What You Actually Download: Shipped Defaults Are Mostly Harmless and Occasionally Catastrophic for Browser LLMs

Language models that run inside web browsers are chosen from leaderboards whose scores were measured under controlled evaluation conditions: greedy or standardized decoding on full precision weights. What a browser user runs is a quantized artifact executing under whatever sampling configuration the packager shipped. We measure this gap directly for seven WebGPU model artifacts on GSM8K, using one harness, one prompt, and one grader across all conditions. We first measure the two explanations practitioners reach for: 4 bit quantization costs a modest 4.4 points at 1.5B on the native reference stack (p=0.045, n=500, uncorrected), and we detect no runtime difference between two 4 bit stacks at matched configuration (90% CI $-3.6$ to $+0.8$ points, n=1319). The remaining, far larger and model dependent factor is configuration. Shipped sampling defaults shift six of seven models by $-1.9$ to $+1.3$ points on matched items, but cost SmolLM2 1.7B 17.3 points, collapsing it from 26.4% to 9.1%. The same temperature 1.0 configuration that breaks SmolLM2 costs Phi 3.5 mini only 1.7 points, so damage is a property of the configuration and the model jointly, not of the configuration alone. Tracing the provenance of all seven shipped configs against upstream files, we identify three failure modes in the vendor to packager pipeline: faithful copying of an opinionated vendor setting (Qwen's repetition penalty of 1.1, $-5.8$ points at two model sizes, $p \approx 0.01$, yet harmless when cross wired onto the 3B), partial copying that drops a parameter (Qwen's top_k never reaches the browser), and void filling when the vendor's machine readable config specifies nothing (three of seven models; harmless twice, catastrophic once). SmolLM2 at its vendor's recommended settings, which exist only in README prose, scores 28.4%, at or above its exam score. We release the harness and all per item results.