This mock example exercises base 1.0 planning and paired collection. It intentionally stays statistically inconclusive: its single deterministic cluster does not establish the independence or precision needed for a release claim. It is not real-agent quality evidence.
Run the base workflow in the controlled comparison guide. Do not overwrite an existing policy, adopt historical outcomes, or resume an interrupted collection.
This fixture is intentionally not a native-Copilot example: its eval selects
the mock executor. For native collection, use your own fixed-model Copilot SDK
evals with no skills, deny-all client tools and supported text/string checks,
then explicitly pass --collector native-copilot to both planning and
collection. The native profile can incur provider usage and is not server
isolation, runtime-version evidence or authenticated assurance. See the
native profile contract.
Do not run native collection to reproduce this offline fixture; it is
paid-capable and this example contains no supported live-source test suite.
The bundled example is not a ready-to-pass assurance fixture: it has no authored reference corpus or review declaration. Examples are unreviewed by default. Test-only reviews elsewhere are explicitly synthetic, not authenticated human review or production review evidence.
To author a separate qualified example, require assurance in the base policy,
declare every native-text endpoint through task requirement/check mappings, and
provide exact-bound per-arm evaluator-only reference labels and finite authored
output envelopes using the existing challenge contracts.
Use compare-assurance-plan --release-policy ... --baseline-eval ... --candidate-eval ... --baseline-references ... --candidate-references ... --output ... to create a new independent contract. Collection and assessment
must explicitly select --assurance-contract plus --release-policy, with
current per-arm sources; assessment also needs --collection-dir.
Review files and per-arm --accept-review-source are separate current inputs.
They do not authenticate a reviewer. Never change an example's unreviewed state
merely to obtain a pass. Fresh exact offline1.0 reports for both arms,
deterministic regrade of retained actual responses, and all base
non-assurance pass conditions are necessary. This example's statistical
limitations cannot be erased by grader qualification.
Only mock/no-skills/native-text/finite-authored-output execution is supported. No live SDK assurance, calibration, paid1.1/billing support or automatic paid validation is selected. Required-assurance base decisions, the API and dashboard remain nonpassing; only explicitly selected outer CLI assessment can pass. See the profile lifecycle, schemas and limitations.
This example uses the public offline mock producer. Native durable collection supports Linux/macOS only; Windows supplied-reader checks do not certify durable writing.