Model Benchmark Lab

Choose a published calculator, build targets from one to four of its reference rows, then choose the candidate references available to the fitting method. Each target's weights must total 100%. The server constructs the synthetic target from the recorded public rows.

Constrained fitting uses nonnegative weights totaling 100%. Convergence is checked with an objective-gap bound; unfinished fits retain their solver status. The experiment has a five-second budget and may return partial results. The closest single reference assigns 100% to its nearest candidate. Distances use scaled G25 Euclidean units multiplied by 100.

Keeping the generating references tests reconstruction within the same panel. Withholding a reference changes the available model. Neither is independent validation. Convergence bounds the numerical objective error; it does not establish unique weights. Similar references may support very different weights at nearly the same distance.

Weight reassigned is half the total absolute weight difference, measured in percentage points. Extra allocation is recovered weight on candidates with an expected weight of zero. In mixed targets it does not identify which source was replaced. Pure-source controls make those allocations easier to inspect.