From shelf to verdict
Most idea generators sample an AI's imagination and get the average of the internet back. The Forge works the other way: it mines mechanisms — first-principles causal structures — out of real books, breeds the ones that have never met, and makes the offspring survive a tournament. Every idea carries its parents' names.
A book is not a topic; it's an argument about why something works. The Forge reads each book's distilled summary and extracts that engine in transferable form: a named mechanism, a one-sentence principle, its preconditions, its failure modes — and a verbatim quote from the text as evidence. No mechanism enters the pool without a citation.
Freshness matters most here. The library's acquisition loop keeps buying what's new, so the newest shelves contribute the newest mechanisms — the scarcest input any ideation system can have. The pool below was mined from books published 2024–2026, some of them ingested the same day.
Organizations route decisions into structures where no individual can be blamed — and therefore nothing can be corrected.
— The Unaccountability Machine, Dan Davies
A resident population can steer its host's behavior toward what feeds the population, not what serves the host.
— The Gut-Brain Paradox, Steven Gundry
People reliably trade effort for visible position in a hierarchy they care about; status only ratchets one way.
— This Is Strategy, Seth Godin
The library's idea graph knows which books already talk to each other — 38,410 edges of similarity. The Forge breeds only pairs with no edge and different genres: mechanisms from corners of the library that have never met. That's where novelty lives — not in the model's temperature, but in the structure of the combination. A health mechanism never crosses another health mechanism; it emigrates into software, strategy, or education.
A diagnostic engine that treats an AI agent's codebase like a gut microbiome: it identifies parasitic patterns — dependencies and feedback loops that steer the system toward what feeds them rather than what serves the user — and prunes them before they compound.
Each cross is instantiated against a rotating audience-and-surface matrix and must pass hard founder constraints before it's even written down:
Forged ideas are exported into an evaluation pipeline: independent judges score them, hard-constraint gates reject what a two-person team can't ship, and a Swiss tournament ranks the survivors. The first live batch — five ideas forged from that day's freshly ingested books — finished like this:
| # | idea | parents | judge avg |
|---|---|---|---|
| 1 | micro-agent-optimizer | gut microbiome × accelerating returns | 4.13 |
| 2 | resilience-parallel-trainer | antifragile loading × parallel computing | 3.50 |
| 3 | nexus-pulse-editor | identity salience × systemic management | 3.50 |
| 4 | calm-signal-cert | 90-second reset × status ratchet | 3.38 |
| — | benchmark: best generic idea, same judges | no book mechanisms | 6.92 |
The judges' verdict was consistent — and useful: "grounded in analogy rather than validated user pain." Maximum novelty arrived with minimum evidence, exactly as theory predicts. The Forge doesn't produce winners; it produces hypotheses too original to have evidence yet. So the pipeline grew a stage: each forged idea's pain-hypothesis now goes to community evidence mining — real complaints from Reddit, Hacker News, app reviews — before it faces the judges. Novelty from the books, validity from the world.
The whole machine runs on pocket lint. Mining a book costs about a tenth of a cent; a forged idea, two-tenths; a full judged tournament, a nickel. The expensive part of having ideas was never the ideas — it was the months spent building the wrong one. That's what the tournament and the evidence stage are for.