reqproof.com continuous correctness audit
Your promises, crystallized.
Your code makes promises no one wrote down. The audit recovers them, formalizes them into checkable requirements, and holds every change to them. What it learns, it keeps — forever.
01 the corpus
The deliverable isn't a report. It's a corpus that compounds.
On jsonparser, a widely used Go JSON library, the audit condensed the codebase into 123 approved requirements — each written in FRETish (NASA FRET lineage), each checked for realizability and vacuity by Kind2 on Z3, each traced to the code and tests that satisfy it.
When Set() writes a value at any keypath, the parser shall not drop or truncate any sibling data.
02 classes close forever
A bug is never one bug. It's a class — and classes close.
CLOSEDthe panic class
The audit surfaced an unchecked access that could panic — the same shape at 8 sites. Formalized as a requirement, all 8 closed the same day. The check now holds every future change to the same bar.
SWEPTthe one that escaped
A silent data-loss defect in Set() escaped with 100% MC/DC — the tests passed the wrong behavior. Caught downstream: postmortem public, requirement strengthened, the whole defect family swept. Misses become evidence too.
03 agents + humans, one loop
Trust is the bottleneck. Evidence is the fix.
Stack Overflow 2025: 46% of developers distrust AI-generated code, 33% trust it. DORA finds ~90% adoption against ~30% trust. The gap closes with evidence, not vibes.
Proof holds both contributors to one bar: agents get machine-checkable obligations; humans review and sign with named signatures; the same audit gates both. Nothing merges on trust alone.
> implement against SW-REQ-089; obligations attached > run the audit — fail on any warning > attach verifier evidence to the diff
signed: named reviewer · human
· the register
Every finding carries a name.
The register is public.
field conditions
04 the engagement
Fixed fee. One component. Four weeks. Then it never stops.
- Pick one critical component — the one you'd least like to be wrong.
- ~4 weeks to an approved requirements corpus with evidence behind every claim.
- Then continuous — the audit runs on every change, and keeps what it learns.