Model specification: Fixed-sample complete validation

Conditionally on frozen training and calibration information:

1. every task, target, decoder, score, and positive shot count $n_{k,\nu }$ is fixed before validation;

2. scores $Z_{k,\nu ,i}\in [-1,1]$ are independent within each task-setting pair and have a common mean $m_{k,\nu }$;

3. all settings for a task test the same complete channel; references are untouched by the decoder, and input preparation and ancillary independence are justified;

4. every attempted trial, including losses and failures, receives its predeclared score;

5. on a joint event $G$, with $\mathrm{Pr}(G^c)\le \beta$,

   $$|m_{k,\nu }-c_{k,\nu }|\le b_{k,\nu },d_{\diamond }({\Gamma }_{\star ,k},{\Gamma }_{0,k})\le {b_k}^{\mathrm{tr}};$$

6. the complete-output access, padding, quantum-reference preparation, and Pauli measurement implementation are admitted and calibrated.

This is additional validation access, not the original classical receiver access ${\mathsf{A}}_0$. No unaccounted adaptive stopping, setting removal, or channel drift is covered.