aahplexx/perfecteval
Provider-agnostic evaluation and review workflows for proving work is done, finding regressions, comparing candidates, diagnosing failures, and maintaining durable quality gates across models, agents, prompts, products, and repositories.
Use when Perfect Eval must independently review a change, artifact, prompt, workflow, or repository result for real failures that the planned evals may have missed.
Use when Perfect Eval must test a user-facing experience from the perspective of someone who did not see it being built and should not rely on implementation knowledge.
Use when Perfect Eval has a reproducible failure, regression, incident, or review finding and must identify root cause, prove the diagnosis, apply an authorized fix, and prevent recurrence.
Use when Perfect Eval identifies repeated false negatives, false positives, review misses, or recurring defect classes and should convert them into concise durable review guidance without overfitting.
Use when a user wants to design, run, compare, diagnose, validate, benchmark, review, or maintain evaluations for an AI model, agent, prompt, workflow, application, feature, bug fix, or repository-backed system; especially for proving work is done, preventing regressions, comparing providers/models, or autonomously iterating until evidence supports completion.
Use when Perfect Eval must iteratively build or repair a target, verify every eval, independently review for missed failures, and converge without infinite agent-review loops.
Use when Perfect Eval must classify the consequence of failure before choosing evaluation depth, review effort, or human approval requirements.
Use when Perfect Eval needs to turn requirements, bugs, product behavior, or repository evidence into concrete capability and regression evals before implementation or comparison work proceeds.