Verification
Overview
Verification (검증, 檢證) is a procedure that confirms with objective evidence whether a claim, system, or output actually meets pre-agreed criteria or specifications. Scientific experiments, software testing, journalism fact-checking, and judicial evidence review all share the same structure—“work that narrows the gap between claim and fact”—differing only in object and means. The credibility of verification is often determined not by the conclusion itself but by the transparency, reproducibility, and falsifiability of the procedure.
Main Content
Distinction Between Verification and Validation
In software engineering, verification and validation are distinguished. Verification asks, “Are we building the product correctly?” while validation asks, “Are we building the correct product?” In other words, verification examines conformity to specifications, whereas validation examines suitability relative to user needs. This distinction applies almost identically in manufacturing quality control, medical device licensing, and regulatory compliance review.
Scientific Verification
The core of scientific verification is falsifiability and reproducibility. Karl Popper held that a claim with no possibility of being falsified cannot be the object of scientific verification. In modern science, preregistration, sample size calculation, multiple-comparison correction, independent replication, and peer review have become standard procedures. Conversely, p-value hacking (p-hacking), selective reporting, and publication bias are cited as representative failure types that undermine the credibility of verification.
Software Verification
Software verification is divided into static verification and dynamic verification. Static verification finds defects without execution, as in code review, static analysis, type checking, and formal verification (model checking, theorem proving). Dynamic verification confirms behavior through actual execution, as in unit, integration, and acceptance testing, fuzzing, and load testing. In practice, regression is suppressed by combining the test pyramid, code coverage, mutation testing, and CI/CD pipeline gates. Safety-critical systems (aviation, rail, medical) are required by standards such as DO-178C and IEC 61508 to document verification artifacts and traceability.
Data and AI Verification
In machine learning, separating training data and evaluation data is basic, and generalization performance is estimated through cross-validation and holdout validation. Recently, as concern grows that benchmark scores alone cannot represent real-world performance, checks for benchmark contamination, distribution shift validation, bias and harm evaluation, red-team testing, and model card disclosure are required together. For generative AI, verification of factuality (hallucination) and verification of source citation have been added as separate evaluation items.
Social Verification
Journalism fact-checking, academic journal peer review, judicial review of evidentiary admissibility, accounting audits, and public institution policy evaluation are verification devices at the level of social institutions. They differ from technical verification in that procedural legitimacy and management of conflicts of interest determine the credibility of results.
Recent Trends
Changes in 2024–2025 fall largely into four branches. First, regulation is making verification mandatory. The EU AI Act requires conformity assessment, risk management, technical documentation, and post-market monitoring for high-risk AI, and countries are also establishing AI verification and certification systems. Second, formal verification is spreading into industrial settings. Mathematical proof-based verification, as in TLA+, Coq/Lean, Rust’s type system, and the seL4 case, is being applied to kernels, distributed systems, and smart contracts. Third, supply chain security verification has become important. SBOM (software bill of materials), reproducible builds, and signing and provenance attestation have become standard practice, and frameworks such as SLSA have spread. Fourth, AI has become both a subject and an object of verification. LLM-based evaluation and verification agents are used for test generation and code review, while deepfake and synthetic content detection and provenance marking (C2PA) have emerged as new challenges for fact-checking. At the same time, responding to benchmark saturation and the reproducibility crisis, and securing transparency of open evaluation datasets, are being discussed as common agendas in academia and industry.
Related Topics
- [[Testing]]
- [[Fact-checking]]
- [[Formal verification]]
- [[Reproducibility]]
- [[Artificial intelligence]]