Discombobulator · legacy → sovereign · re-grow not copy
Drop a decision log — a CSV of the inputs your old system saw and the decision it made (approve/refer/decline, pass/fail, any categories). The forge re-grows a sovereign model that reproduces it, proves how closely on held-out rows it never saw, surfaces every mismatch, and hands you a model you own plus a signed migration receipt. Re-grow, don't copy — a byte-copy reproduces the old flaws.
🔒 100% in your browser — your log never leaves this page, never uploads anywhere.
inputs + the decision column — everything runs locally, nothing is sent
Columns typed automatically. Pick which column is the legacy decision to reproduce — the rest become features.
| column | type | distinct | role |
|---|
The re-grown model vs the legacy decision, on a third of the rows it never saw during induction. The gate IS the safety: every mismatch is surfaced, never hidden.
Sample of the surfaced mismatches (legacy → re-grown) — exactly where a human should look:
| row | legacy said | re-grown said |
|---|
A self-contained model that reproduces the decision, gated by a signed migration receipt. Runs anywhere, no dependencies, no per-call cost. You own it.
Pre-registered in relearn.predictions.json before measurement; checked live now, and re-derived on GitHub's runner in CI.
A byte-identical copy of a legacy system reproduces its undocumented flaws and locks you into the same black box. Re-growing extracts what the system does as rules that survive held-out data, shows you exactly where it can't reproduce the old behavior (so a human confirms the tribal rules in minutes), and leaves you with a small model you own and can re-verify — not another vendor dependency. Nothing uploads; this page has no server.