Developer's Cost Model Self-Corrects After Config File Ambiguity Exposed Hidden Error

A developer building a cost model for streaming a mixture-of-experts AI model off an SSD discovered a critical architectural misreading after the model's built-in validation gate flagged a failure. The error stemmed from an ambiguous field in the model's config.json file — dense_mlp_idx=2 — which could mean either one or two dense layers, affecting the total mixture-of-experts layer count. By parsing the model's GGUF file metadata, the developer confirmed the correct value was two dense layers and 40 mixture-of-experts layers, not 41 as previously assumed. This correction shifted the Gate 2 prediction accuracy from 1.1% off to 3.4% off, pushing it outside the 2% tolerance and causing the model to refuse to output any token-rate estimate. The developer noted that the apparent earlier accuracy of 1.1% was coincidental, the result of two opposing errors cancelling each other out rather than a genuinely correct prediction.