Provisional patent applications are often treated as a quick placeholder: get a filing date on record, then do the real work in the non-provisional twelve months later (37 C.F.R. § 1.78(a)(1)). Because the U.S. does not require claims in a provisional at all (35 U.S.C. § 111(b)(2); 37 C.F.R. § 1.53(c); MPEP § 201.04), many provisionals are filed with no claims, or with a handful of loosely drafted ones.

That is a missed opportunity. A provisional application gives you a priority date only for what it actually discloses. One of the most efficient ways to lock in disclosure of the many feature combinations you may need later is to include a claim set built around multiple dependent claims, and even multiple-multiple dependent claims. In a U.S. non-provisional application, multiple dependent claims carry extra fees and multiple-multiple dependencies are prohibited outright. A provisional application, by contrast, is subject to neither restriction.

First, the terminology

  • Dependent claim: refers back to a single earlier claim and adds a limitation (“The method of claim 1, wherein…”).
  • Multiple dependent claim: refers back, in the alternative, to more than one earlier claim (“The method of claim 1 or claim 2, wherein…” or “The method of any one of claims 1–4, wherein…”).
  • Multiple-multiple dependent claim: a multiple dependent claim that refers back to another multiple dependent claim. For example, claim 5 depends on “any one of claims 1–4,” where claims 3 and 4 are themselves multiple dependent.

Here is how the USPTO treats each form in a non-provisional application:

  • Multiple dependent claims are allowed, but expensive. They must refer to the earlier claims in the alternative only (“claim 1 or claim 2,” never “claims 1 and 2”) (37 C.F.R. § 1.75(c); MPEP § 608.01(n)). An application that contains one triggers a one-time, per-application multiple dependent claim fee under 37 C.F.R. § 1.16(j). As of the USPTO fee schedule effective January 19, 2025, that fee is $925 for large entities, $370 for small entities, and $185 for micro entities.
  • They also inflate the claim count. For excess-claim fees, a multiple dependent claim counts as the number of claims it refers to, and so does every claim that depends from it (37 C.F.R. § 1.75(c); MPEP § 608.01(n)). A single claim reciting “any one of claims 1–10” counts as ten claims, which quickly pushes an application past the 20 claims included in the filing fee, after which each additional claim incurs an excess-claim fee (37 C.F.R. § 1.16(i)).
  • Multiple-multiple dependencies are prohibited outright. Under 35 U.S.C. § 112(e), “a multiple dependent claim shall not serve as a basis for any other multiple dependent claim,” whether it depends on that claim directly or indirectly. The examiner objects to such a claim as improper under 37 C.F.R. § 1.75(c), and it is not examined on the merits until it is rewritten (MPEP § 608.01(n)).

As a result, U.S. practitioners rarely use multiple dependent claims in non-provisional applications, and never use multiple-multiple dependencies.

The real problem: support for combinations

To claim the benefit of a provisional’s filing date, the provisional must provide written description support, under 35 U.S.C. § 112(a), for the invention later claimed (35 U.S.C. § 119(e); 37 C.F.R. § 1.78(a); New Railhead Mfg. v. Vermeer Mfg., 298 F.3d 1290 (Fed. Cir. 2002); MPEP § 211.05). Priority is evaluated claim by claim. If a non-provisional claim recites a combination the provisional does not adequately describe, that claim gets only the later filing date.

The trap is that disclosing each feature individually is not necessarily the same as disclosing every combination of those features. Written description is a fact-specific inquiry into whether the original disclosure, read as a whole by a person of ordinary skill in the art, reasonably conveys that the inventor possessed the claimed invention (MPEP § 2163). There is no categorical rule that combining features from different parts of a specification lacks support, and verbatim disclosure of the exact combination is not always required. But courts have repeatedly found support lacking where the claimed selection or combination could be identified only in hindsight:

  • In re Ruschig, 379 F.2d 990, 994–95 (C.C.P.A. 1967): a broad genus disclosure encompassing numerous possible compounds did not support a claim to a particular species, because the specification lacked “blaze marks” directing a skilled artisan to that specific compound.
  • Purdue Pharma L.P. v. Faulding Inc., 230 F.3d 1320, 1328 (Fed. Cir. 2000): a later-added claim limitation (a ratio between maximum plasma concentration and concentration 24 hours after dosing) lacked written description support even though the ratio could be calculated from data in the disclosed examples, because the specification never identified or directed the reader to that ratio as a feature of the invention.
  • Novozymes A/S v. DuPont Nutrition Biosciences APS, 723 F.3d 1336 (Fed. Cir. 2013): although the original application provided “formal textual support for each individual limitation,” the court, “[t]aking each claim … as an integrated whole rather than as a collection of independent limitations,” found no disclosed species meeting the claimed combination and no “blaze marks” leading to one. It rejected the attempt to build support from “an amalgam of disclosures plucked” from different parts of the application.

Novozymes is the most direct warning for anyone who expects to assemble claims later from separately disclosed options. The lesson of all three cases is the same: if a particular combination matters, the original disclosure should point to it.

The flip side is equally well established: claims filed with an application are part of its disclosure. MPEP § 2163 explains that “a satisfactory description may be found in originally-filed claims or any other portion of the originally-filed specification” (citing In re Koller, 613 F.2d 819 (C.C.P.A. 1980); In re Gardner, 475 F.2d 1389 (C.C.P.A. 1973); and In re Wertheim, 541 F.2d 257 (C.C.P.A. 1976)), and that an applicant may amend the specification to include subject matter found only in an original claim (In re Benno, 768 F.2d 1340 (Fed. Cir. 1985)). A claim that expressly ties features together is exactly the kind of “blaze mark” the cases above found missing. It is not a guarantee: MPEP § 2163 cautions that not every originally filed claim is adequately supported, particularly broad functional claims, and the underlying technical disclosure must still describe and enable what is claimed.

Now consider what happens in the twelve months after a provisional is filed. The product evolves. A competitor launches something close. Your own team demos the technology, publishes a paper, or open-sources part of the codebase. Then, during prosecution, the examiner cites art that forces you to narrow to a specific combination of features. If that combination is not supported in the provisional, intervening art, including your own disclosures, may now be prior art against the narrowed claim.

Why multiple dependencies are the right tool

A claim set with multiple dependencies is, in effect, a compact and explicit map of combinations. “The system of any one of the preceding claims, wherein…” tells the reader in plain terms that this feature is intended to be combined with each of the earlier claimed embodiments. Chain those claims together and a few dozen claims can expressly describe hundreds of feature combinations, something no reasonable amount of prose could do as clearly.

One technical point matters here. A multiple dependent claim does not combine all of the claims it refers to at once. The USPTO treats it “in the same manner as a plurality of single dependent claims,” each containing only the limitations of the particular claim referred to for that embodiment (MPEP § 608.01(n)). In other words, each dependency path is its own embodiment. Build the claim tree so that the combinations you care about actually exist as paths.

The provisional is the ideal place to do this:

  • No claim formalities apply. A provisional may be filed “with or without claims” (35 U.S.C. § 111(b)(2); 37 C.F.R. § 1.53(c)), and provisionals are not examined for patentability (MPEP § 201.04), so the § 112(e) restriction on multiple-multiple dependencies does not constrain what you file.
  • No claim fees. A provisional pays only a basic filing fee (37 C.F.R. § 1.16(d)). The excess-claim and multiple dependent claim fees expressly apply to applications “other than a provisional application” (37 C.F.R. § 1.16(h)–(j)), so a large, richly cross-referenced claim set costs nothing extra in USPTO fees.
  • Everything in the provisional is disclosure. The claims are part of the provisional’s written description and can be relied on for support just like the specification and drawings (see MPEP §§ 211.05, 608.01(l)).

The international angle

A U.S. provisional usually serves as the priority document for foreign and PCT filings within the twelve-month Paris Convention period. Many foreign offices are stricter than the USPTO about combinations. The European Patent Office, for example, grants priority only for subject matter the priority document discloses “directly and unambiguously” (G 2/98), and applies a similarly demanding standard to amendments under Article 123(2) EPC. An amendment that combines features from different embodiments without a clear basis is a classic “intermediate generalization” problem.

The EPO permits multiple dependent claims, including multiple-multiple dependencies, and the dependencies in an originally filed claim set are commonly relied on as a basis for combining the features they link. A provisional with a well-built multiple-dependent claim tree therefore gives your European and other foreign counsel a much stronger record to work from. One caution: some EPO boards have been skeptical that very long, mechanical chains of “any preceding claim” dependencies disclose every theoretically possible combination. The claim tree should reflect combinations you actually intend, not just a formula.

Why this matters even more for AI inventions

AI inventions are stacks of interacting layers: data ingestion and preprocessing, feature engineering, model architecture, training and fine-tuning, inference optimization, deployment, and feedback loops. The commercially important and legally defensible invention is often a particular combination across those layers, not any single layer.

This matters most under 35 U.S.C. § 101 (see MPEP § 2106). After decisions like Recentive Analytics, Inc. v. Fox Corp. (Fed. Cir. 2025), which held that applying generic machine learning to a new data environment is not patent-eligible, AI applicants frequently need to amend toward the specific technical implementation that delivers a concrete improvement. In practice, that usually means combining several dependent features into the independent claim. If the provisional’s claim tree already expressly links those features, the amendment is supported and keeps the original priority date.

Here is a simplified example of how a provisional claim set for a retrieval-augmented generation (RAG) system might be structured:

1. A computer-implemented method comprising: receiving a query;
   retrieving, from a document index, a set of candidate passages;
   and generating, by a language model, a response conditioned on
   the query and at least a subset of the candidate passages.

2. The method of claim 1, wherein the document index is a
   hierarchical index mapping parts of a technical document to
   features of those parts.

3. The method of claim 1 or claim 2, further comprising reranking
   the candidate passages using a cross-encoder model before
   generating the response.

4. The method of any one of claims 1-3, wherein the language model
   is fine-tuned on domain-specific question-answer pairs.

5. The method of any one of claims 1-4, further comprising
   verifying each statement in the response against the retrieved
   passages and suppressing unsupported statements.

6. The method of any one of claims 1-5, wherein the language model
   is quantized and executed on an edge device.

7. A system comprising one or more processors and memory storing
   instructions that, when executed, cause the one or more
   processors to perform the method of any one of claims 1-6.

8. A non-transitory computer-readable medium storing instructions
   that, when executed, perform the method of any one of claims 1-6.

Eight short claims, and the provisional now expressly ties together every combination of hierarchical indexing, reranking, domain fine-tuning, output verification, and edge deployment, in method, system, and media form. Because each dependent claim refers back to every earlier claim, every subset of those features is reachable as a dependency path from claim 1. Claims 4 through 8 are all multiple-multiple dependent claims that could not be filed this way in a U.S. non-provisional, and they don’t need to be. Their job is to establish support.

Practical drafting tips

  1. Use “any one of the preceding claims” liberally for features that genuinely combine with everything before them, and narrower ranges (“any one of claims 3–5”) where only some combinations make technical sense.
  2. Keep each dependent claim to a single feature. Modular claims combine cleanly. A dependent claim that bundles three features only supports those three features together.
  3. Cover every statutory category. Use system and medium claims that incorporate the method claims by multiple dependency (“configured to perform the method of any one of claims 1–15”), so combinations are supported in every claim format.
  4. Nest fallback positions. Add progressively narrower ranges, specific alternatives, and preferred implementations as their own dependent claims rather than burying them in the specification.
  5. Don’t rely on claims alone. Describe your key combinations in the detailed description and figures as well. Explain why they work together and what technical improvement they produce. Claims establish that a combination exists; the specification shows how and why it works, which matters for both enablement and § 101.
  6. Avoid boilerplate as your only bridge. Generic statements such as “any feature may be combined with any other feature” are weak evidence of possession of a specific combination. Explicit claim dependencies are far stronger.
  7. Carry the combinations forward when you convert. Don’t let the disclosure shrink along with the claim set. See the next section.

Carry the combinations into the non-provisional

A provisional’s claim tree only does its job if the non-provisional preserves what it discloses. When you file the non-provisional, its own specification, claims, and drawings as filed become the “original disclosure” against which every later amendment, continuation, and reissue is measured. Claiming the benefit of the provisional does not automatically import the provisional’s content. Incorporating it by reference is not a reliable fix either, because material needed to support a claim (“essential material”) may be incorporated only from a U.S. patent or published application (37 C.F.R. § 1.57(d); MPEP § 608.01(p)).

The practical answer: file the reduced, fee-efficient claim set you want examined, but carry the provisional’s important combinations into the non-provisional specification, for example as an “Illustrative Combinations” section that spells out each meaningful dependency path as an embodiment. Specification text is not subject to the multiple-dependency prohibition or claim fees, so the full combination map survives at no extra cost.

We cover this in detail, along with timing, priority, claim-fee economics, and the other decisions to make when converting, in From Provisional to Non-Provisional: Preserving Support and Priority in the U.S.

The bottom line

A provisional application is cheap insurance, but only for the combinations it actually supports. Spending a few extra hours building a multiple-dependent, and where useful multiple-multiple-dependent, claim tree turns that insurance into broad, explicit coverage of how your invention’s features fit together. When prior art, an examiner, or a competitor forces you to narrow, you will be able to choose your combination without giving up your priority date.

For AI and software inventions, where defensible claims almost always come down to specific technical combinations, this is one of the highest-value steps you can take at the provisional stage.