Kaizen-3C the intent engine
Own your intent — the code is just a projection of it. Verbatim captures the intent behind your software as editable architecture, moves it to any stack, re-defines it when requirements change, and — the step almost everyone skips — verifies it stayed intact. Migration stops being a lossy rewrite and becomes a projection you can re-run.
Captured and verified, intent is IP you can move to any stack and defend to an auditor. Left in people's heads, it's the maintenance cost, the support burden, and the shrinking talent pool of code whose why no one holds — the road that ends in COBOL no one dares touch. Verbatim is the pivot: it turns intent from a liability you carry into an asset you own.
Capture intent as ADRs, SDDs, and specs — before the spend. The model executes a plan a human ratifies, not one it improvised.
Compile intent to the target you need. Migration is a projection of what you meant, not a rewrite from fading memory.
Requirements change → change the intent and regenerate. The why never goes stale, because it's the thing you edit.
Prove the intent survived: testing, fuzzing, a differential oracle, an adversarial envelope. The step almost everyone skips — and our edge.
On DARPA TRACTOR's C→Rust methodology, the most popular transpiler is 100% correct — and the least memory-safe Rust measured. Correct, and unsafe. Verbatim measured the safest of the field, because safety comes from the verification gate, not the model.
An independent self-benchmark against DARPA TRACTOR's published methodology — not a TRACTOR performer, not an official evaluation. Scorecard + harness →
One small C image decoder → safe-Rust and hardened-C, each proven byte-identical to the original across a valid + hostile corpus — in one command, make check. The whole thesis in miniature. github.com/Kaizen-3C/qoi-safe-c
Bug-for-bug fidelity over a defined envelope: the output is byte-for-byte equivalent to your reference over a valid and adversarial corpus, clean under memory sanitizers, and re-runnable on every commit. Not "a model judged it good" — a check anyone can re-run.
C→Rust is one vertical slice of what Verbatim does — the hardest publicly-defined test we could find. The same discipline hardens C in place, modernizes a framework, and re-defines a system when its requirements move. And it's built to make your experts more productive, not redundant: it captures what they know and scales their judgment across the codebase.
The full thesis, the TRACTOR results, and what we don't claim — a short read.