Deployment planner

Do the release review in the same order production will fail.

Open-weight deployments need an evidence trail: exact weights, tokenizer, prompt format, runtime, quantization, eval set, abuse tests, latency budget, cost model, and rollback owner.

Weight intakeRecord model ID, checksum, license, tokenizer, context window, and quantized format.

Runtime matchProve the model loads under the target stack, then test streaming, tool calls, structured outputs, and batching.

Evaluation gateRun task-specific evals plus refusal, jailbreak, sensitive data, and domain-risk checks.

Operations gateMeasure latency, memory pressure, throughput, cost, logging, monitoring, and rollback behavior.