Pattern organs · bring your own data · sovereign
Drop any spreadsheet with a yes/no column — churned, converted, defaulted, passed — and the forge finds the patterns that actually predict it on held-out rows it never saw, then mints the winner as a tiny model you download and own. The alternative to renting an LLM to read your data.
🔒 100% in your browser — your CSV never leaves this page, never uploads anywhere.
or choose a file — everything runs locally, nothing is sent
Columns are typed automatically. Pick which yes/no column to predict — the rest become features (numbers pass through; categories become yes/no indicators; IDs are ignored).
| column | type | distinct | role |
|---|
Every pattern graded on the held-out third of rows (balanced accuracy — never an LLM judge). keep = generalised; drop = failed on rows it never saw.
| pattern | train BA | held-out BA | verdict |
|---|
Breeding ANDs patterns together, selecting on train, judged on held-out. We report what actually happened.
The validated classifier, minted as a tiny model gated by fallforgemint's rule (mints only on a certified beats-the-base receipt). Download a runnable owned-model.mjs you run yourself forever — plus a signed, re-checkable receipt.
On the bundled sample, pre-registered in predictions.json before measurement and checked live now. The same checks run on GitHub's runner in CI.
Embedding every row and asking an LLM (RAG) means a per-call bill, your data leaving the building, and an answer you can't re-check. CSV Forge keeps only rules that survive data they never saw, hands you a tiny model you own and run for nothing, and signs a receipt anyone can re-verify. Nothing uploads — this page has no server.