import tweepy, pandas as pd, os

# Load credentials from environment variables
client = tweepy.Client(bearer_token=os.getenv("TWITTER_BEARER"))

query = "#midterms2024 has:videos -is:retweet"
tweets = client.search_recent_tweets(query=query,
                                     max_results=100,
                                     tweet_fields=["created_at","author_id","public_metrics"],
                                     expansions=["attachments.media_keys"],
                                     media_fields=["url","type"])
df = pd.json_normalize(tweets.data)
df.to_csv("raw_tweets.csv", index=False)
print(f"Fetched {len(df)} tweets")

import cv2, numpy as np, torchaudio
from deepface import DeepFace
from transformers import pipeline

# a) Visual artefacts (blink rate, lighting inconsistency)
def visual_score(video_path):
    cap = cv2.VideoCapture(video_path)
    frames = []
    while cap.isOpened():
        ret, frame = cap.read()
        if not ret: break
        frames.append(frame)
    cap.release()
    # Simple metric: average eye‑aspect‑ratio variance
    # (real implementation would use a pretrained eye‑tracker)
    return np.var([np.mean(frame) for frame in frames])

# b) Audio prosody check (synthetic vs human)
def audio_score(audio_path):
    wav, sr = torchaudio.load(audio_path)
    # Use a pretrained voice‑authenticity model (e.g., wav2vec2‑fine‑tuned)
    model = pipeline("audio-classification", model="microsoft/wav2vec2-base")
    return model(wav.squeeze().numpy())[0]["score"]

# c) LLM fact‑check of the caption/headline
def llm_check(text):
    verifier = pipeline("text-generation", model="google/flan-t5-xl")
    prompt = f"Fact‑check the following claim and return a JSON with fields: {{" \
             f'"claim": "{text}", "verdict": "true/false/needs_more_info", "source": "URL"}}'
    return verifier(prompt, max_new_tokens=200)[0]["generated_text"]

def risk_score(visual, audio, llm):
    # Simple weighted sum – adjust weights to your newsroom's tolerance
    return 0.4*visual + 0.3*audio + 0.3*(1 if "false" in llm.lower() else 0)

# Example usage
vs = visual_score("sample.mp4")
ascore = audio_score("sample.wav")
llm_res = llm_check("Senator X endorses Candidate Y")
final = risk_score(vs, ascore, llm_res)
print("Overall risk:", final)

import requests, json

SLACK_WEBHOOK = os.getenv("SLACK_WEBHOOK_URL")

def send_alert(tweet_url, risk):
    payload = {
        "text": f":warning: *High‑risk deepfake detected*\n>{tweet_url}\nRisk score: `{risk:.2f}`"
    }
    requests.post(SLACK_WEBHOOK, data=json.dumps(payload))

if final > 0.7:
    send_alert("https://twitter.com/user/status/12345", final)