Methodology

How Continuous Correctness Audit turns review into evidence.

Proof scopes one critical path, compares intended behavior against implementation and tests, produces reproducible evidence, and verifies fixes as the codebase changes.

BoundaryPublic safe
Discussed

MC/DC, formal verification, reproducers, traceability, coverage, cadence.

Protected

The proprietary audit engine, orchestration, scoring, and private corpora.

Workflow

The public methodology.

01

Scope the path

Pick the parser, router, policy, SDK, agent loop, feature, or release window where failure would matter.

02

Recover intent

Map what the system is supposed to do from requirements, docs, tickets, tests, code, and product claims.

03

Find drift

Identify where implementation, tests, documentation, and expected behavior no longer agree.

04

Produce evidence

Confirm findings with reproducers, regression tests, expanded coverage, MC/DC, or formal artifacts where applicable.

05

Verify fixes

Retest patches and rerun the evidence corpus on the agreed cadence.

API and infrastructure lenses

Proof looks at the failure modes ordinary review misses.

Spec conformance

Protocol, schema, validation, normalization, and version-compatibility behavior.

Parser and validator edges

Malformed input, ambiguous syntax, size bounds, error handling, and fail-closed behavior.

Routing and planning

Query planning, federation, subgraph behavior, policy routing, retries, and fallback paths.

Concurrency and cancellation

Streaming, subscriptions, cancellation, timeouts, shared state, and resource cleanup.

Resource exhaustion

Inputs or request shapes that can amplify CPU, memory, recursion, fan-out, or downstream work.

SDK regression

Cross-version behavior changes that break users even when internal tests still pass.

Validation gates

What makes a finding deliverable.

GateWhat Proof checks
ReproducibilityThere is a runnable reproducer, test, fixture, replay, trace, or a clearly bounded explanation for why full reproduction is not possible.
SeverityImpact, exploitability, preconditions, required privilege, blast radius, data/security boundary, DoS potential, regression likelihood, confidence, and detection difficulty are considered.
Noise controlFindings are manually reviewed, duplicates are removed, affected surfaces are confirmed, and unvalidated leads are not presented as confirmed issues.
Patch verificationFixes are retested against the original reproducer and adjacent regression cases where feasible.
DisclosureCustomer findings remain private unless publication, advisory language, or field-note reuse is approved.

MC/DC and formal artifacts

Used where they raise confidence.

MC/DC and formal artifacts are used when the reviewed path has decision logic, state transitions, or invariants where ordinary tests do not give enough evidence. They are not claimed for every component by default.

Evidence handoff

The buyer should not depend on Proof to believe the report.

  • Delivered findings include affected surface, severity rationale, reproduction notes, and retest status.
  • Regression tests and reproducers should be inspectable and portable where practical.
  • Proof's internal orchestration remains proprietary; the delivered evidence must still stand on its own.

What this proves

Scoped evidence, not total correctness theater.

  • A finding is reproducible or explicitly bounded by the available evidence.
  • The affected surface, severity, and remediation path are clear.
  • Regression risk is reduced because the evidence can run again.

What it does not prove

No blanket claim that all software is correct.

  • Proof does not claim full-system total correctness.
  • Unreviewed components stay out of scope.
  • Customer-specific findings stay private unless approved for publication.