PRELIMINARY EXPERIMENTAL PROTOCOL — MENS-SCCA
                    × APEX
                              Protocol skeleton for preregistration and confirmatory execution


1. RESEARCH QUESTION
Determine whether longitudinal trajectory instrumentation adds reproducible information beyond final-outcome evaluation.

2. UNIT OF EXPERIMENT
Agent × Task × Run × Condition.

3. DOMAINS
When authorized: Investment Banking, Management Consulting, and Corporate Law.

4. VARIABLES
Outcome variables: final score, pass/fail criteria, expert acceptance, compliance with constraints, and externally verifiable state
changes.
Trajectory variables: observable actions, tools, transformations, preservation, errors, propagation, recovery, consistency, and
transfer.
Controls: task, agent/model, condition, environment, tools, initial information, constraints, rubric version, and evaluator
versions.

5. CONDITIONS
BASELINE, PERTURBATION, RECOVERY, and TRANSFER, as defined in the main proposal.

6. MEASUREMENT
Exploratory constructs: outcome, trajectory integrity, transformation fidelity, failure propagation, recovery, cross-run
consistency, and transfer loss/preservation.
No composite index is assumed valid before stability, sensitivity, redundancy, and reproducibility are established.

7. EVALUATION
Deterministic verification; expert rubric; model-based judge. Disagreements are retained and analyzed.

8. PARTITIONS
DEV for development; PILOT for calibration; HELDOUT for confirmatory evaluation. The final sample size will depend on
authorized task availability and required precision.

9. ANALYSIS
The confirmatory statistical plan will be fixed before confirmatory analysis. Planned analyses include condition differences,
trajectory-outcome association, inter-run stability, expert/judge agreement, sensitivity, redundancy, failure modes, recovery,
and transfer.
A single average will not be used to conceal critical failure categories.

10. CONTAMINATION AND PROVENANCE
Before confirmation, define development tasks, held-out tasks, training exposure, public artifacts, confidential data, and version
controls. Mercor data access remains authorization-dependent.

11. REPRODUCIBILITY RECORD
Each execution should, where authorized, be associated with task_id, run_id, agent/model, condition, environment/version,
tool configuration, rubric version, judge version, trajectory record, final artifact, and evaluation result.

12. SUCCESS AND NON-SUCCESS CRITERIA
Favorable evidence requires stability, relevant discrimination, incremental information beyond outcome, reproducible relation to
expert criteria, and robustness to sensitivity analysis.
Insufficient evidence includes redundancy, replication failure, excessive judge dependence, excessive instrumentation
sensitivity, or no improvement in failure identification/explanation.

13. METHODOLOGICAL ORDER
Question → hypotheses → specification → governance → preregistration → calibration → instrument validation → pilot →
execution → confirmatory analysis → robustness → interpretation.