Skip to content

Methodology

How a Yearover report is made, and when it isn’t.

A report is only worth as much as the filing behind it. This page says exactly where every sentence and every number comes from, what we check before publishing, and what happens when a check does not pass.

Last updated Sep 13, 2026 · parser-v4 · coverage changelog

What a report is, and what it is not

A Yearover report says what a filing says and what changed between two filings. That is the whole scope.

It is not investment advice. There are no buy, sell or hold recommendations, no ratings or scores, no price targets, no predictions and no signals. Not as a feature we have postponed, but as a line we do not intend to cross. What you do with a disclosure is yours to decide.

How a report is made

  1. 01

    Fetch

    Both filings come from SEC EDGAR directly, at no more than ten requests a second, with a User-Agent that says who we are and how to reach us. Nothing is scraped from a third party.

  2. 02

    Section

    Each filing is split into its named items (Item 1A Risk Factors, Item 7 MD&A) and then into paragraphs, so the unit of comparison is a passage a person can read, not a page.

  3. 03

    Match

    Paragraphs in this year’s filing are matched to the paragraph they came from in last year’s. Where no match is confident enough, the passage is reported as new or removed rather than guessed at.

  4. 04

    Number

    Figures are read from the filing’s XBRL data through a tool call. The language model never writes a number. Each figure keeps its tag, its period and the accession number it came from.

  5. 05

    Verify

    Every quote is checked back against the filing text and every figure against its XBRL fact. If a check does not pass, the report stops here.

What “matched against the filing text” means

Filings arrive with typographic quotes, en dashes, non-breaking spaces and line breaks that carry no meaning. Before matching, we normalise those: quotation marks and dashes are standardised and runs of whitespace collapse to one space.

So a quote is matched against the filing text, not character for character. We say it this way deliberately, because the stronger claim would not be true. The words, their order and their meaning are the filing’s; the punctuation may have been normalised.

Where the numbers come from

Every figure in a report is read from the filing’s own XBRL data through a tool call. The language model does not write numbers, and it does not restate them from the surrounding prose.

Each figure keeps three things with it: the XBRL tag it was filed under, the period it covers, and the accession number of the filing it came from. If a figure cannot be tied to all three, it does not appear.

Revenue · FY2025$130,497,000,000us-gaap:Revenues · 2024-01-29 → 2025-01-26 · 0001045810-25-000023

The verification gate

Before a report is published, four things have to hold:

  • Every quoted sentence appears in the filing it is attributed to.
  • Every paragraph is matched to the right paragraph in the earlier filing.
  • Every figure carries an XBRL tag, a period and an accession number.
  • Every section is the section it claims to be.

When one of them fails, the report is held.

It is not published with a gap in it, a hedge, or a confidence score standing in for a check. The filing’s page says the report is held and the coverage changelog records why. A missing report is better than a wrong one.

Why this is built the way it is

The obvious question is why any of this is necessary when a general-purpose model can read a filing. It is necessary because, measured, they do not read them reliably.

Patronus AI benchmarked leading models on 150 questions answered from SEC filings (FinanceBench). Each answer was labelled correct, incorrect or a refusal to answer.

  • GPT-4-Turbo, retrieval over all filingsCorrect: 29 of 150Refused to answer 68% of the time
  • Llama 2, retrieval over all filingsCorrect: 29 of 150Wrong 70% of the time
  • Claude 2, whole filing in contextCorrect: 114 of 150Only with the whole filing in context, which is impractical at 10-K length
  • GPT-4-Turbo, told where to lookCorrect: 128 of 150Which defeats the purpose of automating it

Source: Islam et al., FinanceBench: A New Benchmark for Financial Question Answering, arXiv 2311.11944, Table 2 (n = 150); CNBC, 19 December 2023.

The last row is the important one. Accuracy rises when the model is told exactly where to look, which is the part a person was trying to avoid doing. Tool-called XBRL figures and a verified quote are what close that gap; a longer prompt is not.

Corrections and coverage

Reports are the same for every reader, so a mistake in one is a mistake everyone sees. Corrections are logged publicly with what was wrong, when it was noticed and what changed. They are not edited away quietly.

The coverage changelog records which filings have been processed, which are held and why, and which companies were added. If you find something wrong in a published report, tell us.

Who builds this

Yearover is built by Gorkem Yagci and operated by Shiftbase Labs Ltd, a company registered in England and Wales (no. 17027574) at 71-75 Shelton Street, Covent Garden, London WC2H 9JQ.

A named person stands behind these reports on purpose. Every claim here can be checked against a filing on SEC.gov, and there is a real address to write to when it does not hold: hello@yearover.com.

Yearover is not affiliated with the U.S. Securities and Exchange Commission.