Convexity Redline · what changed in what they file

Sheet 38 · Redline

Companies edit their filings.We read the edits.

A 10-K is 200 pages of which 194 are last year's. The story is in the six that moved: the sentence that got rewritten, the risk factor that appeared, the segment table that broke out a new line. Convexity redlines each filing against the last, grades every 8-K by what its item codes actually mean, and extracts the tables into numbers you can use. The filing is the disclosure; the diff is the news.

Item 1A · Risk Factors · FY25 10-K against FY24sample data
newexport controlsnewconcentration riskgonecovid supply
  1. modified

    We depend on a limited number of third-party foundries to manufacture our products.

    We depend on a single third-party foundry for substantially all of our advanced products, and we have no qualified second source.

  2. added

    New or expanded export controls could restrict our ability to sell our most advanced products into certain markets.

  3. removed

    Pandemic-related disruptions may continue to affect our supply chain and logistics costs.

3 changes in this section1 added · 1 removed · 1 modified
Ingest, redline, grade

Reading a filing is not the job. Finding the edit is.

Search gives you the sentence you asked for. It cannot tell you that the sentence was different last quarter. Convexity keeps every filing parsed and diffed against its predecessor, so the question you actually have, what changed, has an answer that does not depend on you remembering the old language.

the corpus, structured

Ingest

Every 10-K, 10-Q and 8-K lands parsed rather than stored as a blob. Sections are isolated by their Item heading, tables come out as rows and columns carrying the section they sat in, and transcripts are chunked by speaker.

  • Item-level section isolation
  • Structured table extraction with provenance
  • Full-text and semantic search
sentence-level, deterministic

Redline

Each filing is diffed against its predecessor sentence by sentence, per Item section: what was added, what was removed, what was rewritten. Risk-factor themes are tracked across quarters as they appear and disappear. No model touches the diff, so it cannot report a change that is not in the text.

  • Added, removed and modified, per section
  • Risk themes tracked across quarters
  • Same two filings, same redline, every time
graded, then routed

Grade

An 8-K is graded from what its item code means, not from how the headline reads. A financial restatement and a routine results announcement are both 8-Ks, and they are not the same news.

  • Item-code severity map
  • Materiality-ranked change feed
  • Deep links that open the document itself
What gets diffed

Five sections carry the news. Those are the five we isolate.

A 10-K is not one document. It is a set of Item sections with different jobs, and a change in the risk factors means something different from a change in the business description. Each is located by its own Item heading and diffed against the same section of the prior filing, never against the document as a whole.

  1. Item 1ARisk Factorswhat they now say could go wrong
  2. Item 7MD&Ahow management explains the quarter
  3. Item 1Business Descriptionwhat the company says it is
  4. Item 8Financial Statementsthe statements and their notes
  5. Item 3Legal Proceedingswhat is being litigated

Each section carries its own vocabulary, so the tone reading is per section rather than one average for a 200-page document. When a heading cannot be located in both filings, that section returns empty rather than diffing whole documents into noise.

Tables, extracted

A filing table is HTML. It should be numbers.

The segment table you want is buried in a 200-page document among the layout tables that hold the cover page together. Convexity separates them by how numeric they are, so the financial statements win and the boilerplate loses, then keeps each table at its original position in the document.

  • Header row and row grid, not a screenshot
  • Tagged with the Item section it sat in
  • Caption pulled from the nearest heading
  • Ordinal is the document position, so provenance never shifts between runs
Item 8 · table 41 · Revenue by segmentsample data
SegmentFY25FY24Change
Data Center79,60015,005+430%
Gaming11,3509,067+25%
Professional Visualization1,8961,553+22%
Automotive1,091903+21%
4 columns · 4 rowsanchored to Item 8
8-K item grading

Every 8-K looks urgent. The item code says otherwise.

An 8-K carries item codes that name what it discloses. A restatement files under 4.02 and a quarterly results release files under 2.02, and a feed that treats those as one stream is a feed you learn to ignore. Convexity grades on the code, so the ranking survives whatever the headline says.

Selected 8-K item codes and the severity Convexity assigns them
ItemWhat it disclosesGrade
4.02Financial Restatementcritical
2.06Material Impairmentcritical
4.01Auditor Changecritical
5.02Executive Changehigh
2.01Asset Acquisition/Dispositionhigh
2.02Financial Resultsmedium
5.07Shareholder Votemedium
8.01Other Eventslow

Eight of the graded item codes, shown as they are defined in the product.

No model in the loop

A diff is a fact about two documents. It should never be generated.

Both approaches produce a readable list of changes. Everything around the list is the difference:

Where the answer comes from
asking a chatbotModel memory, plus whatever of two long documents fits the context window.
the redlineThe two stored filings, compared by a sequence matcher over sentence lists.
A change that is not there
asking a chatbotRenders exactly as fluently as a real one, and reads identical to the truth.
the redlineCannot occur. Nothing in the path writes prose, so there is nothing to invent with.
Asking twice
asking a chatbotTwo runs, two summaries, and no way to tell which one drifted.
the redlineByte-identical. The diff is a property of the documents, not of a sampling temperature.
When it cannot answer
asking a chatbotAnswers anyway.
the redlineSays it is not comparable, gives the reason, and returns a null as-of.

Every change links back to the document it came from, at the accession the citation names, so a claim in an answer opens the filing that supports it.

Stop re-reading the filing. Read the diff.

Sentence-level redlines across five Item sections, tables extracted with their provenance, and 8-K severity grading, on the same desk as the filings themselves. Redline is on Pro and above.

Founder pricing holds for life while the waitlist is open.