EXPERIMENT X · ALGORITHMS

该不该自己决定Decide or Defer

哪些判断应该交给模型，哪些应该回退给人或规则？Which judgments belong to the model, and which should fall back to a person or a rule?

调研中ResearchingJevJev置信度门控Confidence gating路由Routing
  1. Q提出可证伪的问题State a falsifiable question当前阶段Current stage
  2. R做最小运行路径Build a minimal runnable path未开始Not started
  3. E连接真实验证Validate against reality未开始Not started
  4. P公开证据与限制Publish evidence and limits未开始Not started

为什么值得做Why it is worth doing

意图识别、任务分流这类判断，用生成式模型做既慢又难控制。类型化的概率输出也许能把“要不要自动处理”变成一个可以设阈值的决定。Judgments such as intent recognition and task routing are slow and hard to control when done with a generative model. Typed probability outputs might turn "handle automatically or not" into a decision with a threshold.

Q

假设Hypothesis

HYPOTHESIS

类型化决策加置信度门控，在路由准确率接近生成式结构化输出的同时，显著降低时延与调用成本。Typed decisions plus confidence gating match generative structured output on routing accuracy while cutting latency and call cost significantly.

R

最小装置Minimal setup

MINIMAL SETUP
  1. 01调研类型化概率输出与 RLCDSurvey typed probability outputs and RLCD
  2. 02在同一批分流任务上对照：生成式结构化输出 / 类型化决策 + 门控Compare on the same routing tasks: generative structured output / typed decision + gating
  3. 03记录路由准确率、误判后的回退、时延和成本Record routing accuracy, fallback after misjudgment, latency, and cost
E

什么会证明我错了What would prove me wrong

FALSIFIED IF

类型化决策的误判率明显更高，且无法通过门控回退弥补。If typed decisions misjudge clearly more often, and gating cannot make up for it, the approach does not hold.

P

将会公开What will be published

WILL PUBLISH

任务集、两种方案的对照数据、适用与不适用的任务类型。The task set, comparison data for both approaches, and the task types each suits and does not suit.

前期线索Early leads相关的已有工作，不是这个实验的结论Related earlier work, not a result of this experiment

  • Jev 类型化决策模型Jev typed decision model目前处于调研阶段。Currently at the research stage.

实验日志Run logRUN LOG

还没有记录。实验开始后，每一次运行、失败和修正都会记在实验室里。Nothing logged yet. Once the experiment starts, every run, failure and fix will be recorded in the lab.

去实验室Go to the lab ↗