loading…
The store is bitemporal: every value carries when it was true in the market and when this system first knew it. Nothing is overwritten, so an earlier answer stays answerable after any correction.
the finding
That question decides whether an index over a token good is coherent, and it is answerable before any estimator runs. Every contract in the benchmark, drawn by the sellers offering the identical good:
A benchmark measures price discovery. Only Anthropic sells Opus 5. Only OpenAI sells GPT-6 Astra. There is no second price for the same good, so there is nothing to discover, no dispersion to measure, and no sense in which an average of them is a market rate.
Averaging across different frontier models is worse. It silently assumes those models are substitutes — which is precisely the assumption a buyer is paying not to have to make.
So TIX-FRONTIER-OUT exists in this repository to be
refused. Its contract is deliberately ill-defined, because that fiction
is the point: try to write the contract down and you find there is no single
good to write.