Verifiedon 2026.7.1
Action boundary
Before you act
- Expected result
- A policy routes only eligible work, records its budget basis, exposes a quota state, and names the person who can override it.
- Failure mode
- A cheap default silently receives sensitive, high-stakes, or untested work, or a budget limit fails open with no audit trail.
- Rollback
- Disable the new route and restore the last approved policy while preserving the ledger and breach record.
Routing is a policy decision
A route should answer, in plain language: which work qualifies; what must never take this route; the quality and service thresholds; what budget applies; what happens at a limit; and who can make an exception. “Use local first” is not a policy. It lacks a data boundary, a quality gate, and an owner.
Build the decision ledger
Use observed measurements where you have them and label estimates as estimates. Keep cost/resource calculations legible: quantity, unit basis, time window, assumptions, and source of each number. For local capacity, include the operational costs you chose to count and state what is excluded. A spreadsheet can be honest; a single blended “savings” number often is not.
| Route | Eligible fixture class | Quality gate | Service guard | Budget/quota | On breach |
|---|---|---|---|---|---|
| Candidate A | Named low-risk class | Score threshold | Tail-time threshold | Owner-set limit | Pause and notify owner |
| Candidate B | Named review-required class | Reviewer approval | Queue limit | Owner-set limit | Escalate to human review |
Lab: write a budget scenario
Take a fixed monthly ceiling and a conservative volume assumption. Show the arithmetic in your ledger, then test three cases: normal use, a volume spike, and a failing preferred route. For each, decide whether the system pauses, degrades to a smaller safe task, requires approval, or rejects work. A quota that cannot explain its failure mode is just a surprise waiting for a Friday.
Checkpoint
A provider invoice differs from a planning estimate. What should the team change first?
The ledger’s assumptions and the decision policy, after checking the measurement basis. Do not retrospectively call the old estimate a fact. Preserve the variance; it is evidence for the next review.
Learner artifact — policy gate: Draw input classification → data-policy check → quality eligibility → capacity/service check → budget check → selected route or human escalation. Add a visible audit event after every branch, including rejected work.
Boundary note
No OpenClaw-specific configuration keys are supplied here. Apply a policy only after the implementation path has been verified against the current primary documentation and tested in a non-production environment.
Source provenanceVerification and sources
Review receipt rr_local_models_routing_cost_fixture
- Outcome
- approved
- Method
- source-review
- Reviewer
- academy-specification-review
- Reviewed
Evidence
- academy-spec — curriculum-local-models-contract; snapshot
1c60fa76ac90…
Limitations
- Approval covers the original workload, benchmark, routing, quota, and fallback method for 2026.7.1; it makes no universal provider, model, GPU, performance, privacy, or savings claim.
Lesson checkpoint
Ready to move on?
Mark this lesson complete when you can apply its outcome without relying on the examples above.