Static audit of the supplied comparison pipeline

The findings below concern the supplied source, not an executed reproduction or an assumed later version.

• Matching tolerance is ignored. find_iso_dose_match accepts a tolerance but returns the nearest candidate regardless of its distance.

• The no-overlap fallback is not matching. Pooled percentiles do not establish the common support (eq:commonsupport). Even min–max overlap does not establish coverage of a discrete grid.

• Regrowth at zero is nonzero. The expression max(1, int(regrow_fraction * num_pruned)) activates at least one position in each eligible layer at fraction zero. Equation (eq:regrowcount) repairs the endpoint.

• Quantile and kth-value ties violate exact counts. Magnitude thresholding can retain or delete extra tied positions. Pruning at sparsity one can retain maximum-magnitude entries. The active-weight challenge can delete more than its requested count through the <= threshold rule.

• Eligibility can be lost. The main magnitude-pruning method overwrites masks using all stored weights. If a threshold is zero, previously inactive zero positions may be marked active. A corrected deletion operation restricts eligibility to the current mask and prohibits implicit regrowth.

• The allocation flag is ineffective. The accepted per_layer argument does not implement a global branch. An implementation should supply that branch or reject it explicitly.

• Several configured controls do not govern the default matching pipeline. The primary metric remains L1 even when a norm-choice variable exists. The configured target-dose list is replaced by generated targets. The declared learning-rate grid is not passed to the default adaptation sweep, which uses one rate.

• Sweep sparsity is stage-misaligned. The stored sparsity is obtained after secondary pruning, whereas the displacement metric and acute outcomes concern the pre-challenge checkpoint. Both stages must be named separately.

• The two challenge procedures differ. The main experiment uses a capped overall target sparsity followed by layer-wise quantile pruning. The sweeps remove an active-weight fraction using kth-value thresholding. They need not implement the same integer intervention.

• Bias, activation, and damping changes are unpriced. They can affect predictions without contributing to the weight metric. “No structural changes” should mean mask preservation, not absence of other implementation changes.

• Tanh precedes noise. The deterministic signal is bounded; the stressed hidden state is not, by Proposition prop:tanh.

• Ambient scoring is unsigned. The accumulated score is a sum of absolute minibatch derivatives. It is neither the signed gradient of the averaged loss nor the descent update of Theorem thm:descent.