Your team knows what went into the result. Everyone else has to take your word for it. Colophon closes that gap with a public claim people can inspect for themselves.
A controlled benchmark of the grading prompts behind published LoCoMo scores. Six grading configurations judged the same 240 answers on the same model snapshot.
Changing only the grader moved agreement with the same labels from 60.8% to 87.9%.
A benchmark becomes valuable when people beyond its authors can rely on it.
Give the result a permanent place people can cite, inspect, and return to.
Give customers, reviewers, and competitors something stronger than your word.
Put a clear public record behind a product, research, or procurement decision.
You run the benchmark on the harness you already use, and bring the run back to be sealed. The published claim names who ran it.
Fix the tasks, setups, repetitions and grading before the run, so the method cannot be changed once results are in.
Every planned result is accounted for, failures included, and the claim is sealed with its evidence attached.
npx @colophon-claims/verify@0.2.1 ./bundleThe checker is available on npm.
Checking a published claim is free, and needs nothing from Colophon. Publishing your own claim from this site is not available yet. Until it is, write to ritsu.kai2000@gmail.com to publish.