Day91
§ Reversal

Open instrument · Both sentences always shown

When did they contradict themselves?

Paste two disclosures from the same company — two quarters, two filings, two investor decks. It finds the metrics that vanished, the stories that reversed, and the numbers moving against the narrative wrapped around them.

Why this one shows its working. The Fin-RATE benchmark published by Goldman Sachs and Yale in February 2026 tested language models on three analyst tasks and found accuracy dropped by roughly 18.6% on tracking one company's filings over time and 14.35% on comparing across companies — the two hardest tasks, and exactly what this tool does. So instead of summarising, it applies fixed rules and prints both source sentences next to every finding.
Mode 1

From the filings, without pasting anything

This mode reads nothing. It compares figures the company tagged itself, across its own submissions, and reports only differences that are arithmetic — a period filed twice with two different values, a disclosure that stopped, two measures that normally move together moving apart. There is no threshold to tune and no sentence to misparse, so where this finds something, the finding is a fact about the filings rather than a reading of them.

Mode 2

From two passages you paste

For anything not in the dataset, or not in a filing at all — an investor deck, a press release, a transcript. This mode is heuristic: it matches metric language and compares asserted directions, so it can misread a sentence. That is why it prints both source sentences beside every finding and leaves the judgement to you.

0 flags 0 metrics matched
Earlier
0 words
Later
0 words
Method

What counts as a reversal

A level is not a direction. "Declined to 90%" states where a number landed, not how far it moved, so only a signed delta can contradict prose. Nested metric names are resolved to the most specific match, so "revenue" inside "net revenue retention" is not counted twice.

Scope: English filings, read by fixed rules rather than a model. Feed it the MD&A and the metrics table rather than the whole filing and it will find more. Every finding shows the sentences behind it, so you can check each one.