Single-machine CPU audit suite (reproducible, seeded) that separates world-model perceptual fidelity (next-frame MSE / SSIM / NLL; fixed random-conv filter banks as stand-ins for learned metrics) from control utility (whether a planner can solve the task). Reduced pilot on one env (pixel_nav, 3 seeds × 3 capacities, 12k steps): H1 threshold met — pixel fidelity weakly tracked closed-loop return at |ρ|=0.32<0.4, but this is a weak, tentative exploratory signal only (n=9; 95% CI crosses zero, Holm-adj p=1.0 → not statistically significant, not an established finding); control-relevant diagnostics did not (H2 not supported, ρ=0.26<0.6); INT8 compression did not degrade calibration faster than raw fidelity (H3 not supported). The locked 200k-step × 5-seed × 2-env main run has not been run; this is the preregistered reduced pilot only.
Code: github.com/dafahaha/wm-audit

Solo design & implementation. Open-source tool that verifies whether an AI API relay actually serves the advertised model, by probing statistical behavior (random-number, token, latency distributions) against baselines with chi-square / KS tests and Bayesian updating. A guided offline demo accompanies the tool (run the real audit locally with transit-truth serve); a technical report validates same-family discrimination: gpt-4o vs gpt-4o-mini, TVD up to 0.90, 91.3% in-distribution resubstitution posterior (not held-out).
Code & demo: github.com/dafahaha/transit-truth · live demo

Soft Actor-Critic implemented from scratch in PyTorch (no Stable-Baselines3), run as a pre-registered two-arm controlled study with a logged third-arm (FixedAlpha) amendment on Hopper-v4 (CPU, 120k env steps, 3 seeds/arm): Baseline (reward scale 1.0) vs RewardScaled (0.1) vs FixedAlpha (scale=0.1, α fixed at 0.2; automatic temperature off). Mean return ± SD (final window 105k–120k, n=3/arm): Baseline 1085±698, RewardScaled 1118±623, FixedAlpha 1404±632. Welch p: Baseline–FixedAlpha 0.59, Baseline–RewardScaled 0.95, FixedAlpha–RewardScaled 0.61; one-way ANOVA F=0.22, p=0.81 — no statistically distinguishable effect detected (n=3; the t-intervals overlap heavily). At 60k the reward-scaled arm looked ~3× better, but extending to the pre-registered 120k budget and adding the fixed-temperature control overturned that apparent gain as early seed noise: an honest pre-registered null. Raw CSVs committed, curves reproduced by script.
Code: github.com/dafahaha/embodied-offpolicy-study

Config-driven CLI: export Stable-Baselines3 policies to ONNX/TorchScript, quantize to FP16/INT8, and benchmark latency, throughput, and accuracy across x86 GPU and CPU via ONNX Runtime, with auto-generated HTML reports. An experimental standalone TensorRT engine-builder (FP16/INT8) is included but not yet wired into the benchmark harness (GPU-tested numbers tracked for v2.0). Jetson/edge is a target deployment path, not yet measured on device. A reproducible harness for the accuracy–latency tradeoff of quantized RL policies.