Evidence and Analysis for "Do Retrieved Errors Improve Decisions? A Prospective Audit of Memory-Corrected Visual Planning"

Saved-score data and NumPy analysis accompanying "Do Retrieved Errors Improve Decisions? A Prospective Audit of Memory-Corrected Visual Planning" by Daniel Lobato Garcia. The archive contains 56 complete candidate pools (8 development and 24 evaluation tasks for the memory-source study, plus 24 fresh tasks for the component audit), each with 690 candidates, and the 128-row collection check. It includes predicted and executed feature costs, terminal positions, retrieval records, frozen protocols, analysis code, file checksums, and a data schema. With Python 3.10-3.12 and NumPy 1.26.4, run: python -B reproduce.py. This recomputes both studies' original summaries, development calibration, primary task-bootstrap intervals, support diagnostics, and labeled post-hoc controls. Scope: recomputation from saved scores. The bundle does not regenerate neural predictions, nearest-neighbor searches, candidate searches, or simulator trajectories. No images, feature arrays, model weights, original dataset, or full action trajectories are included. This is not independent scientific replication. Licenses: MIT for Python analysis code; CC BY 4.0 for data and documentation. The archive identifies the file-level license scopes and retains the upstream MIT notice. The author works at Meta on Instagram AI Search. This research is not affiliated with Meta and does not represent Meta's views. OpenAI Codex assisted with implementation, checks, analysis, visualization, literature lookup and manuscript drafting.