RiskDiff

What it is

RiskDiff takes a company's two most recent 10-Ks, pulls Item 1A out of each, and shows what changed between them. The methodology page describes the pipeline in unusual detail: the filing is fetched from sec.gov by CIK, Item 1A is located by one of four HTML strategies and the filing is skipped if all four fail, each prior-year risk factor is matched to a current-year one by TF-IDF cosine similarity, a pair above 0.92 counts as unchanged and below it as modified with a difflib sentence diff, and every change gets a rule-based 1 to 10 score from its status, its similarity, keyword hits from a themes list and how much the text grew. The same input always yields the same score, and the page says so.

On top sits a one-sentence AI brief per comparison with a direction (more concerning, less concerning, mixed, neutral) and up to three themes, badged as AI and persisted so it does not change between visits. Across companies, the directory sorts the 434 filers by summed change score and filters them by nineteen themes; the home page's own summary on 27 October 2026 was that 149 companies had newly added AI risk language and 62 had added tariff exposure. That cross-company view is what the cheaper plan sells.

The vendor is RiskDiff and nothing more is published about it: no company name, no address, no founding date, and the terms link in the footer returns a 404. The product submitted itself to this catalogue from an address at its own domain, signed by an AI-assisted outreach team.

Pricing

$29 a month for Signals, $199 a month for Pro, both month-to-month with a 7-day trial that asks for no card. The billing toggle offers 17% off for annual and prints no annual figure. Signals is the theme index and the company list, in the browser and over MCP, with no company diffs and no history. Pro is every change for every tracked company rather than the top five, year-over-year history, the sentence-level diffs with their confidence labels, and the diff tools over MCP. Bulk export is listed as not included, although the methodology page says Pro subscribers reach complete exports under a /full/ path, and the pricing page says 300+ companies where the home page and the directory say 434.

There is no free plan. What is public is a preview of every comparison: the four counts, the AI brief, one unlocked change with its score, and links to both filings on EDGAR, served as HTML, Markdown and JSON. Nike's 2026 versus 2025 preview shows 0 added, 1 removed, 5 modified and 35 unchanged, unlocks one modified factor, and locks the other five. The full page redirects to sign-in.

Data & coverage

434 US companies, 10-K filings only. The directory does not say how they were chosen; the pricing page's example question refers to S&P 500 companies, and the list reads like large caps. A foreign private issuer filing a 20-F is not here, and neither is a 10-Q, so a risk factor that changed mid-year is invisible until the next annual report.

History is shallower than "historical comparisons" suggests. The site's own llms.txt listed 1,342 comparison pages on 27 October 2026: 404 for 2026 versus 2025, 431 for 2025 versus 2024, 428 for 2024 versus 2023 and 77 for 2023 versus 2022. For most companies that is three comparisons reaching the fiscal-2023 filing; a fifth reach one year further. The index also carries a 2025 versus 2025 and a 2025 versus 2023 pair, which is what a mechanical pipeline does with a filer that changed fiscal year.

Every claim links to an accession number on EDGAR. The comparison page says in its own words that classification is based on text similarity and may include approximations, and the methodology page publishes no list of the filings it skipped.

Integrations

An MCP server at /api/mcp, keyed from the account menu, is the only programmatic access. A POST without a key returns 401 and a hint that an active plan or trial is required. No REST API, no Python client, no webhooks, no spreadsheet add-in, no alerts on any plan page. The public previews are fetchable as index.md and index.json per comparison, which is enough to script the counts and the one unlocked change across all 434 companies without an account.

Limitations

• 434 companies and 10-Ks only. If the filer you cover is mid-cap, foreign or newly public, the tool has nothing on it, and a 10-Q change waits for the annual report.
• The score ranks text change, not risk. A reworded sentence about climate regulation and a new going-concern paragraph can land on the same number; the methodology page says the score does not estimate financial impact.
• TF-IDF matching is approximate, by the vendor's own admission. A factor split into two or merged with another will show as a removal plus an addition.
• No annual price is printed, no legal entity is named, and the terms page is a 404. For a $199 a month subscription, the privacy page and the Stripe receipt are all there is to read.
• The AI model is unnamed, and the "confidence" attached to a match is not defined anywhere on the site.

Alternatives

CapEdge diffs two 10-Ks or 10-Qs as a red-and-green document comparison for any EDGAR filer, free, without scoring or themes; it is the right answer if your company is not among the 434 or you want the whole document rather than Item 1A. BamSEC has document comparison in its Pro plan at $69 a month billed annually, alongside the rest of the filings corpus. sec-api.io returns a named item out of a 10-K through its Extractor API, the raw material for a risk-factor diff you write yourself, and SEC EDGAR MCP splits a filing into its sections for your own model with no paywall. For the numbers rather than the prose, see fundamental data and research platforms.