evaluation-operations
Evaluation Operations
Maintain factual operational control of an evaluation campaign: preserve what happened, keep already approved work moving when requested, and report only what current evidence establishes. This skill supplies the working method. The project supplies its runner, scheduler, manifests, registries, and deterministic checks; do not replace them with a new campaign framework, controller, daemon, schema, or parallel spreadsheet.
Select the authority mode
Use the mode the request actually grants, and keep it throughout the work unless the user changes it.
- Status or diagnosis is read-only. Inspect sources and report the state; do not rewrite source records, launch work, restart work, rescore, or rerun.
- Resume or execute permits only work already approved for its models or targets, data, configuration, budget, resource limits, and other stated bounds. It can use the project's canonical durable records as the operation requires.
- Monitoring is observational unless the request also grants execution or a specific change. Waiting for completion does not authorize dispatch, retries, rescoring, cancellation, or record repair. Define the terminal condition and use a supported sustained-observation route as described below.
Reuse an existing session grant; do not ask for it again. An authorized runtime model or API call needed by the approved campaign is part of that execution and does not need repeat approval. A status request, idle capacity, a prepared configuration, a source location, or a queued-looking record does not authorize execution. Keep external publication, a new costly experiment, a model or API call outside the approved campaign or its limits, scheduler installation, destructive cleanup, and access to an unapproved host in separate scope.
Keep private inputs, results, and source maps outside this package and within the approved sharing scope. Logs, errors, and commands can contain prompts, responses, personal identifiers, local paths, and signed URLs; share only the necessary redacted diagnostics and retain raw evidence in approved storage under its retention policy. Credentials stay in the configured secret mechanism, never in reports or worker packets. Use only approved destinations and providers; do not transmit material to a new provider merely to obtain cheaper routing.
Bind the campaign to its actual records
Before drawing a conclusion or taking an action, find the available project-defined records or explicit coordination that establish campaign scope, priorities, durable attempts and results, raw outputs, live runtime state, and applicable budget, resource, ownership, and validation rules. A project may use a manifest or queue, but a simple approved campaign may instead use a verified single owner and a small task record. Read project binding when those sources are not already clear, disagree, or must be reconstructed.