Technical study/Evaluation · Reproducible study
Constraint Verifier Mutation Audit
A bounded, synthetic study of what different verifier layers catch — and what they miss — when checking optimization solutions.
What does a verifier actually verify?
A candidate optimization solution can be well-formed and still be wrong. It might violate a constraint, report an inconsistent objective, or be feasible without being optimal.
Constraint Verifier Mutation Audit examines those distinctions through a small, inspectable synthetic study. It was prepared with AI assistance under my direction and published as a portfolio artifact.
Four layers, different responsibilities
The study separates verification into four layers:
- Schema: is the candidate structurally well-formed?
- Feasibility: does it satisfy the task’s domain and constraints?
- Objective consistency: does the reported objective match the submitted decision?
- Optimality: does the candidate achieve the known optimum for this bounded task?
The diagram above summarizes these responsibilities. It is an explanatory illustration of the audit, not a production-system architecture.
A deliberately bounded experiment
The published artifact contains 120 bounded synthetic tasks and 1,080 candidate records. It applies declared mutations to examine what the verification layers detect. The repository includes task generation, evaluation, results, and a notebook.
This is a methodological demonstration, not an evaluation of a language model or agent. Its results do not establish global novelty, certification, general reliability, or production performance. The task space is small and synthetic, and the conclusions stay within that scope.
Inspect the work
The public repository contains the implementation, experimental conditions, and limitations. Public inspectability does not imply an open-source license; check the repository’s licensing terms before reuse.