Feed it a legacy system. It works out what the software does (not how it's coded), re-grows a sovereign build that does the same job, proves equivalence on the system's own held-out cases, and migrates the data across. Not a copier — a re-grower: functionally equivalent, structurally superior, owned by you. Everything below runs live in your browser on the real gated kernels.
Proof target: LoanDesk 2011, an undocumented 1,200-record credit-decision engine.
▶ Bring your OWN legacy system. This page proves the method on a sealed demo.
To run it on your own decision log — drop a CSV, re-grow a replacement, download a model you own —
open the live tool → (100% in your browser, nothing uploads).
STAGE 1
Ingest
Inventory the code, config, schema & data; sample the legacy as a black-box oracle.
STAGE 2
Understand
Induce the decision function as two grown organs; flag low-confidence patterns.
STAGE 3
Re-grow
Grow a sovereign, SENTINEL-defended, compressed build from the spec.
STAGE 4
Verify-equivalent
Run the real held-out cases through old vs new; surface every mismatch.
equivalence on the legacy's own held-out cases — higher is more faithful
What didn't come across — flagged, never hidden
id
legacy said
re-grown said
the gate's attribution
The gate returns every case where old and new differ and attributes each to a candidate tribal rule. That short, pinpointed list is what turns blind reverse-engineering into a minutes-long human confirm — the hard-20% made visible, not silently lost.
Structurally superior, not just equivalent
The legacy accepts any call that reaches it. The re-grown build routes every decision as a signed, κ-witnessed SENTINEL packet (Thomas Frumkin's primorial-fold codec) and rejects what the legacy could never see. Live κ-witness check:
—
untampered packet
—
one byte flipped
What v1 is — honestly
v1 is a guided pipeline, not a one-click vending machine. The common rules extract automatically; the gnarly, undocumented ones are flagged for a human to confirm in a moment. Some things don't translate at all (deep platform deps, hardware-bound logic) and would be bridged, not re-grown — flagged up front. Automation deepens as the pattern library grows. The un-forgeable proof is the CI re-run on GitHub's own runner, re-derived from the sealed inputs; this page is the live demonstration.