Spaces:
Running
Running
Update backend/agents/reflection_agent.py
#3
by
munals - opened
backend/agents/reflection_agent.py
CHANGED
|
@@ -51,7 +51,6 @@ class ReflectionAgent:
|
|
| 51 |
passing it downstream to OperationsAgent.
|
| 52 |
"""
|
| 53 |
|
| 54 |
-
# ββ Step 1: Observe ββββββββββββββββββββββββββββββββββββββββββ
|
| 55 |
self._history.append({
|
| 56 |
'risk_level': rr.risk_level,
|
| 57 |
'risk_score': rr.risk_score,
|
|
@@ -64,8 +63,6 @@ class ReflectionAgent:
|
|
| 64 |
corrected_level = rr.risk_level
|
| 65 |
corrected_score = rr.risk_score
|
| 66 |
|
| 67 |
-
# ββ Step 2 & 3: Critique + Correct βββββββββββββββββββββββββββ
|
| 68 |
-
|
| 69 |
# Bias 1: Chronic LOW under-reporting
|
| 70 |
if len(self._history) >= self.BIAS_WINDOW:
|
| 71 |
all_low = all(h['risk_level'] == 'LOW' for h in self._history)
|
|
@@ -73,7 +70,7 @@ class ReflectionAgent:
|
|
| 73 |
if all_low and avg_crowd > self.CROWD_LOW_THRESH:
|
| 74 |
bias_detected = True
|
| 75 |
corrected_level = 'MEDIUM'
|
| 76 |
-
corrected_score = max(rr.risk_score, 0.
|
| 77 |
critique.append(
|
| 78 |
f'CHRONIC_LOW_BIAS: {self.BIAS_WINDOW} consecutive LOW '
|
| 79 |
f'with avg crowd={avg_crowd:.0f} persons. Upgraded to MEDIUM.'
|
|
@@ -83,7 +80,7 @@ class ReflectionAgent:
|
|
| 83 |
if rr.trend == 'rising' and rr.risk_level == 'LOW' and fr.person_count > 20:
|
| 84 |
bias_detected = True
|
| 85 |
corrected_level = 'MEDIUM'
|
| 86 |
-
corrected_score = max(corrected_score, 0.
|
| 87 |
critique.append(
|
| 88 |
f'RISING_TREND_IGNORED: trend=rising, persons={fr.person_count}, '
|
| 89 |
f'but risk=LOW. Upgraded to MEDIUM.'
|
|
@@ -91,20 +88,20 @@ class ReflectionAgent:
|
|
| 91 |
|
| 92 |
# Bias 3: Count-risk mismatch (enhanced with critical threshold)
|
| 93 |
if fr.person_count >= self.CRITICAL_CROWD_THRESH and rr.risk_level == 'LOW':
|
| 94 |
-
#
|
| 95 |
bias_detected = True
|
| 96 |
corrected_level = 'HIGH'
|
| 97 |
-
corrected_score = max(corrected_score, 0.
|
| 98 |
critique.append(
|
| 99 |
f'CRITICAL_COUNT_RISK_MISMATCH: {fr.person_count} persons but risk=LOW. '
|
| 100 |
-
f'This is a severe inconsistency
|
| 101 |
f'Upgraded to HIGH (corrected_score={corrected_score:.3f}).'
|
| 102 |
)
|
| 103 |
elif fr.person_count > self.HIGH_CROWD_THRESH and rr.risk_level == 'LOW':
|
| 104 |
-
#
|
| 105 |
bias_detected = True
|
| 106 |
corrected_level = 'MEDIUM'
|
| 107 |
-
corrected_score = max(corrected_score, 0.
|
| 108 |
critique.append(
|
| 109 |
f'COUNT_RISK_MISMATCH: {fr.person_count} persons but risk=LOW. '
|
| 110 |
f'Thresholds may need recalibration. Upgraded to MEDIUM.'
|
|
@@ -114,7 +111,7 @@ class ReflectionAgent:
|
|
| 114 |
if rr.risk_level == 'HIGH' and fr.person_count < config.REFLECTION_OVER_EST_THRESH:
|
| 115 |
bias_detected = True
|
| 116 |
corrected_level = 'MEDIUM'
|
| 117 |
-
corrected_score = min(corrected_score, 0.
|
| 118 |
critique.append(
|
| 119 |
f'OVER_ESTIMATION: HIGH risk but only {fr.person_count} persons. '
|
| 120 |
f'Downgraded to MEDIUM β possible empty-frame or detection artifact.'
|
|
@@ -123,7 +120,6 @@ class ReflectionAgent:
|
|
| 123 |
if not critique:
|
| 124 |
critique.append('OK: assessment consistent, no bias detected.')
|
| 125 |
|
| 126 |
-
# ββ Step 4: Log ββββββββββββββββββββββββββββββββββββββββββββββ
|
| 127 |
reflection = {
|
| 128 |
'frame_id': rr.frame_id,
|
| 129 |
'original_level': rr.risk_level,
|
|
|
|
| 51 |
passing it downstream to OperationsAgent.
|
| 52 |
"""
|
| 53 |
|
|
|
|
| 54 |
self._history.append({
|
| 55 |
'risk_level': rr.risk_level,
|
| 56 |
'risk_score': rr.risk_score,
|
|
|
|
| 63 |
corrected_level = rr.risk_level
|
| 64 |
corrected_score = rr.risk_score
|
| 65 |
|
|
|
|
|
|
|
| 66 |
# Bias 1: Chronic LOW under-reporting
|
| 67 |
if len(self._history) >= self.BIAS_WINDOW:
|
| 68 |
all_low = all(h['risk_level'] == 'LOW' for h in self._history)
|
|
|
|
| 70 |
if all_low and avg_crowd > self.CROWD_LOW_THRESH:
|
| 71 |
bias_detected = True
|
| 72 |
corrected_level = 'MEDIUM'
|
| 73 |
+
corrected_score = max(rr.risk_score, 0.30) # push into MEDIUM zone (>0.20)
|
| 74 |
critique.append(
|
| 75 |
f'CHRONIC_LOW_BIAS: {self.BIAS_WINDOW} consecutive LOW '
|
| 76 |
f'with avg crowd={avg_crowd:.0f} persons. Upgraded to MEDIUM.'
|
|
|
|
| 80 |
if rr.trend == 'rising' and rr.risk_level == 'LOW' and fr.person_count > 20:
|
| 81 |
bias_detected = True
|
| 82 |
corrected_level = 'MEDIUM'
|
| 83 |
+
corrected_score = max(corrected_score, 0.25) # push into MEDIUM zone (>0.20)
|
| 84 |
critique.append(
|
| 85 |
f'RISING_TREND_IGNORED: trend=rising, persons={fr.person_count}, '
|
| 86 |
f'but risk=LOW. Upgraded to MEDIUM.'
|
|
|
|
| 88 |
|
| 89 |
# Bias 3: Count-risk mismatch (enhanced with critical threshold)
|
| 90 |
if fr.person_count >= self.CRITICAL_CROWD_THRESH and rr.risk_level == 'LOW':
|
| 91 |
+
# 70+ persons with LOW risk β CRITICAL inconsistency β upgrade to HIGH
|
| 92 |
bias_detected = True
|
| 93 |
corrected_level = 'HIGH'
|
| 94 |
+
corrected_score = max(corrected_score, 0.85) # push into HIGH zone (>0.80)
|
| 95 |
critique.append(
|
| 96 |
f'CRITICAL_COUNT_RISK_MISMATCH: {fr.person_count} persons but risk=LOW. '
|
| 97 |
+
f'This is a severe inconsistency. '
|
| 98 |
f'Upgraded to HIGH (corrected_score={corrected_score:.3f}).'
|
| 99 |
)
|
| 100 |
elif fr.person_count > self.HIGH_CROWD_THRESH and rr.risk_level == 'LOW':
|
| 101 |
+
# 50+ persons with LOW risk β upgrade to MEDIUM
|
| 102 |
bias_detected = True
|
| 103 |
corrected_level = 'MEDIUM'
|
| 104 |
+
corrected_score = max(corrected_score, 0.30) # push into MEDIUM zone (>0.20)
|
| 105 |
critique.append(
|
| 106 |
f'COUNT_RISK_MISMATCH: {fr.person_count} persons but risk=LOW. '
|
| 107 |
f'Thresholds may need recalibration. Upgraded to MEDIUM.'
|
|
|
|
| 111 |
if rr.risk_level == 'HIGH' and fr.person_count < config.REFLECTION_OVER_EST_THRESH:
|
| 112 |
bias_detected = True
|
| 113 |
corrected_level = 'MEDIUM'
|
| 114 |
+
corrected_score = min(corrected_score, 0.50) # pull down to MEDIUM zone
|
| 115 |
critique.append(
|
| 116 |
f'OVER_ESTIMATION: HIGH risk but only {fr.person_count} persons. '
|
| 117 |
f'Downgraded to MEDIUM β possible empty-frame or detection artifact.'
|
|
|
|
| 120 |
if not critique:
|
| 121 |
critique.append('OK: assessment consistent, no bias detected.')
|
| 122 |
|
|
|
|
| 123 |
reflection = {
|
| 124 |
'frame_id': rr.frame_id,
|
| 125 |
'original_level': rr.risk_level,
|