Both halves of the cancellation are checkable. The wrong layer count added 256 experts at 14,155,776 bytes each, which is 3,623,878,656 bytes, or 3.62 GB of routed experts [13][1]. Net of the dense layer that extra mixture-of-experts layer displaced, the inflation was +3.54 GB, so the thing it displaced was worth roughly 80 MB [13][2]. Against that sat -5.27 GB of parts the model never attempted: the vision and audio encoders, the multi-token-prediction module, the RMS norms, and the published conversion's choice to keep non-expert tensors at 8 bits rather than 4 [14]. The sum lands 1.73 GB low, reported as 1.1% against a 2% tolerance [15][3].
Back the container out of those two pairs. A 1.73 GB miss at 1.1% implies a published file near 157 GB; a 5.27 GB miss at 3.4% implies near 155 GB [4]. The coverage gap was always about 5 GB on a roughly 155 GB artifact, and the layer-count error was hiding most of it [14][3].
What failed here lives in interpretation, not arithmetic: a config file describes an architecture in the vocabulary of the framework that trained it, so reading one means inferring what a field name meant to its author [20]. `dense_mlp_idx=2` supports "the layer at index 2 is dense" and "layers below index 2 are dense", which give 41 and 40 mixture-of-experts layers [5]. The GGUF header is downstream of that question: the conversion script writes its key/value metadata after it has already resolved every ambiguity of this kind [7]. Forty-two blocks, two of them dense, leaves forty routed layers, which is the second reading [9][8].
The original limitations section addressed this directly. It bounded the effect on bytes per token at 2.4% and called it immaterial to the conclusion [4]. That bound held for the quantity it described, but it said nothing about the validation residual, which moved 2.3 percentage points and crossed a tolerance set at 2% [7][6] - roughly a threefold increase in reported error [6]. For 1.1% to have meant what it looked like, the uncounted parts needed to net to zero and the layer count needed to be right; instead, both were wrong [14][8].
Gate 1 is the part worth copying. It compares computed integers against a published layout with no tolerance at all, and 256 experts at a stride of 1,769,472 bytes is exactly the 452,984,832-byte per-layer blob it checks against [17][5]. Equality has no room inside it for two errors to meet, which is the argument for preferring an exact gate wherever the domain publishes exact integers [16][18]. The load-bearing output here, bytes streamed per token, depends on that stride, and the author reports it improved rather than degraded [19].
Reading the metadata took about forty lines of Python, and the parser was wrong the first time in a way that generalises [9]. GGUF stores string arrays length-prefixed; consuming only the first few elements shifted every subsequent offset, and the parse died later at an invalid-looking type code rather than at the array that caused it [10]. The format reports the failure in the least useful place available. Consume every element even when you only intend to print four [10].
