PUBLIC METHODOLOGY / V0.2
MOEModels separates what a publisher’s metadata establishes, what a reviewed source says, what this project measured, what the engine calculated, and what remains uncertain. Five different things, never merged into one number.
Corrections update the record without erasing its history.
CLAIM CLASS
Computed mechanically from a publisher API response, with every field naming the response field it came from. Exactly as trustworthy as the publisher's own metadata, and never a measurement.
commit revision · per-dtype bytes · routed experts · declared licenceCLAIM CLASS
A directly supported value from a model card, technical report, repository, artifact, or vendor specification, reviewed by a person.
parameter count · context · license · expert topologyCLAIM CLASS
A result this project produced from an identifiable configuration, with raw artifacts retained and a repeatable methodology.
TTFT · throughput · VRAM peak · routing skewCLAIM CLASS
A deterministic output whose inputs, formula, range, and conservative assumptions remain visible.
checkpoint lower bound · usable memory · topology roundingCLAIM CLASS
A reasoned conclusion derived from incomplete evidence and explicitly separated from sourced or measured claims.
likely bottleneck · deployment risk · validation priorityFIT ENGINE / FIRST PRINCIPLES
Fit Check uses exact tensor bytes from a pinned artifact manifest, applies a declared memory reserve, and rounds the result to whole-node topology. Runtime allocations, KV cache, activations, compatibility, throughput, latency, and cost remain unknown until measured.
checkpoint_bytes = pinned_manifest_tensor_bytesusable_vram_per_gpu = advertised_vram × (1 − reserve_bps / 10,000)minimum_gpus = ceil(checkpoint_bytes / usable_vram_per_gpu)topology_gpus = ceil(minimum_gpus / gpus_per_node) × gpus_per_nodeCORRECTIONS / GOVERNANCE
The model release, checkpoint, runtime, hardware, methodology, and retrieval date form a single claim context.
Quantization, templates, reasoning budgets, hardware, and harness versions can change the meaning of a score.
The first engine favors deterministic memory math and conservative envelopes over uncalibrated throughput claims.
A vendor may support research. It cannot purchase a result, confidence class, or placement.