Software Is Free. The Moat Is The Machine.

Patent counsel in Europe bills between EUR 250 and EUR 600 an hour, and a full draft takes four to eight weeks. For a company filing 3 to 10 patents a year, the bottleneck is never their expertise. It is their calendar. So we changed the deal. We do the eighty per cent that is craft, the prior-art research, the 12,000-word specification, the drawings, the translation, and hand over a near-final document. Their twenty per cent is what only a qualified attorney can do: hardening the claim language, polishing it for the jurisdiction, executing the filing. I have built companies across twelve countries and signed documents I did not fully understand. This one I understood. The clock still runs. It just runs on their judgment now, not their typing.

Under 35 USC 101, with Alice and Mayo behind it and the 2024 AI eligibility guidance on top, an abstract idea is not patentable unless it is integrated into a practical application. In Europe, Article 52 of the EPC discards programs for computers as such. The COMVIK test, T 641/00, asks two questions. Is there a non-trivial technical problem. Is the solution technical rather than abstract. Answer no and nothing else matters. Not the quality of the code, not the 11 months you put into the model, not the customer who loves it. That is the whole game. The software is free to everyone the day you ship it. The patent, if you get one, has to be anchored to something with a serial number. A chip, a sensor, a register that measures the physical world and signs what it saw.

The research ran as 4 agents in parallel: closest prior art, adjacent approaches, the regulations pushing the technology, and the white space nobody had claimed. Six Boolean queries each, layered with nine CPC classes. Google Patents and Espacenet for the free pass, PatSnap for the semantic one. Ninety minutes later we had a 5,000-word validation memo that would have taken a human researcher one to two weeks. The finding that mattered was a combination. Per-inference granularity, a multi-PCR normative scheme, hash-chained receipts. No single reference disclosed all three together. Intel WO2021236551 had attestation, but at model load, not per inference. Palantir US10581613 had a hash chain, just not for machine learning. Novelty is not a thing you invent. It is a shape you find in the gap between what already exists.

Three independent claims, mirroring three kinds of infringer. The method claim catches whoever operates the service. The system claim catches whoever builds the device. The computer-readable medium claim catches whoever ships the software. Twenty-eight claims in total, three independent and twenty-five dependent, each layer narrowing step by step, so that when the examiner rejects the broad one there is a fallback still standing. The specification ran to 12,000 words across a fourteen-section template. Every figure described at least 200 words deep in the body. Ten drawings as true vectors, black lines only, 3.5-pixel weights on a 2048 canvas, reference numerals in parentheses, never crossing a line. I read every page twice. It is not literature. It is a fence, and a fence is built with a tape measure.

Novelty is not a thing you invent. It is a shape you find in the gap between what already exists.

PatSnap's automated novelty check labels the closest references. PS-X means one document can kill novelty on its own. PS-Y means two of them combined can kill the inventive step. PS-A means background, and background does not block you. Ours came back PS-A, so we filed without buying the EUR 500 to EUR 1,500 full search report. Earlier, a reference from IBM had surfaced with a claim sitting close to ours. We did not file around it and hope. We promoted the strongest pillar, the piece grounded in a hardware register, out of Claim 7 and into Claim 1, and made the cryptographic binding explicit about what the signature joins. The probability that someone reads your claims before building the same thing is not zero. Most people price it at zero. That is how the same idea gets filed twice, years apart, by two companies that never spoke.

The priority date is the filing date. Everything downstream hangs off that number, and it opens a 12-month window. By month 9 you choose the receiving office, by month 11 you file the PCT application, and roughly 3 months later the international search report lands. At 30 months the national phase opens, and there the arithmetic stops being theoretical. The first filing is cheap. EUR 1,260 at the EPO, USD 2,000 at the USPTO. Five countries at the far end, US, EP, JP, KR and CN, run EUR 25,000 to EUR 70,000, of which translation alone is EUR 7,500 to EUR 20,000. Two or three markets chosen deliberately beat five chosen hopefully. Attorney time to PCT came to 6 to 13 hours against 50 to 100 for a from-scratch draft. You do not own the software. You own the words the office grants you, wrapped around a thing that can be touched.