What 67 commercial contracts filed with the SEC in 2025–26 actually say about liability caps, indemnities, and renewal terms — extracted by LLM agents, linked to every source document, and published as distributions instead of anecdote.
When a lawyer across the table says a term is “not market,” there is no public data to check the claim against. Bloomberg Law and Law Insider sit on the same public corpus this report uses — the material contracts public companies must file as EX-10 exhibits on EDGAR — but both sell single “most standard clause” comparisons behind subscriptions. Neither publishes a percentage. This pilot tests whether LLM extraction can turn that corpus into open clause-prevalence statistics, and whether the resulting numbers are crisp enough to be worth citing.
Across the whole sample only 22% of contracts cap aggregate liability. But the overall number is close to meaningless: a majority of services agreements cap (53%), while license, manufacturing, and supply agreements almost never do — 8%, 0%, and 22% respectively. This is the core product thesis in one chart: a single “market standard clause” answer, which is what incumbents sell, papers over exactly the variation a negotiator needs.
Share of contracts · 8-K item 1.01 EX-10 exhibits, filed Jan 2025 – Aug 2026
One dot per contract (n = 8)
Carve-outs — the liabilities a cap does not protect — are the highest-leverage sentence in a limitation-of-liability section. In the 15 capped contracts:
Contracts carving the item out of the cap (n = 15 capped)
The most common negotiating trap is a clause that is standard in kind but one-sided in direction. In this sample, when consequential-damages waivers and indemnities appear at all, they are usually mutual — a one-way version (9% and 7% of all contracts respectively) is a genuine outlier a negotiator can push back on. Termination for convenience runs the other way: most filed contracts lock both parties in, and when a walk-away right exists it is more often one party’s alone.
Share of contracts (n = 67)
Delaware (30%) and New York (18%) govern nearly half of the sample — but 30% of contracts specify no governing law at all in the filed exhibit. Auto-renewal appears in just 15% of contracts; where it exists, the opt-out notice window runs 60–365 days with a median of 90 — the long tail (two contracts require a full year’s notice) is exactly the kind of term this data exists to flag.
Contracts (n = 67)
The full extracted dataset. Each row links to the underlying exhibit on sec.gov — the citation is the product.
| Filer | Type | Filed | Cap | Basis | CDW | Indemnity | TFC | Auto-renew | Law | Redacted | Conf. | Src |
|---|
EDGAR full-text search for commercial-agreement phrases, restricted to EX-10 exhibits in 8-K filings with an item 1.01 disclosure (entry into a material definitive agreement) — this filter removes most compensation filings before download. 150 exhibits fetched → 87 survived keyword and size triage → 8 parallel LLM extraction agents coded each contract against a fixed 25-field schema → 20 classified out of scope on reading (securities, real estate, M&A) → 67 in-scope rows aggregated to SQLite.
Extraction — the only real cost — ran at roughly a cent or two per contract at current model prices. Human coding of the same fields is typically quoted at dollars per contract-field. That three-orders-of-magnitude gap is why this database can exist now and could not in 2020.
Scale to ~2,000 contracts across 2024–26 filings, add the verification pass, and publish the benchmarks as a browsable public site where every statistic cites its sample and links its contracts — the free, transparent layer the paid tools deliberately don’t ship.