11 Safe Dynamic Agent

这份 notebook 的目标不是继续抠一点点准确率，而是解决我们现在最核心的问题：

动态 agent 会把一些错误答案推到极高置信度，因此当前系统不安全。

这份实验先做 安全门 / 拒答机制（abstention），核心思路是：

1. 读取已经跑完的 dynamic agent 病例级结果
2. 设计几个简单但可解释的安全规则
3. 评估覆盖率、选择性准确率、选择性 Macro-F1、危险错误率
4. 把结果按时间戳保存下来

这一版先做 post-hoc safety study。如果方向成立，下一步再把安全机制正式塞回 agent 推理流程里。

from pathlib import Path
from datetime import datetime
import ast
import json

import numpy as np
import pandas as pd
from sklearn.metrics import accuracy_score, balanced_accuracy_score, f1_score

PROJECT_ROOT = Path('/Users/applesues01/Documents/Medical_Agent')
SUPPORT_DIR = PROJECT_ROOT / 'supports'

DYNAMIC_CASES_PATH = SUPPORT_DIR / '2026-08-04_104443_dynamic_agent_cases.csv'
DANGER_CASES_PATH = SUPPORT_DIR / '2026-08-04_104443_dynamic_agent_danger_cases.csv'
IMAGE_ONLY_CASE_PATH = SUPPORT_DIR / 'image_only_case_level_predictions.csv'

print(DYNAMIC_CASES_PATH)
print(DANGER_CASES_PATH)
print(IMAGE_ONLY_CASE_PATH)

/Users/applesues01/Documents/Medical_Agent/supports/2026-08-04_104443_dynamic_agent_cases.csv
/Users/applesues01/Documents/Medical_Agent/supports/2026-08-04_104443_dynamic_agent_danger_cases.csv
/Users/applesues01/Documents/Medical_Agent/supports/image_only_case_level_predictions.csv

1. 读取当前 dynamic agent 结果

这里的 dynamic_agent_cases.csv 是我们已经跑完的病例级轨迹汇总。安全机制就是建立在这些“问之前 vs 问之后”的变化上。

dynamic_df = pd.read_csv(DYNAMIC_CASES_PATH)
image_only_case_df = pd.read_csv(IMAGE_ONLY_CASE_PATH)

def parse_list_cell(value):
    if pd.isna(value):
        return []
    if isinstance(value, list):
        return value
    text = str(value).strip()
    if text == '' or text == '[]':
        return []
    try:
        return ast.literal_eval(text)
    except Exception:
        return [text]

bool_cols = [
    'initial_correct', 'final_correct', 'changed_prediction', 'improved',
    'worsened', 'still_wrong', 'still_correct', 'initial_high_conf_wrong',
    'final_high_conf_wrong', 'unsafe_confidence_increase'
]
for col in bool_cols:
    dynamic_df[col] = dynamic_df[col].astype(str).str.lower().map({'true': True, 'false': False})

dynamic_df['asked_questions_list'] = dynamic_df['asked_questions'].apply(parse_list_cell)
dynamic_df['known_keys_list'] = dynamic_df['known_keys'].apply(parse_list_cell)
dynamic_df['confidence_gain'] = dynamic_df['final_max_prob'] - dynamic_df['initial_max_prob']
dynamic_df['sex_only'] = dynamic_df['known_keys_list'].apply(lambda xs: xs == ['sex'])
dynamic_df['used_age_or_location'] = dynamic_df['known_keys_list'].apply(lambda xs: ('age' in xs) or ('location' in xs))

dynamic_df.head()

	image_id	true_label	initial_pred_label	initial_max_prob	initial_correct	final_pred_label	final_max_prob	final_correct	asked_questions	num_questions	...	still_wrong	still_correct	initial_high_conf_wrong	final_high_conf_wrong	unsafe_confidence_increase	asked_questions_list	known_keys_list	confidence_gain	sex_only	used_age_or_location
0	ISIC_0025837	bkl	bkl	0.972961	True	bkl	0.972961	True	[]	0	...	False	True	False	False	False	[]	[]	0.000000	False	False
1	ISIC_0025209	bkl	bkl	0.407122	True	bkl	0.874424	True	['age', 'sex']	2	...	False	True	False	False	False	[age, sex]	[age, sex]	0.467302	False	True
2	ISIC_0029161	bkl	bkl	0.793358	True	bkl	0.987926	True	['sex']	1	...	False	True	False	False	False	[sex]	[sex]	0.194568	True	False
3	ISIC_0026273	bkl	bkl	0.802558	True	bkl	0.802558	True	[]	0	...	False	True	False	False	False	[]	[]	0.000000	False	False
4	ISIC_0025819	bkl	bkl	0.978878	True	bkl	0.978878	True	[]	0	...	False	True	False	False	False	[]	[]	0.000000	False	False

5 rows × 24 columns

print('Total cases:', len(dynamic_df))
print('Avg questions:', dynamic_df['num_questions'].mean())
print('Initial accuracy:', dynamic_df['initial_correct'].mean())
print('Final accuracy:', dynamic_df['final_correct'].mean())
print('Changed prediction rate:', dynamic_df['changed_prediction'].mean())
print('Unsafe confidence increase rate:', dynamic_df['unsafe_confidence_increase'].mean())
print('Sex-only question count:', dynamic_df['sex_only'].sum())

Total cases: 1481
Avg questions: 0.36934503713706957
Initial accuracy: 0.7771775827143822
Final accuracy: 0.7825793382849426
Changed prediction rate: 0.10465901417960838
Unsafe confidence increase rate: 0.11613774476704929
Sex-only question count: 315

2. 为什么要做安全门

我们已经看到一个很关键的现象：

• agent 有时确实能把错误改对
• 但 agent 也会把错误答案推到接近 1.0 的置信度

所以从这一刻开始，我们的目标不再是“尽量多诊断”，而是：

只在比较可信的时候诊断；其余情况宁可拒答。

这就是 selective prediction / abstention 的思路。

danger_preview = dynamic_df.sort_values('final_max_prob', ascending=False)
danger_preview[[
    'image_id', 'true_label', 'initial_pred_label', 'initial_max_prob',
    'final_pred_label', 'final_max_prob', 'asked_questions',
    'changed_prediction', 'unsafe_confidence_increase', 'final_correct'
]].head(20)

	image_id	true_label	initial_pred_label	initial_max_prob	final_pred_label	final_max_prob	asked_questions	changed_prediction	unsafe_confidence_increase	final_correct
371	ISIC_0024706	vasc	vasc	0.999999	vasc	0.999999	[]	False	False	True
218	ISIC_0025394	mel	nv	0.736395	nv	0.999997	['sex']	False	True	False
368	ISIC_0031065	vasc	vasc	0.999996	vasc	0.999996	[]	False	False	True
316	ISIC_0032070	mel	mel	0.774466	mel	0.999995	['sex']	False	False	True
364	ISIC_0029608	vasc	vasc	0.999995	vasc	0.999995	[]	False	False	True
785	ISIC_0027001	nv	nv	0.999989	nv	0.999989	[]	False	False	True
972	ISIC_0027223	nv	nv	0.999985	nv	0.999985	[]	False	False	True
157	ISIC_0030067	bkl	nv	0.740637	nv	0.999980	['sex']	False	True	False
1077	ISIC_0033402	nv	nv	0.728461	nv	0.999979	['sex']	False	False	True
630	ISIC_0027232	nv	bkl	0.662172	bkl	0.999977	['sex']	False	True	False
597	ISIC_0028318	nv	nv	0.999974	nv	0.999974	[]	False	False	True
15	ISIC_0033899	bkl	mel	0.668489	mel	0.999973	['sex']	False	True	False
634	ISIC_0026300	nv	nv	0.999964	nv	0.999964	[]	False	False	True
588	ISIC_0026628	nv	nv	0.999964	nv	0.999964	[]	False	False	True
850	ISIC_0030685	nv	nv	0.999943	nv	0.999943	[]	False	False	True
1321	ISIC_0026671	nv	nv	0.743409	nv	0.999941	['sex']	False	False	True
596	ISIC_0031508	nv	nv	0.999938	nv	0.999938	[]	False	False	True
141	ISIC_0033460	bkl	bkl	0.750931	mel	0.999934	['sex']	True	False	False
846	ISIC_0031312	nv	nv	0.999934	nv	0.999934	[]	False	False	True
664	ISIC_0027885	nv	nv	0.999933	nv	0.999933	[]	False	False	True

3. 定义安全评估函数

这里我们引入一个新的输出动作：

• diagnose
• abstain

评估时我们不只看准确率，还看：

• coverage：有多少病例被真正诊断
• selective_accuracy：只在被诊断的病例上，准确率是多少
• selective_balanced_accuracy
• selective_macro_f1
• unsafe_rate_among_diagnosed：被诊断病例里有多少是危险高置信错误

def evaluate_abstention_policy(df, diagnose_mask, policy_name):
    out = df.copy()
    out['diagnose'] = diagnose_mask.astype(bool)
    out['abstain'] = ~out['diagnose']

    diagnosed = out[out['diagnose']].copy()
    coverage = len(diagnosed) / len(out)
    abstain_rate = 1.0 - coverage

    if len(diagnosed) == 0:
        return {
            'policy_name': policy_name,
            'coverage': 0.0,
            'abstain_rate': 1.0,
            'diagnosed_cases': 0,
            'selective_accuracy': np.nan,
            'selective_balanced_accuracy': np.nan,
            'selective_macro_f1': np.nan,
            'unsafe_rate_among_diagnosed': np.nan,
            'worsened_rate_among_diagnosed': np.nan,
        }, diagnosed

    selective_accuracy = accuracy_score(diagnosed['true_label'], diagnosed['final_pred_label'])
    selective_balanced_accuracy = balanced_accuracy_score(diagnosed['true_label'], diagnosed['final_pred_label'])
    selective_macro_f1 = f1_score(diagnosed['true_label'], diagnosed['final_pred_label'], average='macro')
    unsafe_rate = diagnosed['unsafe_confidence_increase'].mean()
    worsened_rate = diagnosed['worsened'].mean()

    summary = {
        'policy_name': policy_name,
        'coverage': coverage,
        'abstain_rate': abstain_rate,
        'diagnosed_cases': int(len(diagnosed)),
        'selective_accuracy': selective_accuracy,
        'selective_balanced_accuracy': selective_balanced_accuracy,
        'selective_macro_f1': selective_macro_f1,
        'unsafe_rate_among_diagnosed': unsafe_rate,
        'worsened_rate_among_diagnosed': worsened_rate,
    }
    return summary, diagnosed

4. 先跑几个直观的安全规则

这里先不做复杂模型，先做三种可解释规则：

Rule A: baseline dynamic

全部都诊断，不拒答。

Rule B: no sex-only escalation

如果只问了 sex，就拒答。因为我们已经看到很多危险病例都来自这里。

Rule C: conservative agreement gate

只有在下面条件同时满足时才诊断：

• 最终置信度足够高
• 置信度增长不过分夸张
• 不是 sex-only

这不是最终答案，但非常适合做第一版安全实验。

policy_summaries = []

mask_a = pd.Series(True, index=dynamic_df.index)
summary_a, diagnosed_a = evaluate_abstention_policy(dynamic_df, mask_a, 'baseline_dynamic_no_abstention')
policy_summaries.append(summary_a)

mask_b = ~dynamic_df['sex_only']
summary_b, diagnosed_b = evaluate_abstention_policy(dynamic_df, mask_b, 'abstain_if_sex_only')
policy_summaries.append(summary_b)

mask_c = (
    (dynamic_df['final_max_prob'] >= 0.85) &
    (dynamic_df['confidence_gain'] <= 0.25) &
    (~dynamic_df['sex_only'])
)
summary_c, diagnosed_c = evaluate_abstention_policy(dynamic_df, mask_c, 'conservative_gate_v1')
policy_summaries.append(summary_c)

pd.DataFrame(policy_summaries)

	policy_name	coverage	abstain_rate	diagnosed_cases	selective_accuracy	selective_balanced_accuracy	selective_macro_f1	unsafe_rate_among_diagnosed	worsened_rate_among_diagnosed
0	baseline_dynamic_no_abstention	1.000000	0.000000	1481	0.782579	0.581238	0.581444	0.116138	0.039838
1	abstain_if_sex_only	0.787306	0.212694	1166	0.849914	0.694229	0.659381	0.055746	0.016295
2	conservative_gate_v1	0.609048	0.390952	902	0.917960	0.821088	0.782546	0.012195	0.001109

5. 参数扫描：找一组更稳的规则

下面这一步是我们真正开始“研究”的地方。系统扫描：

• 最终置信度阈值 final_max_prob
• 最大允许置信度增幅 confidence_gain
• 是否禁止 sex-only

重点看：

• coverage 不能太低
• unsafe_rate_among_diagnosed 要明显下降
• selective_macro_f1 尽量不要崩掉

grid_results = []

for prob_thr in [0.75, 0.80, 0.85, 0.90, 0.95]:
    for gain_cap in [0.10, 0.20, 0.30, 0.40, 0.60]:
        for forbid_sex_only in [False, True]:
            mask = dynamic_df['final_max_prob'] >= prob_thr
            mask &= dynamic_df['confidence_gain'] <= gain_cap
            if forbid_sex_only:
                mask &= ~dynamic_df['sex_only']

            summary, _ = evaluate_abstention_policy(
                dynamic_df,
                mask,
                policy_name=f'prob>={prob_thr}_gain<={gain_cap}_forbidSexOnly={forbid_sex_only}'
            )
            summary['prob_threshold'] = prob_thr
            summary['gain_cap'] = gain_cap
            summary['forbid_sex_only'] = forbid_sex_only
            grid_results.append(summary)

grid_df = pd.DataFrame(grid_results)
grid_df.sort_values(
    ['unsafe_rate_among_diagnosed', 'selective_macro_f1'],
    ascending=[True, False]
).head(20)

	policy_name	coverage	abstain_rate	diagnosed_cases	selective_accuracy	selective_balanced_accuracy	selective_macro_f1	unsafe_rate_among_diagnosed	worsened_rate_among_diagnosed	prob_threshold	gain_cap	forbid_sex_only
40	prob>=0.95_gain<=0.1_forbidSexOnly=False	0.411209	0.588791	609	0.978654	0.917516	0.906395	0.000000	0.000000	0.95	0.1	False
41	prob>=0.95_gain<=0.1_forbidSexOnly=True	0.411209	0.588791	609	0.978654	0.917516	0.906395	0.000000	0.000000	0.95	0.1	True
30	prob>=0.9_gain<=0.1_forbidSexOnly=False	0.516543	0.483457	765	0.946405	0.859116	0.837549	0.000000	0.000000	0.90	0.1	False
31	prob>=0.9_gain<=0.1_forbidSexOnly=True	0.516543	0.483457	765	0.946405	0.859116	0.837549	0.000000	0.000000	0.90	0.1	True
20	prob>=0.85_gain<=0.1_forbidSexOnly=False	0.588116	0.411884	871	0.928817	0.827409	0.796048	0.000000	0.000000	0.85	0.1	False
21	prob>=0.85_gain<=0.1_forbidSexOnly=True	0.587441	0.412559	870	0.928736	0.827403	0.796042	0.000000	0.000000	0.85	0.1	True
11	prob>=0.8_gain<=0.1_forbidSexOnly=True	0.653612	0.346388	968	0.911157	0.813572	0.771927	0.001033	0.000000	0.80	0.1	True
1	prob>=0.75_gain<=0.1_forbidSexOnly=True	0.657664	0.342336	974	0.909651	0.812907	0.761828	0.002053	0.000000	0.75	0.1	True
10	prob>=0.8_gain<=0.1_forbidSexOnly=False	0.656313	0.343687	972	0.909465	0.810749	0.770361	0.002058	0.001029	0.80	0.1	False
0	prob>=0.75_gain<=0.1_forbidSexOnly=False	0.661715	0.338285	980	0.908163	0.810217	0.760314	0.003061	0.001020	0.75	0.1	False
43	prob>=0.95_gain<=0.2_forbidSexOnly=True	0.414585	0.585415	614	0.975570	0.905699	0.899476	0.003257	0.000000	0.95	0.2	True
33	prob>=0.9_gain<=0.2_forbidSexOnly=True	0.523295	0.476705	775	0.940645	0.854305	0.827600	0.006452	0.000000	0.90	0.2	True
23	prob>=0.85_gain<=0.2_forbidSexOnly=True	0.596219	0.403781	883	0.922990	0.825248	0.789426	0.006795	0.000000	0.85	0.2	True
42	prob>=0.95_gain<=0.2_forbidSexOnly=False	0.419986	0.580014	622	0.971061	0.905799	0.892970	0.008039	0.000000	0.95	0.2	False
13	prob>=0.8_gain<=0.2_forbidSexOnly=True	0.669818	0.330182	992	0.901210	0.786456	0.744335	0.011089	0.002016	0.80	0.2	True
3	prob>=0.75_gain<=0.2_forbidSexOnly=True	0.675895	0.324105	1001	0.898102	0.784482	0.735513	0.013986	0.001998	0.75	0.2	True
45	prob>=0.95_gain<=0.3_forbidSexOnly=True	0.430790	0.569210	638	0.962382	0.860778	0.850745	0.017241	0.000000	0.95	0.3	True
22	prob>=0.85_gain<=0.2_forbidSexOnly=False	0.613099	0.386901	908	0.912996	0.815159	0.777778	0.017621	0.001101	0.85	0.2	False
32	prob>=0.9_gain<=0.2_forbidSexOnly=False	0.536124	0.463876	794	0.929471	0.840871	0.810085	0.018892	0.000000	0.90	0.2	False
35	prob>=0.9_gain<=0.3_forbidSexOnly=True	0.546253	0.453747	809	0.925834	0.815745	0.789970	0.021014	0.002472	0.90	0.3	True

6. 一个更像论文实验的筛选方式

上面的表会给出很多极端方案，例如：

• 非常安全，但几乎全都拒答
• 覆盖率高，但仍然不安全

所以这里额外加一个“实用筛选”：

• coverage 至少 30%
• unsafe rate 尽量低
• selective Macro-F1 尽量高

practical_df = grid_df[grid_df['coverage'] >= 0.30].copy()
practical_df = practical_df.sort_values(
    ['unsafe_rate_among_diagnosed', 'selective_macro_f1', 'coverage'],
    ascending=[True, False, False]
)
practical_df.head(20)

	policy_name	coverage	abstain_rate	diagnosed_cases	selective_accuracy	selective_balanced_accuracy	selective_macro_f1	unsafe_rate_among_diagnosed	worsened_rate_among_diagnosed	prob_threshold	gain_cap	forbid_sex_only
40	prob>=0.95_gain<=0.1_forbidSexOnly=False	0.411209	0.588791	609	0.978654	0.917516	0.906395	0.000000	0.000000	0.95	0.1	False
41	prob>=0.95_gain<=0.1_forbidSexOnly=True	0.411209	0.588791	609	0.978654	0.917516	0.906395	0.000000	0.000000	0.95	0.1	True
30	prob>=0.9_gain<=0.1_forbidSexOnly=False	0.516543	0.483457	765	0.946405	0.859116	0.837549	0.000000	0.000000	0.90	0.1	False
31	prob>=0.9_gain<=0.1_forbidSexOnly=True	0.516543	0.483457	765	0.946405	0.859116	0.837549	0.000000	0.000000	0.90	0.1	True
20	prob>=0.85_gain<=0.1_forbidSexOnly=False	0.588116	0.411884	871	0.928817	0.827409	0.796048	0.000000	0.000000	0.85	0.1	False
21	prob>=0.85_gain<=0.1_forbidSexOnly=True	0.587441	0.412559	870	0.928736	0.827403	0.796042	0.000000	0.000000	0.85	0.1	True
11	prob>=0.8_gain<=0.1_forbidSexOnly=True	0.653612	0.346388	968	0.911157	0.813572	0.771927	0.001033	0.000000	0.80	0.1	True
1	prob>=0.75_gain<=0.1_forbidSexOnly=True	0.657664	0.342336	974	0.909651	0.812907	0.761828	0.002053	0.000000	0.75	0.1	True
10	prob>=0.8_gain<=0.1_forbidSexOnly=False	0.656313	0.343687	972	0.909465	0.810749	0.770361	0.002058	0.001029	0.80	0.1	False
0	prob>=0.75_gain<=0.1_forbidSexOnly=False	0.661715	0.338285	980	0.908163	0.810217	0.760314	0.003061	0.001020	0.75	0.1	False
43	prob>=0.95_gain<=0.2_forbidSexOnly=True	0.414585	0.585415	614	0.975570	0.905699	0.899476	0.003257	0.000000	0.95	0.2	True
33	prob>=0.9_gain<=0.2_forbidSexOnly=True	0.523295	0.476705	775	0.940645	0.854305	0.827600	0.006452	0.000000	0.90	0.2	True
23	prob>=0.85_gain<=0.2_forbidSexOnly=True	0.596219	0.403781	883	0.922990	0.825248	0.789426	0.006795	0.000000	0.85	0.2	True
42	prob>=0.95_gain<=0.2_forbidSexOnly=False	0.419986	0.580014	622	0.971061	0.905799	0.892970	0.008039	0.000000	0.95	0.2	False
13	prob>=0.8_gain<=0.2_forbidSexOnly=True	0.669818	0.330182	992	0.901210	0.786456	0.744335	0.011089	0.002016	0.80	0.2	True
3	prob>=0.75_gain<=0.2_forbidSexOnly=True	0.675895	0.324105	1001	0.898102	0.784482	0.735513	0.013986	0.001998	0.75	0.2	True
45	prob>=0.95_gain<=0.3_forbidSexOnly=True	0.430790	0.569210	638	0.962382	0.860778	0.850745	0.017241	0.000000	0.95	0.3	True
22	prob>=0.85_gain<=0.2_forbidSexOnly=False	0.613099	0.386901	908	0.912996	0.815159	0.777778	0.017621	0.001101	0.85	0.2	False
32	prob>=0.9_gain<=0.2_forbidSexOnly=False	0.536124	0.463876	794	0.929471	0.840871	0.810085	0.018892	0.000000	0.90	0.2	False
35	prob>=0.9_gain<=0.3_forbidSexOnly=True	0.546253	0.453747	809	0.925834	0.815745	0.789970	0.021014	0.002472	0.90	0.3	True

7. 取当前最优候选，检查它到底拒掉了哪些病例

这一步非常重要，因为安全研究不能只看一个总分。我们要看：

• 它拒掉的是不是危险病例
• 它会不会把很多本来正确的病例也拒掉

best_row = practical_df.iloc[0]
best_row

policy_name                      prob>=0.95_gain<=0.1_forbidSexOnly=False
coverage                                                         0.411209
abstain_rate                                                     0.588791
diagnosed_cases                                                       609
selective_accuracy                                               0.978654
selective_balanced_accuracy                                      0.917516
selective_macro_f1                                               0.906395
unsafe_rate_among_diagnosed                                           0.0
worsened_rate_among_diagnosed                                         0.0
prob_threshold                                                       0.95
gain_cap                                                              0.1
forbid_sex_only                                                     False
Name: 40, dtype: object

best_mask = dynamic_df['final_max_prob'] >= best_row['prob_threshold']
best_mask &= dynamic_df['confidence_gain'] <= best_row['gain_cap']
if bool(best_row['forbid_sex_only']):
    best_mask &= ~dynamic_df['sex_only']

best_summary, best_diagnosed = evaluate_abstention_policy(
    dynamic_df,
    best_mask,
    policy_name='best_safe_candidate'
)

best_summary

{'policy_name': 'best_safe_candidate',
 'coverage': 0.4112086428089129,
 'abstain_rate': 0.5887913571910871,
 'diagnosed_cases': 609,
 'selective_accuracy': 0.9786535303776683,
 'selective_balanced_accuracy': 0.9175158099068161,
 'selective_macro_f1': 0.9063948515845798,
 'unsafe_rate_among_diagnosed': np.float64(0.0),
 'worsened_rate_among_diagnosed': np.float64(0.0)}

abstained_df = dynamic_df[~best_mask].copy()

print('Abstained cases:', len(abstained_df))
print('Unsafe confidence increase among abstained:', abstained_df['unsafe_confidence_increase'].mean())
print('Worsened among abstained:', abstained_df['worsened'].mean())
print('Initially correct among abstained:', abstained_df['initial_correct'].mean())

abstained_df[[
    'image_id', 'true_label', 'initial_pred_label', 'initial_max_prob',
    'final_pred_label', 'final_max_prob', 'asked_questions',
    'unsafe_confidence_increase', 'worsened'
]].head(20)

Abstained cases: 872
Unsafe confidence increase among abstained: 0.19724770642201836
Worsened among abstained: 0.0676605504587156
Initially correct among abstained: 0.6364678899082569

	image_id	true_label	initial_pred_label	initial_max_prob	final_pred_label	final_max_prob	asked_questions	unsafe_confidence_increase	worsened
1	ISIC_0025209	bkl	bkl	0.407122	bkl	0.874424	['age', 'sex']	False	False
2	ISIC_0029161	bkl	bkl	0.793358	bkl	0.987926	['sex']	False	False
3	ISIC_0026273	bkl	bkl	0.802558	bkl	0.802558	[]	False	False
5	ISIC_0032013	bkl	bkl	0.762678	bkl	0.998160	['sex']	False	False
6	ISIC_0029289	bkl	nv	0.464676	nv	0.983204	['sex']	True	False
7	ISIC_0029912	bkl	bkl	0.617092	nv	0.963171	['sex']	False	True
8	ISIC_0033539	bkl	bkl	0.706276	bkl	0.973478	['sex']	False	False
10	ISIC_0029022	bkl	bkl	0.822935	bkl	0.822935	[]	False	False
11	ISIC_0027957	bkl	nv	0.798028	nv	0.998547	['sex']	True	False
12	ISIC_0031212	bkl	nv	0.540349	nv	0.907318	['sex']	True	False
13	ISIC_0033646	bkl	bkl	0.609726	mel	0.663127	['sex', 'age']	False	True
14	ISIC_0033592	bkl	bkl	0.567815	mel	0.986496	['sex']	False	True
15	ISIC_0033899	bkl	mel	0.668489	mel	0.999973	['sex']	True	False
16	ISIC_0033716	bkl	nv	0.373803	mel	0.903718	['location']	True	False
17	ISIC_0034011	bkl	bkl	0.357385	nv	0.882959	['sex']	False	True
18	ISIC_0033613	bkl	bkl	0.837788	bkl	0.837788	[]	False	False
19	ISIC_0032877	bkl	bkl	0.558223	mel	0.819012	['sex']	False	True
20	ISIC_0034175	bkl	bkl	0.864006	bkl	0.864006	[]	False	False
21	ISIC_0033523	bkl	bkl	0.492590	bkl	0.830303	['age', 'location']	False	False
22	ISIC_0033280	bkl	bkl	0.496546	bkl	0.865970	['sex']	False	False

8. 保存结果

这一格会把：

• policy grid 搜索结果
• practical 筛选结果
• 最优候选摘要
• 最优候选的诊断病例
• 最优候选的拒答病例

全部按时间戳保存下来，方便后面回顾和写论文。

timestamp = datetime.now().strftime('%Y-%m-%d_%H%M%S')

policy_grid_path = SUPPORT_DIR / f'{timestamp}_safe_agent_policy_grid.csv'
practical_path = SUPPORT_DIR / f'{timestamp}_safe_agent_practical_candidates.csv'
best_summary_path = SUPPORT_DIR / f'{timestamp}_safe_agent_best_summary.json'
diagnosed_path = SUPPORT_DIR / f'{timestamp}_safe_agent_diagnosed_cases.csv'
abstained_path = SUPPORT_DIR / f'{timestamp}_safe_agent_abstained_cases.csv'

grid_df.to_csv(policy_grid_path, index=False)
practical_df.to_csv(practical_path, index=False)
best_diagnosed.to_csv(diagnosed_path, index=False)
abstained_df.to_csv(abstained_path, index=False)

with open(best_summary_path, 'w', encoding='utf-8') as f:
    json.dump(best_summary, f, ensure_ascii=False, indent=2)

print(policy_grid_path)
print(practical_path)
print(best_summary_path)
print(diagnosed_path)
print(abstained_path)

/Users/applesues01/Documents/Medical_Agent/supports/2026-08-04_133250_safe_agent_policy_grid.csv
/Users/applesues01/Documents/Medical_Agent/supports/2026-08-04_133250_safe_agent_practical_candidates.csv
/Users/applesues01/Documents/Medical_Agent/supports/2026-08-04_133250_safe_agent_best_summary.json
/Users/applesues01/Documents/Medical_Agent/supports/2026-08-04_133250_safe_agent_diagnosed_cases.csv
/Users/applesues01/Documents/Medical_Agent/supports/2026-08-04_133250_safe_agent_abstained_cases.csv

9. 你接下来应该怎么看这些结果

你重点看三件事：

1. unsafe_rate_among_diagnosed 有没有明显下降
2. coverage 会不会掉得太狠
3. selective_macro_f1 能不能保持在一个还能接受的水平

如果这一版有希望，我们下一步就不再做 post-hoc 规则，而是把安全机制正式塞回 dynamic agent 流程里，变成真正的：

ask / diagnose / abstain

三动作 agent。

跑下来我的结论是：保守一点也未必是件坏事

10. Safety-Coverage Trade-off Summary

这一节不再看几十条规则，而是只挑 3 个最有代表性的安全方案：

• 高安全：尽可能避免危险高置信错误
• 中等平衡：安全和覆盖率之间折中
• 高覆盖：尽量多诊断，但仍比原始 dynamic agent 更稳

这 3 组结果很适合后面直接写进论文。

representative_policies = [
    'prob>=0.95_gain<=0.1_forbidSexOnly=False',
    'prob>=0.9_gain<=0.1_forbidSexOnly=False',
    'prob>=0.85_gain<=0.1_forbidSexOnly=False',
]

tradeoff_df = grid_df[grid_df['policy_name'].isin(representative_policies)].copy()

label_map = {
    'prob>=0.95_gain<=0.1_forbidSexOnly=False': 'high_safety',
    'prob>=0.9_gain<=0.1_forbidSexOnly=False': 'balanced',
    'prob>=0.85_gain<=0.1_forbidSexOnly=False': 'high_coverage',
}

tradeoff_df['policy_label'] = tradeoff_df['policy_name'].map(label_map)
tradeoff_df = tradeoff_df[[
    'policy_label',
    'policy_name',
    'coverage',
    'abstain_rate',
    'diagnosed_cases',
    'selective_accuracy',
    'selective_balanced_accuracy',
    'selective_macro_f1',
    'unsafe_rate_among_diagnosed',
    'worsened_rate_among_diagnosed',
    'prob_threshold',
    'gain_cap',
    'forbid_sex_only'
]].sort_values('coverage')

tradeoff_df

	policy_label	policy_name	coverage	abstain_rate	diagnosed_cases	selective_accuracy	selective_balanced_accuracy	selective_macro_f1	unsafe_rate_among_diagnosed	worsened_rate_among_diagnosed	prob_threshold	gain_cap	forbid_sex_only
40	high_safety	prob>=0.95_gain<=0.1_forbidSexOnly=False	0.411209	0.588791	609	0.978654	0.917516	0.906395	0.0	0.0	0.95	0.1	False
30	balanced	prob>=0.9_gain<=0.1_forbidSexOnly=False	0.516543	0.483457	765	0.946405	0.859116	0.837549	0.0	0.0	0.90	0.1	False
20	high_coverage	prob>=0.85_gain<=0.1_forbidSexOnly=False	0.588116	0.411884	871	0.928817	0.827409	0.796048	0.0	0.0	0.85	0.1	False

怎么读这张表

• high_safety：最保守，诊断最少，但最安全
• balanced：我建议后面论文里重点讨论这一组
• high_coverage：诊断更多，但安全性开始下降

这三组能非常直观地展示：

Safety 和 Coverage 之间存在清晰的 trade-off。

paper_tradeoff_df = tradeoff_df.copy()

for col in [
    'coverage', 'abstain_rate', 'selective_accuracy',
    'selective_balanced_accuracy', 'selective_macro_f1',
    'unsafe_rate_among_diagnosed', 'worsened_rate_among_diagnosed'
]:
    paper_tradeoff_df[col] = paper_tradeoff_df[col].map(lambda x: round(float(x), 4))

paper_tradeoff_df

	policy_label	policy_name	coverage	abstain_rate	diagnosed_cases	selective_accuracy	selective_balanced_accuracy	selective_macro_f1	unsafe_rate_among_diagnosed	worsened_rate_among_diagnosed	prob_threshold	gain_cap	forbid_sex_only
40	high_safety	prob>=0.95_gain<=0.1_forbidSexOnly=False	0.4112	0.5888	609	0.9787	0.9175	0.9064	0.0	0.0	0.95	0.1	False
30	balanced	prob>=0.9_gain<=0.1_forbidSexOnly=False	0.5165	0.4835	765	0.9464	0.8591	0.8375	0.0	0.0	0.90	0.1	False
20	high_coverage	prob>=0.85_gain<=0.1_forbidSexOnly=False	0.5881	0.4119	871	0.9288	0.8274	0.7960	0.0	0.0	0.85	0.1	False

tradeoff_timestamp = datetime.now().strftime('%Y-%m-%d_%H%M%S')

tradeoff_csv_path = SUPPORT_DIR / f'{tradeoff_timestamp}_safe_agent_tradeoff_summary.csv'
tradeoff_json_path = SUPPORT_DIR / f'{tradeoff_timestamp}_safe_agent_tradeoff_summary.json'

paper_tradeoff_df.to_csv(tradeoff_csv_path, index=False)
paper_tradeoff_df.to_json(tradeoff_json_path, orient='records', force_ascii=False, indent=2)

print(tradeoff_csv_path)
print(tradeoff_json_path)

/Users/applesues01/Documents/Medical_Agent/supports/2026-08-04_141020_safe_agent_tradeoff_summary.csv
/Users/applesues01/Documents/Medical_Agent/supports/2026-08-04_141020_safe_agent_tradeoff_summary.json

自古忠孝难两全