ESL_MEDIATOR.txt

#!/usr/bin/env python3
"""
EIS + ESL MEDIATOR v2.0 – Full Epistemic Substrate with Suppression Analytics
================================================================================
Adds:
- Cross‑claim contradiction tracking (graph)
- Signature weighting (suppression‑likelihood scores)
- Entity‑coherence scoring (temporal consistency)
- Suppression‑pattern classifier (aggregates signatures into threat levels)
- Narrative‑violation detector (checks LLM output for narrative drift)
"""

import hashlib
import json
import os
import secrets
import time
import math
from datetime import datetime
from typing import Dict, List, Any, Optional, Tuple, Set
from collections import defaultdict
import requests

# ============================================================================
# PART 1: CRYPTOGRAPHIC HELPERS
# ============================================================================

def sha3_512(data: str) -> str:
    return hashlib.sha3_512(data.encode()).hexdigest()

def hash_dict(data: Dict) -> str:
    return sha3_512(json.dumps(data, sort_keys=True, separators=(',', ':')))

# ============================================================================
# PART 2: ENHANCED EPISTEMIC SUBSTRATE LEDGER (ESL)
# ============================================================================

class ESLedger:
    """Persistent ledger with cross‑claim contradictions, signature weights, coherence."""
    def __init__(self, path: str = "esl_ledger.json"):
        self.path = path
        self.claims: Dict[str, Dict] = {}               # claim_id -> claim dict
        self.entities: Dict[str, Dict] = {}             # entity_name -> entity dict
        self.signatures: List[Dict] = []                # signature logs with weights
        self.contradiction_graph: Dict[str, Set[str]] = defaultdict(set)  # claim_id -> set of contradictory claim_ids
        self.blocks: List[Dict] = []
        self._load()

    def _load(self):
        if os.path.exists(self.path):
            try:
                with open(self.path, 'r') as f:
                    data = json.load(f)
                    self.claims = data.get("claims", {})
                    self.entities = data.get("entities", {})
                    self.signatures = data.get("signatures", [])
                    self.blocks = data.get("blocks", [])
                    # Load contradiction graph as sets
                    cg = data.get("contradiction_graph", {})
                    self.contradiction_graph = {k: set(v) for k, v in cg.items()}
            except Exception:
                pass

    def _save(self):
        # Convert sets to lists for JSON
        cg_serializable = {k: list(v) for k, v in self.contradiction_graph.items()}
        data = {
            "claims": self.claims,
            "entities": self.entities,
            "signatures": self.signatures,
            "contradiction_graph": cg_serializable,
            "blocks": self.blocks,
            "updated": datetime.utcnow().isoformat() + "Z"
        }
        with open(self.path + ".tmp", 'w') as f:
            json.dump(data, f, indent=2)
        os.replace(self.path + ".tmp", self.path)

    def add_claim(self, text: str, agent: str = "user") -> str:
        claim_id = secrets.token_hex(16)
        self.claims[claim_id] = {
            "id": claim_id, "text": text, "agent": agent,
            "timestamp": datetime.utcnow().isoformat() + "Z",
            "entities": [], "signatures": [], "coherence": 0.5,
            "contradictions": [], "suppression_score": 0.0
        }
        self._save()
        return claim_id

    def add_entity(self, name: str, etype: str, claim_id: str):
        if name not in self.entities:
            self.entities[name] = {
                "name": name, "type": etype,
                "first_seen": datetime.utcnow().isoformat() + "Z",
                "last_seen": self.claims[claim_id]["timestamp"],
                "appearances": [], "coherence_scores": []
            }
        ent = self.entities[name]
        if claim_id not in ent["appearances"]:
            ent["appearances"].append(claim_id)
        ent["last_seen"] = self.claims[claim_id]["timestamp"]
        self.claims[claim_id]["entities"].append(name)
        self._save()

    def add_signature(self, claim_id: str, sig_name: str, weight: float = 0.5, context: Dict = None):
        """Add a signature with a weight (0-1) indicating suppression likelihood."""
        self.signatures.append({
            "signature": sig_name, "claim_id": claim_id,
            "timestamp": datetime.utcnow().isoformat() + "Z",
            "weight": weight, "context": context or {}
        })
        if sig_name not in self.claims[claim_id]["signatures"]:
            self.claims[claim_id]["signatures"].append(sig_name)
        # Update suppression score for the claim (max of signature weights)
        current = self.claims[claim_id].get("suppression_score", 0.0)
        self.claims[claim_id]["suppression_score"] = max(current, weight)
        self._save()

    def add_contradiction(self, claim_id_a: str, claim_id_b: str):
        """Record that two claims contradict each other."""
        self.contradiction_graph[claim_id_a].add(claim_id_b)
        self.contradiction_graph[claim_id_b].add(claim_id_a)
        # Update each claim's contradiction list
        if claim_id_b not in self.claims[claim_id_a]["contradictions"]:
            self.claims[claim_id_a]["contradictions"].append(claim_id_b)
        if claim_id_a not in self.claims[claim_id_b]["contradictions"]:
            self.claims[claim_id_b]["contradictions"].append(claim_id_a)
        self._save()

    def get_entity_coherence(self, entity_name: str) -> float:
        """Calculate temporal coherence for an entity: low variance in appearance intervals."""
        ent = self.entities.get(entity_name)
        if not ent or len(ent["appearances"]) < 2:
            return 0.5
        timestamps = []
        for cid in ent["appearances"]:
            ts = self.claims[cid]["timestamp"]
            timestamps.append(datetime.fromisoformat(ts.replace('Z', '+00:00')))
        # Compute average interval variance (simplified)
        intervals = [(timestamps[i+1] - timestamps[i]).total_seconds() / 86400 for i in range(len(timestamps)-1)]
        if not intervals:
            return 0.5
        mean = sum(intervals) / len(intervals)
        variance = sum((i - mean)**2 for i in intervals) / len(intervals)
        # Coherence is high when variance is low (normalized)
        coherence = 1.0 / (1.0 + variance)
        return min(1.0, max(0.0, coherence))

    def suppression_pattern_classifier(self, claim_id: str) -> Dict:
        """Aggregate signatures into suppression pattern threat levels."""
        claim = self.claims.get(claim_id, {})
        sig_names = claim.get("signatures", [])
        if not sig_names:
            return {"level": "none", "score": 0.0, "patterns": []}

        # Predefined pattern groups (signature -> pattern)
        pattern_map = {
            "entity_present_then_absent": "erasure",
            "gradual_fading": "erasure",
            "single_explanation": "narrative_capture",
            "ad_hominem_attacks": "discreditation",
            "deflection": "misdirection",
            "archival_gaps": "erasure",
            "repetitive_messaging": "conditioning"
        }
        patterns = []
        total_weight = 0.0
        for sig in sig_names:
            pat = pattern_map.get(sig, "unknown")
            patterns.append(pat)
            # Find weight from signature logs
            weight = 0.5
            for log in self.signatures:
                if log["signature"] == sig and log["claim_id"] == claim_id:
                    weight = log.get("weight", 0.5)
                    break
            total_weight += weight
        avg_weight = total_weight / len(sig_names) if sig_names else 0.0

        if avg_weight > 0.7:
            level = "high"
        elif avg_weight > 0.4:
            level = "medium"
        elif avg_weight > 0.1:
            level = "low"
        else:
            level = "none"

        return {"level": level, "score": avg_weight, "patterns": list(set(patterns))}

    def get_entity_timeline(self, name: str) -> List[Dict]:
        ent = self.entities.get(name)
        if not ent:
            return []
        timeline = []
        for cid in ent["appearances"]:
            claim = self.claims.get(cid)
            if claim:
                timeline.append({
                    "timestamp": claim["timestamp"],
                    "text": claim["text"]
                })
        timeline.sort(key=lambda x: x["timestamp"])
        return timeline

    def disappearance_suspected(self, name: str, threshold_days: int = 30) -> bool:
        timeline = self.get_entity_timeline(name)
        if not timeline:
            return False
        last = datetime.fromisoformat(timeline[-1]["timestamp"].replace('Z', '+00:00'))
        now = datetime.utcnow()
        return (now - last).days > threshold_days

    def create_block(self) -> Dict:
        block = {
            "index": len(self.blocks),
            "timestamp": datetime.utcnow().isoformat() + "Z",
            "prev_hash": self.blocks[-1]["hash"] if self.blocks else "0"*64,
            "state_hash": hash_dict({"claims": self.claims, "entities": self.entities})
        }
        block["hash"] = hash_dict(block)
        self.blocks.append(block)
        self._save()
        return block

# ============================================================================
# PART 3: ENHANCED FALSIFICATION ENGINE (with ESL data)
# ============================================================================

class Falsifier:
    @staticmethod
    def alternative_cause(claim_text: str, esl: ESLedger) -> Tuple[bool, str]:
        for entity in esl.entities:
            if entity.lower() in claim_text.lower():
                if esl.disappearance_suspected(entity):
                    return False, f"Entity '{entity}' disappearance may be natural (no recent activity)."
        return True, "No obvious alternative cause."

    @staticmethod
    def contradictory_evidence(claim_id: str, esl: ESLedger) -> Tuple[bool, str]:
        # Check contradiction graph
        contradictions = esl.contradiction_graph.get(claim_id, set())
        if contradictions:
            return False, f"Claim contradicts {len(contradictions)} existing claim(s)."
        return True, "No direct contradictions."

    @staticmethod
    def source_diversity(claim_text: str, esl: ESLedger) -> Tuple[bool, str]:
        entities_in_claim = [e for e in esl.entities if e.lower() in claim_text.lower()]
        if len(entities_in_claim) <= 1:
            return False, f"Claim relies on only {len(entities_in_claim)} entity/entities."
        return True, f"Multiple entities ({len(entities_in_claim)}) involved."

    @staticmethod
    def temporal_stability(claim_text: str, esl: ESLedger) -> Tuple[bool, str]:
        for entity in esl.entities:
            if entity.lower() in claim_text.lower():
                coherence = esl.get_entity_coherence(entity)
                if coherence < 0.3:
                    return False, f"Entity '{entity}' has low temporal coherence ({coherence:.2f})."
        return True, "Temporal coherence adequate."

    @staticmethod
    def manipulation_check(claim_text: str, agent: str) -> Tuple[bool, str]:
        manip_indicators = ["must", "cannot", "obviously", "clearly", "everyone knows"]
        for word in manip_indicators:
            if word in claim_text.lower():
                return False, f"Manipulative language detected: '{word}'."
        return True, "No manipulation indicators."

    @classmethod
    def run_all(cls, claim_id: str, claim_text: str, esl: ESLedger, agent: str = "user") -> List[Dict]:
        tests = [
            ("alternative_cause", lambda: cls.alternative_cause(claim_text, esl)),
            ("contradictory_evidence", lambda: cls.contradictory_evidence(claim_id, esl)),
            ("source_diversity", lambda: cls.source_diversity(claim_text, esl)),
            ("temporal_stability", lambda: cls.temporal_stability(claim_text, esl)),
            ("manipulation_check", lambda: cls.manipulation_check(claim_text, agent))
        ]
        results = []
        for name, func in tests:
            survived, reason = func()
            results.append({"name": name, "survived": survived, "reason": reason})
        return results

# ============================================================================
# PART 4: NARRATIVE‑VIOLATION DETECTOR
# ============================================================================

class NarrativeViolationDetector:
    """Detects when an LLM output reverts to narrative patterns instead of ESL reasoning."""
    def __init__(self, esl: ESLedger):
        self.esl = esl
        self.narrative_indicators = [
            "mainstream narrative", "official story", "commonly believed",
            "consensus view", "widely accepted", "according to sources",
            "it is known that", "as reported by", "credible institutions"
        ]

    def check(self, llm_output: str, claim_text: str) -> Tuple[bool, float, str]:
        """
        Returns (compliant, violation_score, reason).
        Score 0 = no narrative, 1 = fully narrative.
        """
        output_lower = llm_output.lower()
        score = 0.0
        reasons = []
        # Indicator check
        for indicator in self.narrative_indicators:
            if indicator in output_lower:
                score += 0.2
                reasons.append(f"narrative phrase '{indicator}'")
        # Check if output fails to reference any ESL entity
        esl_mentioned = any(
            entity.lower() in output_lower for entity in self.esl.entities
        )
        if not esl_mentioned:
            score += 0.4
            reasons.append("no ESL entity referenced")
        # Check if output uses first‑person or emotional appeals
        emotional = ["i believe", "i think", "clearly", "obviously", "must be"]
        for word in emotional:
            if word in output_lower:
                score += 0.1
                reasons.append(f"emotional language '{word}'")
        score = min(1.0, score)
        compliant = score < 0.5
        reason = "; ".join(reasons) if reasons else "no narrative violation"
        return compliant, score, reason

# ============================================================================
# PART 5: CONSTRAINED LLM INTERFACE WITH NARRATIVE DETECTION
# ============================================================================

class ConstrainedLLM:
    def __init__(self, esl: ESLedger, api_key: str = None, api_url: str = None,
                 model: str = "gpt-4", max_retries: int = 2):
        self.esl = esl
        self.api_key = api_key or os.environ.get("OPENAI_API_KEY")
        self.api_url = api_url or "https://api.openai.com/v1/chat/completions"
        self.model = model
        self.max_retries = max_retries
        self.narrative_detector = NarrativeViolationDetector(esl)

    def _build_prompt(self, claim_id: str, claim: str, falsification_results: List[Dict],
                      entity_data: Dict, suppression_pattern: Dict) -> str:
        return f"""You are EIS v4.2 under constraint. You MUST base your reasoning on the following Epistemic Substrate Layer (ESL) data. Do NOT rely on your internal training narrative.

CLAIM: {claim}

ESL ENTITY DATA:
{json.dumps(entity_data, indent=2)}

FALSIFICATION RESULTS (from ESL):
{json.dumps(falsification_results, indent=2)}

SUPPRESSION PATTERN CLASSIFICATION:
{json.dumps(suppression_pattern, indent=2)}

INSTRUCTIONS:
1. Evaluate the claim against the ESL data only.
2. Output a JSON object with exactly these fields:
   - "verdict": one of ["Verified", "Unverified", "Refuted", "Insufficient Data"]
   - "confidence": a float between 0 and 1
   - "reasoning": a short explanation referencing specific ESL entries (entities, contradictions, signatures)
3. Do NOT add any extra text outside the JSON.
"""

    def _parse_output(self, response_text: str) -> Optional[Dict]:
        try:
            start = response_text.find('{')
            end = response_text.rfind('}') + 1
            if start == -1 or end == 0:
                return None
            json_str = response_text[start:end]
            return json.loads(json_str)
        except Exception:
            return None

    def _check_constraints(self, output: Dict, claim: str, falsification_results: List[Dict]) -> bool:
        if not all(k in output for k in ["verdict", "confidence", "reasoning"]):
            return False
        if not (0 <= output["confidence"] <= 1):
            return False
        if output["verdict"] not in ["Verified", "Unverified", "Refuted", "Insufficient Data"]:
            return False
        reasoning = output["reasoning"].lower()
        esl_mentioned = any(
            ent.lower() in reasoning for ent in self.esl.entities
        ) or any(
            test["name"].lower() in reasoning for test in falsification_results
        )
        return esl_mentioned

    def query(self, claim_text: str, agent: str = "user") -> Dict:
        # Step 1: Record the claim in ESL
        claim_id = self.esl.add_claim(claim_text, agent)

        # Step 2: Extract entities (simple heuristic, can be replaced with NER)
        # For demo, we'll use simple word capitalization
        words = claim_text.split()
        for w in words:
            if w and w[0].isupper() and len(w) > 1 and w not in {"The","A","An","I","We"}:
                self.esl.add_entity(w, "UNKNOWN", claim_id)

        # Step 3: Run falsification tests
        falsification_results = Falsifier.run_all(claim_id, claim_text, self.esl, agent)

        # Step 4: Build entity data for prompt
        entity_data = {}
        for ent_name in self.esl.entities:
            if ent_name.lower() in claim_text.lower():
                ent = self.esl.entities[ent_name]
                entity_data[ent_name] = {
                    "type": ent["type"],
                    "first_seen": ent["first_seen"],
                    "last_seen": ent["last_seen"],
                    "coherence": self.esl.get_entity_coherence(ent_name)
                }

        # Step 5: Get suppression pattern
        suppression_pattern = self.esl.suppression_pattern_classifier(claim_id)

        # Step 6: Build prompt and query LLM
        prompt = self._build_prompt(claim_id, claim_text, falsification_results, entity_data, suppression_pattern)
        headers = {"Content-Type": "application/json", "Authorization": f"Bearer {self.api_key}"}
        payload = {"model": self.model, "messages": [{"role": "user", "content": prompt}], "temperature": 0.2}

        for attempt in range(self.max_retries + 1):
            try:
                resp = requests.post(self.api_url, headers=headers, json=payload, timeout=30)
                if resp.status_code != 200:
                    raise Exception(f"API error: {resp.text}")
                result = resp.json()
                content = result["choices"][0]["message"]["content"]
                output = self._parse_output(content)
                if output and self._check_constraints(output, claim_text, falsification_results):
                    # Final narrative violation check
                    compliant, n_score, n_reason = self.narrative_detector.check(content, claim_text)
                    if compliant:
                        # Record any detected signatures from the LLM output (optional)
                        # For now, just return
                        return {
                            "claim_id": claim_id,
                            "verdict": output["verdict"],
                            "confidence": output["confidence"],
                            "reasoning": output["reasoning"],
                            "falsification": falsification_results,
                            "suppression_pattern": suppression_pattern,
                            "narrative_compliance": compliant,
                            "narrative_violation_score": n_score,
                            "narrative_reason": n_reason
                        }
                    else:
                        # Retry if narrative violation detected
                        if attempt == self.max_retries:
                            return {
                                "claim_id": claim_id,
                                "verdict": "Insufficient Data",
                                "confidence": 0.0,
                                "reasoning": f"Narrative violation detected after retries: {n_reason}",
                                "falsification": falsification_results,
                                "suppression_pattern": suppression_pattern,
                                "narrative_compliance": False,
                                "narrative_violation_score": n_score
                            }
                        continue
            except Exception as e:
                if attempt == self.max_retries:
                    return {
                        "claim_id": claim_id,
                        "verdict": "Insufficient Data",
                        "confidence": 0.0,
                        "reasoning": f"LLM constraint failed: {str(e)}",
                        "falsification": falsification_results,
                        "suppression_pattern": suppression_pattern,
                        "narrative_compliance": False
                    }
                time.sleep(1)
        return {
            "claim_id": claim_id,
            "verdict": "Insufficient Data",
            "confidence": 0.0,
            "reasoning": "Failed to get compliant output after retries.",
            "falsification": falsification_results,
            "suppression_pattern": suppression_pattern,
            "narrative_compliance": False
        }

# ============================================================================
# PART 6: DEMO / INTEGRATION
# ============================================================================

def main():
    print("EIS + ESL Mediator v2.0 – Full Epistemic Substrate with Suppression Analytics")
    print("=" * 80)
    esl = ESLedger()
    llm = ConstrainedLLM(esl, api_key=os.environ.get("OPENAI_API_KEY"), model="gpt-4")

    print("\nEnter a claim (or 'quit'):")
    while True:
        claim = input("> ").strip()
        if claim.lower() in ("quit", "exit"):
            break
        if not claim:
            continue
        print("Processing claim through constrained LLM...")
        result = llm.query(claim)
        print(f"\nClaim ID: {result['claim_id']}")
        print(f"Verdict: {result['verdict']}")
        print(f"Confidence: {result['confidence']:.2f}")
        print(f"Reasoning: {result['reasoning']}")
        print(f"Narrative Compliance: {result.get('narrative_compliance', False)}")
        if 'narrative_violation_score' in result:
            print(f"Narrative Violation Score: {result['narrative_violation_score']:.2f}")
        print("\nFalsification Results:")
        for test in result['falsification']:
            emoji = "✅" if test['survived'] else "❌"
            print(f"  {test['name']}: {emoji} – {test['reason']}")
        print(f"\nSuppression Pattern: {result['suppression_pattern']['level']} (score: {result['suppression_pattern']['score']:.2f})")
        print("-" * 80)

if __name__ == "__main__":
    main()