Tune model-right-sizer's wording knobs (`eval/tuning/knobs.py`) against a REAL, already-measured build's actual token spend — not the synthetic benchmark `model-right-sizer-prompt-tuning` searches, nor a fresh build per candidate. Picks a task from `overfitting_guard.py`'s `HOLDOUT_TASKS` registry (real actual/budgeted pairs already recorded), dispatches 3 INDEPENDENT BLIND dry-runs per candidate (no calibration-ledger access), averages the budget across draws (single-draw noise can flip within/over-budget calls), maps to real actuals, scores via `classify_budget_adherence` + `score_candidate`, diagnoses the miss pattern, proposes ONE wording change, and re-runs to check improvement. n stays fixed per task — flags rather than silently continues once squeezing looks like overfitting. Use when someone says "tune the knobs against this blueprint/build", "iterate the dry run with no prior context against the real actuals", "keep tuning until N%", or "how close does a blind estimate get to the actual cost".