Why can I not reproduce the exact output my model gave in production?
機会
Even at temperature zero, the same prompt sent to the same model twice can return different tokens. A January 2026 arXiv paper traced the root cause to dynamic batching: GPU reduction kernels change their internal arithmetic tree when batch size changes, and floating-point addition is not associative, so two requests processed under different batch sizes take different numerical paths and diverge. Serving infrastructure has no obligation to record or expose the batch context that produced a given output. This makes it impossible to replay a failed inference, write a regression test that pins exact behavior, or reconstruct what a model actually said during a production incident. The only solution demonstrated so far imposes a 61 percent throughput cost.
重要な理由
Deterministic replay is the minimum bar for treating a model call as an auditable computation rather than a black box you trust but cannot inspect.
機会をどう評価するか
Opportunity Scoreは測定値ではなく、私自身の見解です。どれほど痛みを伴うか、どれほど頻繁に影響を与えるか、そして今日時点で解決策がいかに少ないか。スコアが高いほど、構築する価値が高いと私は考えています。
それが現れたときにどれほどの痛みをもたらすか。
実際にどれほど頻繁に人々がそれに直面するか。
今日時点で、それに対する優れたツールがいかに少ないか。