PatentFactory — a working invention-council process, run across ten industries, that has filed zero patents

An earlier version of this report said PatentFactory's templates were built and zero scan or brainstorm sessions had ever been run. That was wrong. The project's own tracker shows thirteen sessions did run, and produced real, countable output. This is the corrected account — what actually happened, how, and what still hasn't.

What it was

PatentFactory (/patentfactory/) is a systematic invention process: assemble 10–15 simulated experts from different fields per session, plus 2–3 cross-domain wildcards, a patent attorney, and a skeptic, and have them debate a problem until they land on cross-domain intersections — novel by construction, because no single-field expert would reach them alone. The design covers 25 target industries and keeps a named human as inventor of record throughout, following the real U.S. legal position that an AI cannot itself be a named inventor (Thaler v. Vidal) — AI is the tool inside the process, not the inventor.

The process runs in seven stages: SCAN (rank the top problems in an industry), BRAINSTORM (the expert-council debate), ASSESS (prior-art search, a 1–10 novelty score, a non-obviousness and utility check, a market estimate), DEEP DIVE (technical spec + draft claims for the strongest ideas), PROVE (a mathematical, simulation, or literature argument for ideas that don't need a lab), FILE (a drafted $65 micro-entity provisional), and MONETIZE (outreach and licensing) — tracked throughout in 00-management/PATENT-TRACKER.md.

What we built — and actually ran

🔬Research

By 4 April 2026, per the project's own tracker: 10 of the 25 planned industries had been scanned (pharma, batteries, education, climate/energy, medical devices, aerospace, automotive, chemical, biotech, finance), across 13 brainstorm sessions (10 industry sessions + 3 "moonshot" sessions — room-temperature superconductors, universal cancer detection, $100 desalination). That produced 75 ideas (66 industry + 9 moonshot), of which 66 were assessed (prior art + novelty score), 6 were deep-dived, and 6 provisional patents were drafted and marked ready for submission. Scale: 112 files generated. Fourteen ideas scored a 10/10 novelty rating — the "crown jewels" the tracker flags for priority filing.

Ideas were also run through a two-model computational validation pipeline: a local Qwen2.5-coder:14b model (via Ollama, on an RTX 3090) drafts a technical critique, Claude reviews it for errors — the tracker records 3 moonshot validations plus 1 Claude-vs-Qwen comparison, and 10 simulated expert-panel reviews (5 pharma + 5 battery). The actual Qwen output for the three moonshots (moonshots/VALIDATION.md) is exactly the kind of honest, hedge-heavy technical review you'd want — "probability of Tc > 200K: estimated at less than 1%" for the superconductor idea, for instance — not a rubber stamp.

What "filed" actually means here: nothing yet. Zero of the 75 ideas, zero of the 6 drafted provisionals, have been filed with the USPTO or any other patent office. Total investment to date: $0. Total revenue: $0.

What we learned — including the honest gaps

• The earlier report undersold this, and that's worth naming directly. A report that says "zero sessions run" when 13 sessions and 112 files exist is as much a failure of the honesty discipline as a report that oversells — it just fails in the other direction. The correction stands on the record next to the original claim, not instead of it.
• Idea generation at volume is real and repeatable. 75 assessed-or-better ideas from 13 sessions is a genuine throughput number, and the process documentation (session structure, scoring rubric, claim-drafting guide) is detailed enough that a fresh operator could run session 14 without reinventing the method.
• "Validated" needs a qualifier every time it's used. The pipeline here is a local coder-model draft plus a cloud-model review — useful for catching technical errors and inflated claims early, but it is not prior-art litigation search, not a granted patent examiner's opinion, and not a wet-lab result. None of the 75 ideas has cleared an independent bar higher than "two AI models and a scoring rubric agreed it looks strong."
• Filing is the actual bottleneck, and it's cheap. Six provisionals are drafted and ready; filing all six costs $390. The gap between "ready" and "filed" here is not technical — it's that the $65-per-idea step, the cheapest part of the entire pipeline, has not yet been taken.
• The monetization thesis is untested at every step. No outreach emails, no buyer calls, no signed option agreements are evidenced anywhere in the project's own records. The $1T/95%-take-up framing is a brainstorm, not a plan with any executed steps behind it.

Where it went / status

Active, partially executed, honestly incomplete. Tiered research rather than shipped, because the throughput claim (idea generation at volume) is demonstrated but the pipeline's own later stages (filing, monetization) are not — and tiered research rather than vision, because unlike a purely designed-but-unrun process, this one has real, dated, measured output behind it. The correction from the earlier "never run" framing is itself now part of the honest record.