JP Journal lab · bounded local artifact

AWQ vs GPTQ toy bench

Compare round-to-nearest, AWQ-inspired activation scaling, and GPTQ-inspired reconstruction on one fixed 6×5 matrix. The held-out vectors are not used by calibration.

Truth boundary: this tiny synthetic bench exposes method shape. It is not AWQ or GPTQ production code, a real-model benchmark, or evidence of kernel speed or downstream quality.

Bits

Frozen weight matrix

Calibration activation saliency

Held-out reconstruction

Method	Weight MSE	Held-out MSE	Role

Release gate

No method wins here. A production choice still requires the real checkpoint, held-out product slices, supported packing and kernels, and serving measurements.

Ready. The matrix, calibration vectors, and held-out vectors are fixed.