Proof
DOC NO. PRF-EXP-3  ·  SHEET 1 OF 1  ·  REQPROOF.COM
DocumentCONTINUOUS CORRECTNESS AUDIT — PUBLIC PROSPECTUS
RevisionEXP-3
Date of issue2026-08
IssuerPROOF (REQPROOF.COM)
ClassificationPUBLIC
StatusIN FORCE
52Checks
4Evidence classes
11Languages
0Unverified claims

Software, held to what was promised.

Statement of purpose · §0.1

Security firms audit whether your software can be broken into. We audit whether it works.

Recital 1 · Stack Overflow Developer Survey, 2025
46%Distrust AI-generated code
33%Trust it
Recital 2 · DORA, 2025
90%Have adopted AI tooling
30%Report little or no trust
Now, therefore, verification is issued on the following terms.

Composition of one audit

Ref. label, right

A Continuous Correctness Audit is a standing, machine-checked statement of what one component of your software is required to do, and evidence — re-runnable evidence — that it does it.

Every quantity on the label opposite is taken from a real, publicly inspectable audit. Nothing on it is estimated, projected, or asserted by its author without a check behind it. Where the method failed, the failure is on the label too.

Disclosure The published miss (1) escaped under 100% MC/DC coverage. Coverage is a floor, not a claim. See §03.

The Gate

Exhibit A

The audit runs in your CI, on every change. When a requirement cannot be shown to hold, the pipeline exits non-zero and the change does not merge. There is no dashboard to ignore and no report to file away. The gate is the deliverable.

Exhibit A — Gate output, CI terminal capture Representative
$ proof audit --fail-level warn
continuous correctness audit · 52 checks · 4 evidence classes

  requirement_approval ............. PASS   123 approved
  formal_proof / kind2 ............. PASS
  formal_proof / z3 ................ PASS
  coverage / mcdc .................. PASS   100.0%
  hazard_analysis .................. PASS   per requirement
  finding_register ................. FAIL   1 open

FINDING  F-XXXX  requirement not held under boundary input
         reproducer: findings/F-XXXX/repro.sh  ·  signed: named reviewer

exit status 1

The Evidence

Exhibit B

Four classes of evidence are admitted: formal proof, condition-level coverage, runnable reproducers, and named signatures. Author-asserted metadata is not evidence and is not admitted. Everything below re-runs from the repository.

Exhibit B — Evidence tree, repository layout Re-runnable
evidence/
├── requirements/            123 approved · FRETish
├── proofs/
│   ├── kind2/               model-checked
│   └── z3/                  SMT-solved
├── coverage/
│   └── mcdc/                condition-level, avionics bar
├── hazards/                 per-requirement analysis
├── findings/                7 · reproducers attached
├── misses/
│   └── 0001-set-silent-loss/postmortem — published
└── signatures/              named reviewer, per finding

The Public Audit — jsonparser

Register

We audited buger/jsonparser, a widely used Go JSON library, in public. The full register — requirements, findings, and the miss — is open for independent inspection at portal.reqproof.com/projects/jsonparser.

Findings register — summary
RefItemQtyDisposition
REQ Requirements, approved 123 IN FORCE
F-ALL Findings, confirmed real bugs 7 REPRODUCERS ATTACHED
F-PANIC   of which: one panic class 8 sites CONFIRMED
MISS-001 Silent data loss in Set() — escaped under 100% MC/DC 1 PUBLISHED · RE-RUNNABLE
On MISS-001 The one defect our method did not catch is documented with the same rigor as the seven it did: postmortem published, reproduction re-runnable, register public. A method that cannot show you its misses is asking for faith. Faith is not an admitted evidence class.

The Method

Parts list

Standard instruments. Aviation-grade lineage. Applied to ordinary software, which is where the money actually is.

Part No.DesignationFunctionOrigin / Note
MTH-01KIND2 Model checker — formal verification of requirements Proves, not tests
MTH-02Z3 SMT solver — formal verification of requirements Proves, not tests
MTH-03FRETISH Structured requirement language NASA flight-software lineage
MTH-04MC/DC Condition-level coverage measurement Avionics certification bar · 11 languages incl. Go, Rust, Solidity, Zig
MTH-05HAZARD ANALYSIS What happens when each requirement fails Performed per requirement

The Engagement

Article 7

Terms are short because the deliverable is not a document. It is a gate that keeps running after we leave.

7.1
Fee. The fee is fixed. It does not float with hours, headcount, or enthusiasm.
7.2
Scope. One component. The one whose failure you least want to explain.
7.3
Schedule. Approximately four (4) weeks to install. Thereafter the audit stands in your CI, continuously, on every change.
7.4
Findings. Every finding ships with a runnable reproducer and the signature of a named reviewer. See §06.

The Signature

Attestation

No finding in this practice is issued anonymously. A named reviewer signs every finding and answers for it — by name, not by brand. If we are wrong, you know exactly who was wrong.

Request for Scoping

Form PRF-1

Complete the form below. Submission constitutes a request for scoping and nothing else.

Filed requests are read by a person. Response is a scoping conversation, not a sequence.
Received. Your request has been entered into the register.
A named human will respond.