Cost is a requirement, not a guess. One records what a workload is expected and allowed to consume, binds it to exact evidence, and refuses to invent precision it does not have.
Bind cost to exact evidence
A cost estimate cites the provider price snapshot and its observation time, together with the workload assumptions behind it. Planning records fixed, marginal, expected, and worst-case consumption so a later reader can see which number each conclusion used.
When apply refreshes an observation and the refreshed estimate exceeds a bound recorded in the reviewed plan, apply stops for approval or requires a new plan. A price, capacity, or quota change is a material input; it is never silently absorbed into an existing immutable plan.
Budgets compose over the resource graph
Every budget attaches to a level of the resource graph and inherits downward:
Fleet -> Tenant -> Deployment -> Runtime -> Component -> Scope -> OperationA child budget cannot widen its parent. The dimensions a budget may constrain include CPU time and share, wall deadline, memory, allocation rate, buffer bytes, task and thread counts, handles, storage and IOPS, network, observation cardinality, external calls, and monetary cost.
A provider profile that proves one stateless process does not thereby prove a multi-instance stateful host or a tenant-wide quota; capacity, isolation, persistence, and reconciliation remain separate claims. See Resources and lifecycle and Multi-tenant platforms.
Cost is a selection axis, not the first one
After hard feasibility and assurance, Planning applies default lexicographic objectives. Required availability and latency come before operational and monetary cost; artifact size, cold start, and runtime services come after. Cost cannot override a hard semantic, security, authority, compatibility, or recovery constraint.
Selection and cost evidence are recorded in the immutable plan. Follow Plan, build, and apply and Deployment for how a reviewed plan is applied and how quotas, preemption, scaling, and exhaustion are enforced as typed outcomes.
Author quantities with typed units
Authored quantities use one compact, dimensioned form. Fixed elapsed durations
and calendar periods remain distinct, so a Month or Year is not silently
treated as a fixed number of seconds.
monthly_cost: <= 5_000USD
availability: >= 99.95%
throughput: 100MiB/s
payload_limit: 16MiB
memory: 2GiBThe quantity forms above are illustrative end-product authoring. Canonical normalization may change presentation units, but it cannot silently truncate or reinterpret a dimension.
Agent and external-effect budgets
Bounded autonomous work carries explicit budgets for tokens, cost, elapsed time, context, tool calls, child work, network use, and external effects. Reaching an external spend limit is a declared resource-exhaustion outcome, not a retry, a domain fault, or permission to widen the budget. See Workflows and agents.
Benchmark budgets
AbsoluteBudget and RegressionBudget in
Benchmark and performance evidence bind measured
behavior to typed budget outcomes. A benchmark budget is evidence only for its
exact workload, environment, and harness; it never substitutes for a resource
or monetary constraint.
Inspect cost without a universal score
One never invents precision or a universal score. Missing price, capacity, or observation evidence remains an explicit missing-evidence disposition rather than a made-up number. One Cloud presents plans and costs from the same owner records without becoming a second authority for them.
Normative owner & evidence
The canonical resource and quota owner is Resources. Cost selection and the lexicographic objectives are owned by Planning, with quota enforcement in Deployment and observation limits in Observability.
Demonstrated today: the resource lifecycle and quota contracts, Planning's cost and selection records, and provider price and observation providers. Illustrative: the typed-quantity block above and any specific monetary figure or budget bound not recorded by an owning plan.