Explore how policy updates reach active workflows, what a controller can enforce, and what evidence remains afterward. Clear questions. Testable outcomes.
Proposed research, open to evaluation.The emphasis is on observable behavior at defined control boundaries—and an accurate account of what remains outside them.
Questions to test, with the assumptions left visible.
How should updates affect retries, parallel branches, and handoffs? Define the authoritative policy version at each checkpoint, then test delayed updates and decisions made near a policy change.
Propagation delay · stale decisions · checkpoint behavior
External work may continue after permission changes. Test how the controller distinguishes a requested cancellation, a confirmed stop, completed side effects, and an outcome it cannot yet establish.
Status accuracy · reconciliation · repeat-operation risk
Connect the policy, selected action, and cost information available at the time. Ask whether an independent reviewer can reconstruct a decision, including a missing acknowledgement or a late usage report.
Evidence coverage · missing records · overhead
Could independent providers compete for eligible tasks on price, availability, and verifiable execution requirements? Study strategic bids, observable quality, non-completion, and cancellation terms. Incentive guarantees would depend on explicit assumptions.
Mechanism design · supplier incentives · settlement
Start with controlled workloads and explicit policy changes. Compare an initial-authorization baseline with checks at the proposed boundaries, using the same workload and timing of injected events.
Specify the controlled actions, policy authority, worker capabilities, and what counts as an observed outcome.
Repeat the same policy change before dispatch, during an external attempt, and before result release.
Compare action traces, unresolved events, reported usage, and the evidence a reviewer can reconstruct.
Report timing and failure assumptions, workload variation, and unsupported actions alongside the results.
Success criteria belong to the agreed workflow. These proposed measurements make the trade-offs and unresolved cases explicit.
| Change to introduce | Behavior to inspect | Measure or report |
|---|---|---|
| Revoke a destination | Check the next dispatch and transfer, including parallel work that started earlier. | Policy-application delay; decisions using an older version; prior transfers. |
| Reduce the budget | Include pending external work when deciding whether another attempt is eligible. | Settled and reserved exposure; late charges; blocked or approved attempts. |
| Lose a provider response | Preserve an unknown outcome until evidence supports a retry or reconciliation. | Unresolved duration; duplicate-operation risk; status accuracy. |
| Change a release rule | Evaluate a completed result before forwarding it through the controller. | Release decisions and latency; missing evidence; completed external side effects. |
Compare like with like. Hold workload volume, quality requirements, and execution endpoints constant where possible. Include retry, failure, and evidence-recording overhead.
Report uncertainty. A missing provider acknowledgement is an unknown outcome. An estimated charge is not a settled cost. The evaluation should retain those distinctions.
This agenda describes proposed questions. It does not claim published results, verified novelty, universal enforcement guarantees, or measured customer outcomes.
A proposed product design.
Capabilities to define and evaluate for an agreed workflow.
An open question with no committed outcome or availability.