Freeze scored fixtures and pass the six-execution non-scored pilot

[ohad6k]

The scored benchmark cannot be the place where we discover that fixture reset, blinding, or redaction is broken.

Freeze the private Ditto Proof v1 task pack, then pass a separate six-execution, non-scored pilot before any paid or scored run.

Acceptance criteria

• Create primary and held-out variants for work/done, design, and writing.
• Freeze every task as a clean Git commit plus a deterministic content hash before the pilot.
• Keep scored fixture content and the private operator rubric outside Git until the final evidence package is approved.
• Use separate pilot-work, pilot-design, and pilot-write fixtures that can never enter v1 aggregates.
• Run each pilot fixture once in cold and +Ditto conditions: six isolated executions total.
• Capture every required field without manually patching records.
• Prove deterministic reset, clean condition separation, opaque review IDs, artifact hashing, and append-only storage.
• Catch seeded secret, username, local-path, profile-text, and receipt canaries.
• Produce a sanitized pilot package labelled non-scored and non-comparable.
• Record the observed provider cost and quota, then get explicit approval before any later scored execution.

Failure policy

A failed pilot is diagnosed and rerun from clean pilot fixtures after the harness is fixed. Pilot outputs never become public performance evidence.