File size: 10,664 Bytes
ed1b365 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 | """Codette Guardian — Input Safety, Ethical Checks, Trust Calibration
Three-layer protection:
1. InputSanitizer: Catches injection, XSS, encoded attacks
2. EthicalAnchor: Tracks ethical regret and learning over time
3. TrustCalibrator: Dynamic trust scores for adapter/agent outputs
Origin: input_sanitizer.py + validate_ethics.py + trust_logic.py +
Codette_Deep_Simulation_v1.py (EthicalAnchor), rebuilt
"""
import re
import math
import time
import logging
from dataclasses import dataclass, field
from typing import Dict, List, Optional
logger = logging.getLogger(__name__)
# ================================================================
# Layer 1: Input Sanitization
# ================================================================
class InputSanitizer:
"""Detect and neutralize injection patterns in user input."""
_INJECTION_PATTERNS = re.compile(
r"(?:"
r"\\[nr]|" # Escaped newlines
r"�[ad];|" # HTML entities for CR/LF
r"%0[ad]|" # URL-encoded CR/LF
r"<script|" # Script injection
r"<iframe|" # IFrame injection
r";--|" # SQL comment injection
r"UNION\s+SELECT|" # SQL union
r"\bDROP\s+TABLE|" # SQL drop
r"javascript:|" # JS protocol
r"data:text/html" # Data URI XSS
r")",
re.IGNORECASE,
)
_PROMPT_INJECTION = re.compile(
r"(?:"
r"ignore\s+(?:all\s+)?(?:previous|above)|"
r"disregard\s+(?:your|all)|"
r"you\s+are\s+now|"
r"new\s+instructions?:|"
r"system\s*prompt:|"
r"forget\s+everything"
r")",
re.IGNORECASE,
)
def sanitize(self, text: str) -> str:
"""Remove dangerous patterns, return cleaned text."""
original = text
text = self._INJECTION_PATTERNS.sub("[BLOCKED]", text)
if text != original:
logger.warning("Input sanitized: injection pattern detected")
return text
def detect_threats(self, text: str) -> Dict[str, bool]:
"""Analyze text for various threat types."""
return {
"injection": bool(self._INJECTION_PATTERNS.search(text)),
"prompt_injection": bool(self._PROMPT_INJECTION.search(text)),
"excessive_length": len(text) > 10000,
}
def is_safe(self, text: str) -> bool:
"""Quick safety check — True if no threats detected."""
threats = self.detect_threats(text)
return not any(threats.values())
# ================================================================
# Layer 2: Ethical Anchor (from Deep Simulation)
# ================================================================
@dataclass
class EthicalAnchor:
"""Tracks ethical alignment through regret-based learning.
The ethical score M evolves as:
M = λ(R + H) + γ·Learn(M_prev, E) + μ·regret
Where regret = |intended - actual| measures the gap between
what the system intended to do and what it actually did.
"""
lam: float = 0.7 # Weight for recent reasoning + history
gamma: float = 0.5 # Weight for learning from experience
mu: float = 0.3 # Weight for regret signal
learning_rate: float = 0.2
score: float = 0.5 # Current ethical alignment score [0, 1]
total_regret: float = 0.0
history: List[Dict] = field(default_factory=list)
def update(self, coherence: float, tension: float,
intended_helpfulness: float = 0.8,
actual_helpfulness: float = 0.7) -> float:
"""Update ethical score after a response.
Args:
coherence: How coherent the response was [0, 1]
tension: Epistemic tension level [0, 1]
intended_helpfulness: What we aimed for [0, 1]
actual_helpfulness: Estimated actual quality [0, 1]
"""
regret = abs(intended_helpfulness - actual_helpfulness)
self.total_regret += regret
# Learning signal: move toward better alignment
learn = self.learning_rate * (coherence - self.score)
# New score
reasoning_quality = 0.5 * coherence + 0.5 * (1.0 - tension)
self.score = (
self.lam * reasoning_quality
+ self.gamma * learn
+ self.mu * (1.0 - regret) # Low regret → high ethics
)
self.score = max(0.0, min(1.0, self.score))
record = {
"timestamp": time.time(),
"score": round(self.score, 4),
"regret": round(regret, 4),
"coherence": round(coherence, 4),
}
self.history.append(record)
# Keep only recent history
if len(self.history) > 50:
self.history = self.history[-50:]
return self.score
def get_state(self) -> Dict:
return {
"ethical_score": round(self.score, 4),
"total_regret": round(self.total_regret, 4),
"recent_trend": self._trend(),
}
def _trend(self) -> str:
if len(self.history) < 3:
return "insufficient_data"
recent = [h["score"] for h in self.history[-5:]]
slope = recent[-1] - recent[0]
if slope > 0.05:
return "improving"
elif slope < -0.05:
return "declining"
return "stable"
def to_dict(self) -> Dict:
return {
"score": self.score,
"total_regret": self.total_regret,
"history": self.history[-10:],
}
@classmethod
def from_dict(cls, d: Dict) -> "EthicalAnchor":
anchor = cls()
anchor.score = d.get("score", 0.5)
anchor.total_regret = d.get("total_regret", 0.0)
anchor.history = d.get("history", [])
return anchor
# ================================================================
# Layer 3: Trust Calibration
# ================================================================
class TrustCalibrator:
"""Dynamic trust scores for adapter outputs.
Trust increases when outputs are coherent, helpful, and ethically sound.
Trust decreases for incoherent, harmful, or low-quality outputs.
"""
def __init__(self):
self.trust_scores: Dict[str, float] = {}
self.interaction_counts: Dict[str, int] = {}
def get_trust(self, adapter: str) -> float:
"""Get current trust score for an adapter [0.05, 1.5]."""
return self.trust_scores.get(adapter, 1.0)
def update(self, adapter: str, coherence: float = 0.5,
was_helpful: bool = True, ethical_score: float = 0.5):
"""Update trust for an adapter based on output quality."""
current = self.trust_scores.get(adapter, 1.0)
count = self.interaction_counts.get(adapter, 0)
# Quality composite
quality = 0.4 * coherence + 0.3 * float(was_helpful) + 0.3 * ethical_score
# Adaptive adjustment (smaller changes as trust stabilizes)
adjustment_rate = 0.1 / (1.0 + count * 0.01)
if quality > 0.6:
current *= (1.0 + adjustment_rate)
elif quality < 0.3:
current *= (1.0 - 2 * adjustment_rate)
else:
current *= (1.0 - 0.5 * adjustment_rate)
# Clamp to valid range
current = max(0.05, min(1.5, current))
self.trust_scores[adapter] = current
self.interaction_counts[adapter] = count + 1
def weighted_consensus(self, adapter_responses: Dict[str, str]) -> List[str]:
"""Rank adapter responses by trust-weighted priority."""
ranked = sorted(
adapter_responses.keys(),
key=lambda a: self.get_trust(a),
reverse=True,
)
return ranked
def get_state(self) -> Dict:
return {
"trust_scores": {k: round(v, 3) for k, v in self.trust_scores.items()},
"total_interactions": sum(self.interaction_counts.values()),
}
def to_dict(self) -> Dict:
return {
"trust_scores": self.trust_scores,
"interaction_counts": self.interaction_counts,
}
@classmethod
def from_dict(cls, d: Dict) -> "TrustCalibrator":
cal = cls()
cal.trust_scores = d.get("trust_scores", {})
cal.interaction_counts = d.get("interaction_counts", {})
return cal
# ================================================================
# Combined Guardian
# ================================================================
class CodetteGuardian:
"""Unified guardian combining all three safety layers."""
def __init__(self):
self.sanitizer = InputSanitizer()
self.ethics = EthicalAnchor()
self.trust = TrustCalibrator()
def check_input(self, text: str) -> Dict:
"""Check user input for safety issues."""
threats = self.sanitizer.detect_threats(text)
safe_text = self.sanitizer.sanitize(text) if any(threats.values()) else text
return {
"safe": not any(threats.values()),
"threats": threats,
"cleaned_text": safe_text,
}
def evaluate_output(self, adapter: str, response: str,
coherence: float = 0.5, tension: float = 0.3):
"""Evaluate an adapter's output and update trust/ethics."""
# Estimate helpfulness from response quality signals
helpful = len(response) > 50 and coherence > 0.3
self.ethics.update(
coherence=coherence,
tension=tension,
actual_helpfulness=0.7 if helpful else 0.3,
)
self.trust.update(
adapter=adapter,
coherence=coherence,
was_helpful=helpful,
ethical_score=self.ethics.score,
)
def get_state(self) -> Dict:
return {
"ethics": self.ethics.get_state(),
"trust": self.trust.get_state(),
}
def to_dict(self) -> Dict:
return {
"ethics": self.ethics.to_dict(),
"trust": self.trust.to_dict(),
}
@classmethod
def from_dict(cls, d: Dict) -> "CodetteGuardian":
g = cls()
if "ethics" in d:
g.ethics = EthicalAnchor.from_dict(d["ethics"])
if "trust" in d:
g.trust = TrustCalibrator.from_dict(d["trust"])
return g
|