docs: clarify Explore harness metric fit (#1730)

Author: huangruiteng

README.md

@@ -197,6 +197,12 @@ credentials, or AB/canary gates.
 LoopX will plan before writing state, then create ordered P0/P1/P2 todos that
 make the algorithm, infra, validation, and human gates visible.
 
+This Explore-style harness works best when the task has a measurable offline
+eval, baseline, treatment, and guardrail metrics. It is not recommended for
+open-ended tasks or scenarios where the evaluation metric cannot be quantified
+reliably, because the loop depends on comparable metrics to decide whether an
+exploration branch actually improved the result.
+
 <details>
 <summary>Example visible todos</summary>
 

README.zh-CN.md

@@ -348,6 +348,10 @@ Result Layer 把探索中的 `node`、`edge`、`finding` 记录成 public-safe
 Harness 读取这张图、open todo、ownership、capability 和 write-scope 信息，生成只读的
 todo branch plan 或 worker lane plan。
 
+适用边界：Explore Harness 依赖可量化的评估指标来比较 baseline、treatment 和
+guardrail。对于开放任务，或者评估指标无法准确量化的场景，不推荐直接使用这个
+harness；这类问题更适合先沉淀成人工判断、定性研究或更小的可度量子问题。
+
 ```bash
 loopx explore node --goal-id <goal-id> --title "Map the next frontier"
 loopx explore finding --goal-id <goal-id> --title "Confirmed reusable contract" --node <node-id>