# Model decision worksheet examples

These synthetic examples show a pending decision and an unknown serving route.

Every identifier, role, and setup detail is invented. No model was called, no result was measured, and no run receipt exists. Keep the examples separate from records of your own runs.

## Synthetic example: a decision stays pending

Workload: classify fictional support messages and preserve the supplied ticket ID.
Acceptance rule: copy the ID exactly, use only facts stated in the message, and leave absent facts unknown.
Known failure cases: an invented category and a paraphrased requested action.

Model display name: Example model A, invented for this worksheet.
Exact API model ID: `example-model-a`, an invented identifier.
Serving route or host: Example host A, a fictional route.
Region: Unknown.
Prompt and configuration version: `example-prompt-v1`. Parameters not set for a run.
Fixture, schema, and checker versions: `example-fixture-v1`, `example-schema-v1`, and `example-checker-v1`. All invented labels.
Unknown setup fields: supported parameters and region.

Run date and receipt reference: Not measured. No run date or receipt exists.
Attempted, completed, and accepted counts: Not measured.
Cost basis: Unavailable. No usage, rate source, or invoice supplied.
Price source, retrieval date, and applicable route: Unknown.
Latency boundary and sample count: Not measured.
Source-documented facts: None supplied.
Observed outputs and check results: Not measured.
Receipt provenance limits: this invented setup has no run receipt.
Interpretation and limits: the acceptance rule is written, but there is no result to support a production choice.

Provider-published replacement: Unknown. No notice supplied.
Approved choice: Pending.
Decision and supporting cases: investigate the two listed failure types before selecting a configuration.
Next action owner: fixture reviewer, a fictional role.
Reopen trigger: a reviewed fixture and a dated receipt become available.
Next review date: Not chosen.

## Synthetic example: the serving route is unknown

Workload: draft a support-reply suggestion for human review, using fictional messages.
Acceptance rule: preserve the stated issue and proposed action without adding account facts.
Known failure cases: invented account status and an unsupported resolution.

Model display name: Example model B, invented for this worksheet.
Exact API model ID: `example-model-b`, an invented identifier.
Serving route or host: Unknown. The model name alone does not identify the host.
Region: Unknown.
Prompt and configuration version: `example-prompt-v2`. Parameters not recorded.
Fixture, schema, and checker versions: Not recorded.
Unknown setup fields: route, region, parameters, and fixture versions.

Run date and receipt reference: unavailable because no evaluation was performed.
Attempted, completed, and accepted counts: Not measured.
Cost basis: Unavailable. No rate applies until its route and source are established.
Price source, retrieval date, and applicable route: Unknown.
Latency boundary and sample count: Not measured.
Source-documented facts: None supplied.
Observed outputs and check results: Not measured.
Receipt provenance limits: no attempts or output logs are available.
Interpretation and limits: neither a model identity nor a published retirement notice can be matched to a serving route yet.

Provider-published replacement: Unknown. No notice supplied.
Approved choice: Pending.
Decision and supporting cases: identify the route before interpreting a provider notice or planning a comparison.
Next action owner: application maintainer, a fictional role.
Reopen trigger: the exact serving route is recorded with a dated source.
Next review date: Not chosen.