Estimator contract
The useful number is not the most precise-looking one. It is the one whose limits you can inspect.
01 · Throughput
Token time is the slower of memory time and compute time, plus exposed per-layer overhead. Before measured calibration, that overhead is explicitly zero and every decode result is downgraded.
t/token = max(bytes / (bandwidth × efficiency), FLOPs / (peak × MFU)) + overhead
02 · Capacity
Weights, position-dependent KV storage, and backend activation allowance must fit inside the configuration's usable memory—not its sticker capacity.
03 · Boundaries
Factor-level ambiguity publishes no number. The calculator names the missing execution path or compute input instead of displaying a dash, a zero, or an invented estimate.