Reviews & Ghost Evaluation

What it is

The quality layer:

  • Reviews — human / agent approval gates over changes.
  • The Review Engine — an automated, domain-by-domain analyzer.
  • Ghost evaluation — offline A/B testing of prompts and models against shadow candidates.

Operator surface

  • Reviews — review_create (CodeReview, DesignReview, SecurityReview, ComplianceReview, TaskReview, General), review_list, review_get, review_pending, review_assign, review_approve, review_reject, review_cancel, review_report_generate.
  • Review Engine — review_engine_start, review_engine_domains, review_engine_status, review_engine_findings, review_engine_run_sweep, review_engine_full_system, plus schedule and submission tools; and UX review (ux_review_*).
  • Ghost — ghost_evaluation_list / _get / _stats, ghost_evaluation_trigger, ghost_evaluate_from_history, ghost_promotion_candidates, ghost_prompt_history.

REST: ReviewsController, ReviewEngineController, UxReviewController, GhostEvaluationController.

How it works

Reviews persist as ReviewRecords tied to the entity under review and can block mission completion at a constraint gate. The Review Engine runs a ReviewEngineSession that enumerates project targets and produces ReviewEngineFindings via per-target LLM analysis, each domain routing through its ReviewEngine:{Domain} site preset (with schedules for recurring sweeps). Ghost evaluation replays real or synthetic prompts against shadow models, scoring them (GhostEvaluation, PromptEvolution, LearningPrompt) so better prompts / models can be surfaced as promotion candidates without touching live traffic.