.

FAQ

Is the default workflow offline? Yes, only the explicitly documented deterministic fixtures. Provider, download, upload, GPU, and remote-code paths are separate opt-ins.

Are model benchmarks comparable? Only with identical dataset/evaluator revisions, prompts, pooling, dimensions, task sets, and contamination policy.

Can I publish the private-alpha artifacts? No. Publication requires explicit authority and later acceptance gates.

Is RunPod supported? The Phase 20–23 CLI and offline/read-only/dry-run contracts exist. Live Pod or volume mutation is unavailable pending qualification; production recipes, spot, generic exec/log REST, direct Secrets, Serverless, and fleet are also unavailable. No live evidence or spend was produced. See RunPod execution and support status.

Next: migration and release or the support policy.