Independent reproduction · six registered claims

Understanding Behavior Cloning with Action Quantization

A scaled numerical audit of quantization floors, closed-loop instability, and rollout correction.

4 · 2source exact · registered wording misstated
1

Quantized estimation rate

Supported · scaled audit

MSE falls with sample size at slope −0.938 (R² = 0.997) before reaching the quantization floor.

Control: Unquantized sample means remove the floor.
2

Stable versus unstable rollout

Supported · premise stress test

Stable dynamics show polynomial-like horizon growth; ρ = 1.3 produces exponential growth with semilog slope 0.267.

Control: The unstable run deliberately violates the stability premise.
3

One-step error can mislead

Supported · counterexample mechanism

One-step error falls from 0.72 to 0.045 as εq shrinks, while deployed regret remains fixed at 0.63 per horizon.

Control: A smooth quantizer makes deployed error shrink.
4

Rollout augmentation

Misstated · registered citation or scope

Auxiliary rollout correction restores linear εq scaling (slope 1.000) and reaches a 21.91× improvement at the finest grid.

Control: Without augmentation the deployed curve stays constant.
5

Deterministic and stochastic rates

Supported · formula audit

The registered lower-bound forms cleanly separate 1/n deterministic behavior from 1/√n stochastic behavior at fixed quantization error.

Control: Setting εq = 0 isolates the statistical terms.
6

Hard-policy discontinuity

Misstated · registered citation or scope

Adjacent-grid TV change is 5.93e−5 for a soft binned policy but remains exactly 1 for the learned hard rule.

Control: Grid refinement shrinks the soft jump, not the hard jump.