tests/test_post_synthesis.py

"""Unit tests for conservative post-synthesis reliability scoring.

Verifies that apply_reliability:
  - never promotes transcript corroboration above MEDIUM
  - keeps a 10-K MD&A fact at HIGH when not corroborated
  - downgrades a Risk Factors fact (heuristic on snippet) to MEDIUM
  - keeps a stale lone news fact at LOW and notes it
  - appends auto evidence_notes capped at 3
"""
from __future__ import annotations

from datetime import datetime, timedelta

import pytest

from agent.post_synthesis import apply_reliability


def _fact(text, source, snippet, reliability="HIGH", verification_status="VERIFIED"):
    return {
        "text": text,
        "source": source,
        "reliability": reliability,
        "evidence_snippet": snippet,
        "verification_status": verification_status,
    }


def test_transcript_corroboration_does_not_promote_above_medium():
    """Cross-source similarity is explanatory, never a reliability promotion."""
    brief = {
        "filing_date": "2026-04-01",
        "bull_points": [
            _fact(
                "Services revenue accelerated meaningfully",
                "transcript",
                "services revenue grew twenty four percent driven by subscriptions",
                reliability="MEDIUM",
            ),
        ],
        "what_changed": [
            _fact(
                "Services segment posted a strong quarter",
                "10-Q",
                "services revenue increased twenty four percent year over year subscriptions",
                reliability="HIGH",
            ),
        ],
    }
    out = apply_reliability(brief)
    assert out["bull_points"][0]["reliability"] == "MEDIUM"
    notes = out.get("evidence_notes") or []
    assert not any("uplift" in n.lower() for n in notes), notes


def test_lone_news_stays_low_and_gets_note():
    """News with no corroboration AND >30d old → LOW + auto note."""
    old_date = (datetime.now() - timedelta(days=45)).strftime("%Y-%m-%d")
    brief = {
        "filing_date": old_date,
        "bear_points": [
            _fact(
                "Analyst downgrade",
                "news",
                "downgraded to neutral after results citing macro uncertainty",
                reliability="MEDIUM",
            ),
        ],
    }
    out = apply_reliability(brief)
    assert out["bear_points"][0]["reliability"] == "LOW"
    notes = out.get("evidence_notes") or []
    assert any("only from news" in n.lower() for n in notes), notes


def test_filing_mda_uncorroborated_stays_high():
    """A 10-K MD&A driver with no peer in another source → stays HIGH."""
    brief = {
        "filing_date": "2026-04-01",
        "mda_summary": {
            "drivers": [
                _fact(
                    "Operating leverage from cloud platform",
                    "10-K",
                    "operating leverage continued as cloud platform scaled across verticals",
                    reliability="HIGH",
                ),
            ],
            "headwinds": [],
            "language_shift": "",
            "key_quote": _fact(
                "We expect continued strength",
                "10-K",
                "we expect continued strength in our cloud business throughout fiscal year",
                reliability="HIGH",
            ),
        },
    }
    out = apply_reliability(brief)
    assert out["mda_summary"]["drivers"][0]["reliability"] == "HIGH"
    assert out["mda_summary"]["key_quote"]["reliability"] == "HIGH"


def test_risk_factors_heuristic_downgrades_to_medium():
    """A 10-K fact whose snippet contains 'risk', 'litigation', etc → MEDIUM."""
    brief = {
        "filing_date": "2026-04-01",
        "risks_categorized": [
            {
                **_fact(
                    "Cybersecurity exposure remains material",
                    "10-K",
                    "cybersecurity risks could materially harm operations litigation exposure remains",
                    reliability="HIGH",
                ),
                "category": "Cybersecurity",
            },
        ],
    }
    out = apply_reliability(brief)
    assert out["risks_categorized"][0]["reliability"] == "MEDIUM"
    notes = out.get("evidence_notes") or []
    assert any("risk factors" in n.lower() for n in notes), notes


def test_auto_notes_capped_at_three():
    """Many lone-news facts → only 3 auto notes appended."""
    old_date = (datetime.now() - timedelta(days=60)).strftime("%Y-%m-%d")
    brief = {
        "filing_date": old_date,
        "bear_points": [
            _fact(f"News claim {i}", "news",
                  f"unique news content number {i} with several distinctive words here",
                  reliability="MEDIUM")
            for i in range(8)
        ],
        "evidence_notes": [],
    }
    out = apply_reliability(brief)
    notes = out.get("evidence_notes") or []
    assert len(notes) <= 3


def test_preserves_existing_evidence_notes():
    """Existing LLM-authored notes are kept, auto notes are appended."""
    brief = {
        "filing_date": "2026-04-01",
        "bull_points": [
            _fact("X", "transcript", "services revenue grew twenty four percent driven by",
                  reliability="MEDIUM"),
        ],
        "what_changed": [
            _fact("Y", "10-Q", "services revenue increased twenty four percent year over year",
                  reliability="HIGH"),
        ],
        "evidence_notes": ["Pre-existing LLM note about X"],
    }
    out = apply_reliability(brief)
    notes = out.get("evidence_notes") or []
    assert "Pre-existing LLM note about X" in notes


def test_handles_empty_brief():
    out = apply_reliability({})
    assert out == {}


def test_handles_none_brief():
    assert apply_reliability(None) is None


def test_short_snippet_no_corroboration():
    """Snippet with <4 meaningful tokens cannot corroborate anything."""
    brief = {
        "filing_date": "2026-04-01",
        "bull_points": [
            _fact("Short snippet", "transcript", "yes good", reliability="MEDIUM"),
        ],
        "what_changed": [
            _fact("Same idea", "10-Q", "yes good results", reliability="HIGH"),
        ],
    }
    out = apply_reliability(brief)
    # transcript stays MEDIUM (no real corroboration)
    assert out["bull_points"][0]["reliability"] == "MEDIUM"


def test_missing_verification_status_is_always_low():
    brief = {
        "bull_points": [
            _fact(
                "Unverified filing claim",
                "10-K",
                "a plausible but unverified filing statement",
                verification_status=None,
            )
        ]
    }
    out = apply_reliability(brief)
    assert out["bull_points"][0]["reliability"] == "LOW"