1"""Consensus types for multi-model verification.
2
3ModelAnswer / VerificationAudit / compare_answers extracted from the
4tical-code ortools core so this package has zero external dependencies.
5"""
6from collections import Counter
7from dataclasses import dataclass, field
8from typing import Dict, List, Optional
9import time
10
11
12@dataclass
13class ModelAnswer:
14    """One model's reply in a multi-model comparison."""
15    model_id: str
16    provider_name: str
17    content: str
18    tool_calls: Optional[List[Dict]] = None
19    latency_seconds: float = 0.0
20    error: Optional[str] = None
21
22
23@dataclass
24class VerificationAudit:
25    """All answers + consensus analysis."""
26    prompt: str
27    answers: List[ModelAnswer] = field(default_factory=list)
28    consensus: str = ""
29    divergence_score: float = 0.0   # 0.0 unanimous, 1.0 fully divergent
30    recommendations: List[str] = field(default_factory=list)
31    timestamp: float = field(default_factory=time.time)
32
33    def unanimous(self) -> bool:
34        contents = [a.content.strip() for a in self.answers if not a.error]
35        if len(contents) < 2:
36            return False
37        return len(set(contents)) == 1
38
39    def error_count(self) -> int:
40        return sum(1 for a in self.answers if a.error)
41
42    def best_answer(self) -> Optional[ModelAnswer]:
43        ok = [a for a in self.answers if not a.error]
44        return min(ok, key=lambda a: a.latency_seconds) if ok else None
45
46    def to_summary(self) -> str:
47        lines = [
48            "=== Verification Audit ===",
49            f"Prompt: {self.prompt[:100]}",
50            f"Models: {len(self.answers)}  Errors: {self.error_count()}",
51            f"Unanimous: {self.unanimous()}  Divergence: {self.divergence_score:.2f}",
52            f"Consensus: {self.consensus[:200]}",
53            "Per-model:",
54        ]
55        for a in self.answers:
56            status = f"ERROR {a.error[:80]}" if a.error else f"{a.latency_seconds:.1f}s"
57            lines.append(f"  {a.model_id} ({a.provider_name}): {status}")
58            if not a.error:
59                lines.append(f"    {a.content[:200]}")
60        if self.recommendations:
61            lines.append("Recommendations:")
62            lines.extend(f"  - {r}" for r in self.recommendations)
63        return "\n".join(lines)
64
65
66def compare_answers(answers: List[ModelAnswer]) -> VerificationAudit:
67    """Produce a consensus audit from a list of model answers."""
68    ok = [a for a in answers if not a.error]
69    audit = VerificationAudit(prompt="", answers=list(answers))
70
71    if not ok:
72        audit.consensus = "All models failed."
73        audit.divergence_score = 1.0
74        audit.recommendations.append("Check provider health - all models returned errors.")
75        return audit
76
77    if len(ok) == 1:
78        audit.consensus = ok[0].content
79        audit.divergence_score = 0.0
80        return audit
81
82    contents = [a.content.strip().lower() for a in ok if a.content]
83    unique = set(contents)
84    if len(unique) == 1:
85        audit.consensus = ok[0].content
86        audit.divergence_score = 0.0
87        return audit
88
89    audit.divergence_score = (len(unique) - 1) / len(ok)
90    majority_content, _ = Counter(contents).most_common(1)[0]
91    audit.consensus = ok[contents.index(majority_content)].content
92
93    if audit.divergence_score > 0.5:
94        audit.recommendations.append("High divergence - verify manually before acting.")
95    elif audit.divergence_score > 0.3:
96        audit.recommendations.append("Models disagree - add extra verification.")
97    return audit
98