Every claim has a document behind it.

What counts as a source

SEC filings — 10-Qs, 10-Ks, S-1s, 8-K acquisition and results disclosures. Company investor relations releases. Live pricing pages, diffed week over week. Job board APIs. Nothing enters an issue because another publication reported it.

Where a figure comes from a vendor dataset rather than a filing, the signal says so in the FACT block. A labelled secondary source is acceptable. An unlabelled one is not.

AI-drafted, human-approved

An agent pipeline collects the documents and drafts the FACT blocks. A human reads every issue before it ships and cuts more than it keeps. Nothing publishes automatically, and there is no configuration in which it could — auto-publish would put the one asset this product cannot replace, its credibility, on an unattended process.

Inference is labelled as inference

Filings state figures; they rarely state ratios. When a signal reports a multiple, a spread or a share, that arithmetic is ours and the entry says so — including which direction the real number sits if the disclosed input is a floor or a range.

Corrections are public

If a link is wrong or a number is misread, reply and it is corrected in the next issue with the correction stated plainly, and added to the public log. A corrections page nobody ever needs is a claim. A corrections page with entries in it is evidence.

The tools have no backend

The Check, the Grid, the Mirror, the Diff, the Reversal, the Battlecard and the benchmark tool all run entirely in your browser. There is no upload endpoint, no account and no server to send anything to. This is a deliberate architectural choice: a product about other companies' disclosures should not be accumulating yours.

Two things reach outside the browser and both are read-only. Pages you ask the Mirror or the Diff to fetch travel via a public relay, because a browser cannot read another origin — only the URL travels, never anything you typed. And the MCP endpoint is a stateless worker that reads this site's own JSON and logs nothing. Neither holds a record of what you looked at.

The benchmark that shaped how these tools work

In February 2026, researchers at Goldman Sachs and Yale published the Fin-RATE benchmark. It tested language models on the three things analysts actually do: pull a fact from one filing, compare disclosures across companies, and track how one company's filings change over time. Accuracy dropped by roughly 18.6% on the longitudinal task and 14.35% on the cross-company one — the two hardest, and the two that produce any real edge.

Those are precisely the tasks The Reversal and The Grid perform. So neither of them summarises anything. They apply fixed, published rules, print the sentences they read, and mark every output with how good its sourcing is. Where a rule can't answer, the cell stays empty and hands you the EDGAR search instead of a plausible guess.

That is also why a human approves every issue before it ships, and why coverage grows slowly. The bottleneck isn't a limitation being apologised for — it is the reason the output is worth putting in front of a board.

Tell us where the method breaks.

There is nothing to buy yet, so there is no demo to book. What is useful is the claim we graded wrong, the company we haven’t covered, or the rule that produced a bad answer.