Pandago/graphstrike-model-training

1#!/usr/bin/env python32"""End-to-end test for Round 2 implementation.3 4Tests:51. Platform-specific episode loading62. New tool actions (GET_POLICY, REVERSE_IMAGE_SEARCH, ANALYZE_BIO, CHECK_IP)73. Platform-adaptive scoring84. Hidden signals revelation9"""10 11from pathlib import Path12import sys13 14sys.path.insert(0, str(Path(__file__).parent))15 16from server.environment import FakeGangEnvironment17from models import FakeGangAction, ActionType18 19 20def test_round2():21    """Run comprehensive Round 2 test."""22 23    print("=" * 70)24    print("ROUND 2 END-TO-END TEST")25    print("=" * 70)26 27    env = FakeGangEnvironment()28 29    # Test 1: Instagram episode (even seed)30    print("\n[Test 1] Instagram Episode (seed=0)")31    print("-" * 70)32    obs = env.reset(task="easy", seed=0)33    print(f"✓ Platform: {obs.platform}")34    assert obs.platform == "Instagram", f"Expected Instagram, got {obs.platform}"35    print(f"✓ Steps remaining: {obs.steps_remaining}")36    print(f"✓ Starting visible: {len(obs.visible_account_ids)} accounts")37 38    # Test 2: GET_POLICY action39    print("\n[Test 2] GET_POLICY Action")40    print("-" * 70)41    action = FakeGangAction(action_type=ActionType.GET_POLICY)42    obs = env.step(action)43    print(f"✓ Message: {obs.message[:200]}")44    assert "Instagram" in obs.message or "threshold" in obs.message.lower(), "Policy not returned"45    assert obs.steps_remaining == 30, "GET_POLICY should not consume steps"46 47    # Test 3: INSPECT to find accounts48    print("\n[Test 3] INSPECT Action")49    print("-" * 70)50    acc_id = obs.visible_account_ids[0]51    action = FakeGangAction(action_type=ActionType.INSPECT, account_id=acc_id)52    obs = env.step(action)53    print(f"✓ Inspected: {acc_id}")54    print(f"✓ Steps remaining: {obs.steps_remaining}")55    assert obs.steps_remaining == 29, "INSPECT should consume 1 step"56 57    # Check that profile exists58    profile = next((p for p in obs.visible_accounts if p.account_id == acc_id), None)59    assert profile is not None, f"Profile for {acc_id} not found"60    print(f"✓ Profile created: fake_risk={profile.fake_risk_score:.3f}")61 62    # Test 4: REVERSE_IMAGE_SEARCH (hidden signal revelation)63    print("\n[Test 4] REVERSE_IMAGE_SEARCH Action")64    print("-" * 70)65    photo_before = profile.photo_reuse_score66    print(f"  Before: photo_reuse_score = {photo_before:.3f}")67 68    action = FakeGangAction(action_type=ActionType.REVERSE_IMAGE_SEARCH, account_id=acc_id)69    obs = env.step(action)70    print(f"✓ Steps remaining: {obs.steps_remaining}")71    assert obs.steps_remaining == 28, "REVERSE_IMAGE_SEARCH should consume 1 step"72 73    profile = next((p for p in obs.visible_accounts if p.account_id == acc_id), None)74    photo_after = profile.photo_reuse_score75    print(f"  After: photo_reuse_score = {photo_after:.3f}")76    print(f"✓ Signal revealed (changed: {photo_before != photo_after})")77 78    # Test 5: ANALYZE_BIO79    print("\n[Test 5] ANALYZE_BIO Action")80    print("-" * 70)81    bio_before = profile.bio_template_score82    print(f"  Before: bio_template_score = {bio_before:.3f}")83 84    action = FakeGangAction(action_type=ActionType.ANALYZE_BIO, account_id=acc_id)85    obs = env.step(action)86    assert obs.steps_remaining == 27, "ANALYZE_BIO should consume 1 step"87 88    profile = next((p for p in obs.visible_accounts if p.account_id == acc_id), None)89    bio_after = profile.bio_template_score90    print(f"  After: bio_template_score = {bio_after:.3f}")91    print(f"✓ Signal revealed (changed: {bio_before != bio_after})")92 93    # Test 6: CHECK_IP (expensive action)94    print("\n[Test 6] CHECK_IP Action")95    print("-" * 70)96    steps_before = obs.steps_remaining97    action = FakeGangAction(action_type=ActionType.CHECK_IP, account_id=acc_id)98    obs = env.step(action)99    print(f"✓ Steps consumed: {steps_before - obs.steps_remaining}")100    assert steps_before - obs.steps_remaining == 2, "CHECK_IP should consume 2 steps"101    print(f"✓ Message: {obs.message[:150]}")102 103    # Test 7: Snapchat episode (odd seed)104    print("\n[Test 7] Snapchat Episode (seed=1)")105    print("-" * 70)106    obs = env.reset(task="easy", seed=1)107    print(f"✓ Platform: {obs.platform}")108    assert obs.platform == "Snapchat", f"Expected Snapchat, got {obs.platform}"109 110    action = FakeGangAction(action_type=ActionType.GET_POLICY)111    obs = env.step(action)112    print(f"✓ Message: {obs.message[:200]}")113    assert "Snapchat" in obs.message or "threshold" in obs.message.lower()114 115    # Test 8: Platform-adaptive scoring116    print("\n[Test 8] Platform-Adaptive Scoring")117    print("-" * 70)118 119    # Reset to Instagram120    obs = env.reset(task="easy", seed=0)121    action = FakeGangAction(action_type=ActionType.GET_POLICY)122    obs = env.step(action)123 124    # Inspect and flag an account125    acc_id = obs.visible_account_ids[0]126    action = FakeGangAction(action_type=ActionType.INSPECT, account_id=acc_id)127    obs = env.step(action)128 129    profile = next((p for p in obs.visible_accounts if p.account_id == acc_id), None)130    print(f"  Account: {acc_id}")131    print(f"  fake_risk_score: {profile.fake_risk_score:.3f}")132    print(f"  status: {profile.status}")133    print(f"✓ Risk computed with platform-adaptive weights")134 135    # Test 9: SUBMIT with platform-specific rewards136    print("\n[Test 9] SUBMIT with Platform Rewards")137    print("-" * 70)138 139    # Flag gang members if we can identify them140    obs = env.reset(task="easy", seed=2)141 142    # Inspect a few accounts143    for acc_id in obs.visible_account_ids[:5]:144        action = FakeGangAction(action_type=ActionType.INSPECT, account_id=acc_id)145        obs = env.step(action)146 147    # Flag high-risk accounts148    flagged_count = 0149    for profile in obs.visible_accounts:150        if profile.fake_risk_score > 0.6 and flagged_count < 5:151            action = FakeGangAction(action_type=ActionType.FLAG, account_id=profile.account_id)152            obs = env.step(action)153            flagged_count += 1154 155    print(f"  Flagged: {len(obs.flagged_ids)} accounts")156 157    action = FakeGangAction(action_type=ActionType.SUBMIT)158    obs = env.step(action)159    print(f"✓ Episode complete: done={obs.done}")160    print(f"✓ Final reward: {obs.reward:.3f}")161    print(f"✓ Message: {obs.message[:200]}")162 163    print("\n" + "=" * 70)164    print("ALL TESTS PASSED ✓")165    print("=" * 70)166    print("\nRound 2 implementation verified:")167    print("  ✓ Platform-specific episodes (Instagram/Snapchat)")168    print("  ✓ GET_POLICY action (0 steps)")169    print("  ✓ REVERSE_IMAGE_SEARCH (1 step)")170    print("  ✓ ANALYZE_BIO (1 step)")171    print("  ✓ CHECK_IP (2 steps)")172    print("  ✓ Hidden signals revelation")173    print("  ✓ Platform-adaptive scoring")174    print("  ✓ Complete episode flow")175 176 177if __name__ == "__main__":178    test_round2()179