During a second architecture review for a risk analytics team, the team rejects an earlier proposal to use ECS services with one task definition per job type and manual scaling. A security exception is about to expire, and the replacement must be enforceable by policy rather than by wiki instructions. The existing environment still has a situation where tens of thousands of independent Monte Carlo jobs run nightly, can use Spot, and need queue priorities and retry policies, but the new decision record must maximize compute efficiency without building a scheduler. Jobs use container images and write results to S3. The solution also must leave clear audit evidence for why the tempting shortcut was not selected. Which recommendation best addresses the risk in the earlier proposal?
