I study how agents fail, where decision models help, and how to build dependable 3D systems. These experiments connect measured results to the larger problems I’m working on.
Comparing ways to challenge coding agents: finite rules, competing evidence, scope boundaries, and archive-specific inference.
Finite rule and convention tasks stayed at 2/2 solved in the selected local probes. Scope-conflict and archive-inference revisions reduced observed solves to 0/5 and 3/10.
Comparing general LLMs and a decision-model hybrid across retrieval, tool routing, arithmetic, classification, and other tasks.
In a 60-run synthetic benchmark, Decisions + Astra completed 20/20 tasks at 22% lower estimated cost than Astra. Gains concentrated in bounded choices; many file-based tasks still needed Astra.
Comparing an original boundary representation with an independent geometry engine.
54 OCCT comparisons passed. Maximum volume difference: 2.73 × 10⁻¹² mm³ across the tested cases.
Measuring how well a trained splat reproduces synthetic views of an apartment mesh.
After scene-specific cleanup: 34.60 dB PSNR and 0.9718 SSIM on the same 12 synthetic evaluation views.
Fine-tuning YOLOv8s and connecting piece detections to a structured board state.
61 logged epochs. Final validation mAP@50-95: 82.16%; no physical-gameplay claim.