The check failed. Why did the AI remember the answer?

Описание и применение

При ошибке проверки ответа ИИ старый код сохранял ответ и передавал его в следующий запрос, что приводило к «памяти» неверной информации. Новая версия исправляет это, не повторяя ответ, если проверка не завершена.

Задача

When a check on an AI assistant’s answer fails, the answer may still be stored and sent to the next model call, causing the assistant to “remember” an unverified response.

Результат

The patch ensures that answers failing validation are not automatically included in future model calls, preventing the assistant from using unverified information as context.

Рабочий процесс

• 1. Configure a check that can pass, reject, or throw.
• 2. Run the assistant with a test model that returns a distinctive marker text.
• 3. Observe the conversation history sent to the model after the check outcome.
• 4. Compare old logic (replays answer on error) vs. fixed logic (does not replay).

Доказательства

In our controlled comparison, the old logic raised an error and saved the answer anyway. The next model call received it.

Источник

The fixed logic preserves that distinction. It does not declare the answer disproved, and it does not automatically admit the answer as context for later work.

Источник