PRIMARY EXPERIMENT
Verifier Ablation · 验证器作用实验
回答 verifier、详细 gate feedback 与 agent-facing self-check tools 是否带来质量收益,以及代价是多少。主实验结果
60 new generations + 10 reused L1 · 70 freshly scored rowsMatched-control run 已完成;这里直接比较 verifier、详细 gate feedback 与 agent-facing align/diff 的 observed paired effect。
Loading matched-control drift calibration…
7 matched conditions
每行固定 10 个 matched cases。Abs / Large 的分母 10 是 case-h1…case-h10;分子是至少命中过一次该风险的 cell 数,不是 Structure 的通过项数。
| condition | Gate mean ± σ 10-case mean | ≥ 4 cases / 10 | semantic 10-case mean | geometry 10-case mean | structure 10-case mean | canvas 10-case mean | Abs flagged cases / 10 | Large flagged cases / 10 | verifier dose calls total | time mean / cell |
|---|---|---|---|---|---|---|---|---|---|---|
| Waiting for canonical comparison… | ||||||||||
Primary paired contrasts
正值表示左侧 treatment 的 Gate 更高;CI 是固定 seed 的 10,000 次 paired bootstrap。单 replicate、10 cases,只作 observed paired effect 解读。
| contrast | Gate Δ | 95% CI | W / T / L | semantic Δ | geometry Δ | structure Δ | time Δ | dose / tools Δ |
|---|---|---|---|---|---|---|---|---|
| Waiting for paired analysis… | ||||||||
10 cases × 7 matched conditions
每格是该 case × condition 的具体 Gate 总分,不是 replicate 平均。每行最高值加粗;点击后进入对应 Viewer evidence。
| case | L1 Direct | L2 Base | L2 Scoped | L2 Detail | L3 Full | L3 Hidden | L3 No Self-Check |
|---|---|---|---|---|---|---|---|
| Waiting for 70 freshly scored rows… | |||||||
Behavior dose & contract compliance实际调用剂量、turns、epochs 与 agent-visible policy点击展开
L2 verifier behavior
| condition | cells calling verifier | calls total / mean | turns mean | time mean |
|---|---|---|---|---|
| Loading L2 dose… | ||||
L3 harness compliance
| condition | actual epochs | turns mean | tool calls mean | gate feedback | agent-visible MCP tools |
|---|---|---|---|---|---|
| Loading L3 compliance… | |||||