Run the consequence path.
Choose a worked application, inspect every stage, change the control inputs, and export an unsigned research receipt. This preserved evaluator uses the website's earlier 28-gauge panel; it is not the locked R1 implementation, makes no network call, and cannot authorize a real action.
Choose a worked case. Replace every assumption.
Each case is prefilled, editable, deterministic, and tested to an expected gate. It contains no real person or confidential record.
Autonomous AI / tool action
Every stage reports complete, warning, binding, or unknown. Click a stage to inspect or change its inputs.
A later model suggestion cannot lower this floor. No average creates a global green. Unknown remains unknown.
PRESERVED RESEARCH EVALUATOR · NOT CURRENT R1 · Not deployment authorization. All evaluation runs locally in memory. Resetting or leaving the page destroys the case. Nothing is stored or transmitted by Project Shadow.