Merlise achieves breakthrough results on SciFact-Open
Merlise
Financial · Failure mode

The sentence said improved. The filing did not agree.

May 20264 min read

Most people never read a filing directly. They read a headline, a summary written by an analyst, or a paragraph a model generated from the numbers. A layer between the filing and the reader is convenient, and a layer is also a chance for the figure to drift, rounded wrong, pulled from the wrong period, or left standing after a correction changed the number underneath it.

Revenue grew. Loss narrowed. Guidance increased. Three phrases, and only the filing knows if any of them are still true.

Two different kinds of mismatch

A direct mismatch is easy to spot once you look: an article says the company reported 2.4 million and the corrected filing says 1.9 million. An interpretive mismatch is harder. An article says performance improved while the filed figures show the opposite, or show enough uncertainty that "improved" overstates what happened.

To Merlise, a sentence like "revenue grew" reads as a structured claim, not simply prose. Two periods sit hidden inside the comparison. Pull the filed values behind them, and check whether the implied comparison actually holds.

A verdict and a trail, not one or the other

The short verdict might read "disputed: narrative does not tie to current filing." Behind it sits the original source, the corrected source if one exists, and the exact line item that changed. A writer gets a specific place to look instead of a vague sense that something is off.

This also helps with prose a model produced. Smooth financial writing is easy to generate and easy to mistake for accurate, because fluency and correctness are not the same property. Merlise sits between the generated paragraph and the filings it claims to summarize, checking the connection instead of assuming it.

The evidence ledger

The same claim by claim view the product shows on a live document, built from this case.

VERIFICATION LEDGERNarrative claim, checked against the filing
"Revenue improved meaningfully compared with the prior quarter."Numeric
31%
Filed figures show a smaller gain than the sentence implies once the restated quarter is used.Disputed
10-Q and 8-K
"The company reported 2.4 million in quarterly revenue."Numeric
28%
Ties to the original filing but not the corrected figure.Disputed
10-Q, filed May 12

How it resolves

The resolving record

NeuroOne's own filing chain, the original 10-Q and the later 8-K, provides the clearest current example of a precisely quantified gap between a stated figure and a later corrected one.

NeuroOne Medical Technologies, SEC EDGAR