Break the retry storm
Keep the work

Compare a bounded retry policy with a coordinated candidate. Verify the same result, recovery and tenant isolation before accepting fewer requests

Run a controlled incident

Synthetic faults · real queue · durable receipts

Candidate

Ready for a synthetic comparison

ContractAwaiting run

Outputs, effects, scope and deadlines

Baseline operations—

Requests plus receipt lookups

Candidate operations—

Recovery lookups included

Evidence by scenario

Scenario	Baseline	Candidate	Independent check
Run a comparison to inspect the evidence

What these numbers mean

Operation counts describe executed work, not measured energy or CO₂. This workspace has not measured physical energy. Carbon claims require a declared SCI boundary, energy methodology, hardware allocation and the overhead of AI and CI

Faults are injected into the application fixture. They are not claims of bugs in BullMQ or bull-board. GitLab Duo execution remains subject to a separate platform validation

NET BENEFIT, SAME USEFUL WORK

Environmental assessment

Awaiting comparison

Energy, hardware allocation and optimization overhead are assessed separately from operation counts

Boundary and evidence

One functional unit is a verified completed document. Every failed and unresolved case still contributes to the batch cost. The model includes runtime support, reserved idle capacity, allocated hardware emissions and the AI/CI cost of optimization

Intervals disclose uncertainty. Illustrative values never establish a real saving; supplied evidence must be reviewed independently. A faster finish does not eliminate idle energy when the host stays reserved

SAME WORK, EQUAL RESERVATION WINDOWS

Recorded power estimates

Loading local evidence

Review completed local measurements without starting a new workload

What the estimates establish

Every comparison requires the complete independently verified workload, six balanced orders, equal reservation windows and raw files matching the current source. Reviewing evidence does not start a sampler or change a policy

Intervals cover timestamp resolution and display rounding. Instrument accuracy and background activity are not quantified. CPU and combined estimates overlap and must not be added

These values do not establish whole-host electricity, carbon savings or a complete SCI result. Missing runtime components, hardware allocation and AI/CI overhead remain separate requirements. File hashes bind bytes; they do not authenticate the measurement source

CORRECTNESS, TIME TARGET, ENERGY

Retry policy selection

Loading local evidence

Compare complete evidence before choosing a policy. Missing or overlapping evidence prevents selection

How the decision works

The selector verifies source, results, time targets and environmental evidence. It compares energy per verified document, including amortized AI/CI overhead. A winner requires uncertainty intervals that do not overlap

Selection is conditional on supplied evidence and limited to evaluated eligible policies. It does not apply a policy, certify physical energy or deploy anything