Two Projects, One Problem — What PlannerCritic and AdversarialDebate Each Got Wrong

Описание и применение

Обе системы, PlannerCritic и AdversarialDebate, пытались обнаружить ошибки LLM, но потерпели неудачу из‑за того, что доверенные детерминированные слои перестали срабатывать, делая метрики лучше, чем они были. Добавление простых проверок формы вывода (Gate Canary и Capitulation Detector) выявило эти скрытые сбои и восстановило надёжность.

Задача

Both systems aimed to detect when an LLM’s judgment is wrong but failed because the deterministic safety layers they trusted stopped firing silently, making metrics look better while safety slipped.

Результат

Adding lightweight, shape‑based checks uncovered silent failures in deterministic safety layers, restoring reliable safety metrics and preventing misleading improvements.

Рабочий процесс

• 1. Identify failure mode where deterministic layer stops firing silently.
• 2. Design a lightweight secondary check that inspects the shape of the output (gate canary or transcript analysis).
• 3. Implement the check as a CLI command or transcript parser.
• 4. Integrate the check into the CI/CD pipeline and production monitoring.
• 5. Validate the check by running fixture pairs or transcript samples and ensuring no false negatives.
• 6. Deploy the updated version (v0.2.3 for PlannerCritic, v0.2.2 for AdversarialDebate).
• 7. Monitor both LLM and deterministic layers for continued health.

Доказательства

The Gate Canary — ten fixture pairs, one per gate class, each with a known-good plan and a known-bad plan. The CLI command plancritic gates canary --check runs them in under a second at zero LLM cost and exits 1 if any gate stopped firing on its bad plan.

Источник

The capitulation detector — transcript-level inspection independent of convergence scores.

Источник

Both projects shipped their fixes in late August 2026 and validated in production.

Источник