ML Production

Use this contract for the concrete task. Apply a short relevant check for a small change; expand investigation when the failure, risk or requested artifact warrants it. Resolve facts from the repository and runtime before imposing a process. Delegation is optional and should answer a bounded unresolved question.

Release and operation contract

Identify the artifact, deployment target, traffic/resource budget, quality acceptance criteria and rollback owner. Dataset release and evaluation-set construction can precede training; serving needs a runnable model artifact.

• Bind model, preprocessing/tokenizer, data/prompt versions, dependencies and configuration to a release identity. Detect train/serve skew with representative inputs.
• Validate model output and failure behavior on required slices, not only endpoint health. Canary/shadow decisions need quality and resource evidence plus a reversible rollout.
• Measure queueing, batch behavior, memory/KV-cache pressure, warmup and tail latency on the actual hardware/traffic shape before choosing scaling or quantization.
• Test quantized/compressed artifacts against the task metric and supported kernels; smaller weights do not guarantee faster serving or acceptable quality.
• Define observability for input/output quality, drift, errors, tool calls and cost as applicable. Protect sensitive logs and distinguish model failure from serving/infrastructure failure.
• Preserve dataset composition, provenance, labels, deduplication/contamination checks, slice coverage and release documentation. Schema validity alone does not establish content quality.
• Exercise rollback/restore and record artifact compatibility. Do not equate deployment, health checks and live quality acceptance.

Evidence and output

Produce a release/serving decision record, dataset release manifest, benchmark, incident diagnosis or rollout evidence. Include the artifact/config identity, measured workload, quality/resource results, unresolved limits and recovery path. Use existing backend/devops integration rather than copying an unrelated service scaffold.