The evidence, not the argument
This site is the public proof package behind Corograph's benchmark numbers. Every page except this one is generated from a single export of the benchmark's own store: 2,692 runs across9 arms and 11 models in Epoch-4, 2,480 of them valid and scored, with every question, answer sheet, judged verdict and transcript published beside the figures. Every number on the site links to the run that produced it.
Who ran it
This is Corograph's own benchmark. It was designed, run, scored and published by Jasdeep Jaitla, the founder of Metaphori, Inc., the company that builds Corograph. We built the tool under test and we ran every arm, including the comparison group. That is a conflict of interest, and the answer to it is the package itself: nothing on this site is a summary of something you cannot open. The Procedurepage states what was measured and how, the Limitations page collects every caveat, and the Data page points at the files and the export that made them.
We did not write the questions. Each battery was authored by a frontier model working as an agent over the pinned repository with no access to Corograph, its index or its outputs; the gold answers were drafted by the same agent and then verified against the checkout, with and without Corograph, with corrections recorded with file:line evidence before the batteries were frozen. The full sequence, the review loop that produced the exemplars, and the dates are in How the questions and the grading key were made.
The argument for why context is the thing to measure, and the story of how this benchmark came to exist, are not on this site. They belong to the product and the company, and they live there.
MetaphoriA different approach to interpretability, from the outside in.The company. Metaphori, Inc. is an AI research company working on how semantic structure shapes attention in transformer models, and the maker of Corograph, TrueMemory, ContextMax and ClaudeFu. The longer essay on context, and on why this benchmark exists, is on the Metaphori blog.metaphori.ai ↗ · metaphori.ai/blog ↗What this site records about you
A cookieless counter. The site loads a small script from insight.corograph.com, a server we run, which records page views, the referrer, and a handful of named events: an early-access link clicked, a run directory opened, a figure's data downloaded, a figure control changed, a card's details opened. For a sample of visits the same server also records where on the page clicks land and how far the page is scrolled, kept only as an aggregate heat map over the page; no visit is played back. It sets no cookie, stores no IP address, and carries no identifier that survives the day. There is no advertising or third-party analytics on this site. The early-access links lead to corograph.com, which is where any form lives; this site collects nothing itself. Local and preview builds of the site load no counter at all.
Reproducing it
The whole tree, this site included, is one repository:github.com/Corograph/benchmarks. The site is a static build over epochs/E004/summary/runs.json; the figures, the scorecard and the cards are all functions of that one file, so they cannot disagree with each other or with the runs. Re-exporting the epoch from the benchmark's store reproduces the tree byte for byte.