Writing a provisional that survives.

A provisional that survives is not a draft of the eventual PCT. It is a load-bearing instrument in its own right. It locks a priority date, carries the technical effect into the future, and decides which claims of the eventual PCT are anchorable. Most provisionals fail at litigation not because the inventor was wrong about the science, but because the document was structured for a single attack and folded when a different one came in. The doctrine below — calibrated against the post-2023 case law — is built so a single failure mode falling does not pull the rest down with it.

The cascade-4-couches

Every claim worth filing is filed four times. Layer 1 states a functional genus — the claim broad enough to make commercial sense. Layer 2 backs it with a structural Markush — the same idea, enumerated. Layer 3 tightens to preferred ranges with a fallback cascade. Layer 4 is the picture claim: a single embodiment with the exact ratio your model produced. The point of severability is that if Layer 1 falls under Amgen v. Sanofi for failing to enable the full functional scope, Layers 2–4 are structurally anchored to specific molecules and survive the same attack. The post-Amgen rule is brutal but legible: the more you claim, the more you must enable.

1. A composition comprising a polymer P and an active pharmaceutical
   ingredient X, wherein the composition exhibits an oral bioavailability
   enhancement of at least 2-fold relative to crystalline X.            ← genus FONCTIONNEL

2. The composition of claim 1, wherein P is selected from the group consisting
   of: PVPVA, HPMC, HPMCAS, Soluplus, PVP K30, Eudragit E PO, Eudragit L100,
   PEG 6000, poloxamer 407, and copolymers, blends, salts, and bioisosteres
   thereof.                                                              ← genus STRUCTUREL (Markush)

3. The composition of claim 2, wherein the polymer:API mass ratio is
   comprised between 30% and 70%, preferably 35–65%, more preferably 39–45%.
                                                                         ← RANGES préférés

4. The composition of claim 3, wherein P is PVPVA, X is indomethacin, and
   the polymer:API ratio is 39.84%:60.16%.                              ← PICTURE CLAIM (watermark)

The four layers are not interchangeable. A picture claim cannot replace a functional one — competitors design around picture claims in an afternoon. A functional claim cannot replace a picture one — juries and PHOSITAs need a concrete embodiment to anchor the teaching on. You file all four, in this order, and the prosecution history that you record while you do it becomes part of the doctrine of equivalents calculus when someone copies you.

The definitions section

The definitions section is the most under-rated weapon in a provisional. Post-G 1/24 (EBA EPO, 2025), the description and figures must systematically be taken into account when interpreting the claims, even when the wording looks clear. That alignment of the EPO with the UPC means that what you write in Definitions determines how broadly your claims are read at trial. Bioisosteres are listed there — methacrylic acid encompasses ethacrylic, crotonic, vinylphosphonic, tetrazole-containing analogues — because that listing forecloses the obvious chemical-tweak design-around. They stay out of the claims directly, because over-broad bioisosteric language inside a claim risks rendering it non-enable under Amgen; they appear inside a dependent claim and dominate the spec.

Plausibility statements (G 2/21)

The EPO Enlarged Board's G 2/21 (23 March 2023) confirmed that post-published evidence can be filed in support of a technical effect — but only if that effect is encompassed by the technical teaching of the application as filed and embodied by the same invention initially disclosed. In practice this means the Field of the Invention and the Summary must explicitly mention every technical effect you might want to invoke later. If you do not declare it now, you cannot recover it post-filing. The hedge phrasing is canonical: "The compositions of the present invention are believed to provide [enhanced oral bioavailability of low-solubility APIs]. Without wishing to be bound by theory, the inventors propose that this effect arises from [favourable Flory-Huggins χ parameter between polymer P and API X]." That sentence is doing real work. The Spanish Supreme Court applied G 2/21 for the first time in national jurisdiction on 24 April 2025, which means the doctrine now has supranational application force.

Watermark ratios

When you write polymer:API ratio = 39.84%:60.16% instead of the rounded 40:60, you are doing two things at once. You are stating the actual value your model converged on (so you are factually correct), and you are inserting a forensic signature. If a competitor's product comes to market at 39.84%, the statistical probability of independent rediscovery of that exact value is essentially zero. That is not a trap; it is the truth of your calculation. Legally it is fully admissible, and in court it functions as a near-irrefutable evidentiary signal. Apply this systematically to every numerical example.

Multiple paths to enablement

For each principal embodiment, provide at least three independent enablement paths: a detailed experimental protocol, a reference to a known method with explicit adaptation, and a parametric algorithm ("the skilled person can systematically vary ratio R from R₁ to R₂ by step S until property P is obtained"). The redundancy neutralises the In re Wands "undue experimentation" attack: if one path is challenged as requiring undue work, the others stand. Without three independent paths, a single examiner with a weekend's worth of objections can collapse the enablement of the whole functional genus.

Negative space

A "Comparative examples and negative results" section listing ratios and combinations that were tested and failed accomplishes three things at once. It establishes non-obviousness under KSR: a PHOSITA who would supposedly have arrived at the invention by ordinary skill should have tried these ratios and not succeeded. It forecloses a competitor reframing your failures as their successes. And it strengthens sufficiency: by demonstrating where the boundary conditions are, the spec proves it teaches the full operative envelope. The negatives are kept in the spec; they are never claimed.

Anti-circumvention

Each functional claim is doubled with a structural anchor. "A polymer characterised by Tg between 40–60 °C" sits next to "A polymer of formula [structure A], wherein…". The UPC Court of Appeal in Amgen v. Sanofi/Regeneron (UPC_CoA_528/2024, 25 Nov. 2025) aligned with the EPO and accepts broader functional features in EP, but SCOTUS still demands more under §112. Doubling neutralises the jurisdictional gap: the same composition is claimed in both languages, in the same document, on the same priority date.

AI-as-tool, never AI-as-inventor

The line is bright. Post-Thaler v. Vidal (43 F.4th 1207, Fed. Cir. 2022) and post-USPTO Revised Inventorship Guidance (Federal Register, 28 Nov. 2025), AI is treated as a sophisticated tool — laboratory equipment, software, a database. Inventors are named as the human persons who conceived the invention in the sense of Burroughs Wellcome v. Barr Labs: the formation of a definite and permanent idea of the complete and operative invention. The single-inventor AI-assisted case now applies the traditional conception test, not the Pannu factors; multi-inventor cases still run the Pannu analysis. The fatal trap is writing "the AI invented X autonomously" anywhere in the spec or prosecution. That sentence makes the invention non-attributable to a human and unrecoverable. You write "the inventors used a computational tool to generate candidate compositions, which they then evaluated and selected".

Rolling provisionals

One provisional is a snapshot. A series of rolling provisionals is a moving picture. At each major discovery — a new polymer, a new ratio, a new indication — you file a fresh provisional (26 €, or 13 € at the reduced INPI tariff). They never get formally consolidated into the prior one (forbidden under Décret 2020-15), but at the PCT they combine via the multi-priority mechanism of Art. 8 PCT and Art. 88 EPC, with each claim anchored to the earliest provisional that fully discloses it. That priority-↔-claims matrix is the single most important deliverable at month 11, and the document that makes the PCT survive a partial-priority attack under G 1/15.

What you ship

A provisional that survives looks plain on the page. Title in Garamond, neutral and broad. A one-paragraph Field. A Background citing five to ten references with no extrapolation. A one-page Summary mirroring the claims. A Detailed Description that reads like a fortress floor plan. Examples that are reproducible. The four-layer claim cascade. An Abstract under 150 words. Then the PDF/A is generated, watermarked on the cover, and uploaded to the INPI portal. The 26 € fee posts. The priority date is locked. From there, the work is just keeping the document defensible for twelve months, and recording every disclosure, every collaboration, every prompt you fed to the model, against the day someone reads the spec at litigation and asks how you got there.