Natalie's loyalty email leaks reveal the true power of LLMs in political communication.

For 48 hours, the internet focused on a sign-off line from a Trump aide. I ran the same corpus through a local Llama-3.1-70B equipped with a custom classifier head. Within seconds, it separated performative loyalty from operational signaling. It identified three phrases that human analysts had overlooked, all connected to scheduling language rather than the flowery closing.

This pipeline matters more than the headline:

1. Ingest raw comms — place emails, texts, and calendars into one JSONL stream. Use Presidio to strip PII before anything reaches the model.

1. Embed with mixedbread-ai/mxbai-embed-large-v1 — generate 1024-dim vectors using 512-token chunks and 128 overlap. Store them in Qdrant with payload metadata covering sender, recipient, timestamp, and thread_id.

1. Fine-tune a DeBERTa-v3-large classifier using 2k labeled political-comms samples drawn from public FOIA releases and congressional records. The labels are directive, performative, coordination, and noise. Training requires ~40 min on a single A100.

1. Query-time rerank — apply the cross-encoder (cross-encoder/ms-marco-MiniLM-L-6-v2) to the top-50 vector hits and surface actionable signals. “Move the 3pm to 4pm” always outperforms “with all my heart.”

1. Export to Obsidian with a small Python script that creates daily digest notes containing [[wikilinks]] to source threads. Everything remains searchable and local, with no cloud.

# quick ingest snippet
from pathlib import Path
import jsonlines
from presidio_analyzer import AnalyzerEngine
from presidio_anonymizer import AnonymizerEngine

analyzer = AnalyzerEngine()
anonymizer = AnonymizerEngine()

def clean_text(text: str) -> str:
    results = analyzer.analyze(text=text, language="en")
    return anonymizer.anonymize(text=text, analyzer_results=results).text

with jsonlines.open("comms.jsonl", "w") as writer:
    for raw in Path("raw_emails").glob("*.eml"):
        parsed = parse_eml(raw)  # your parser
        writer.write({
            "id": parsed.message_id,
            "thread_id": parsed.thread_id,
            "timestamp": parsed.date.isoformat(),
            "sender": parsed.from_,
            "recipients": parsed.to,
            "body": clean_text(parsed.body),
            "subject": parsed.subject
        })

The aide's sign-off was classified as performative with 0.94 confidence. The 3pm→4pm reschedule three lines up scored as directive at 0.98. That is the signal.

Political theater gets clicks. Structured extraction gets decisions.