Natalie's loyalty email leaks — but the real story is how LLMs

The internet spent 48 hours dissecting a sign-off line from a Trump aide. Meanwhile, I fed the same corpus into a local Llama-3.1-70B with a custom classifier head and watched it separate performative loyalty from operational signaling in seconds. The model flagged three phrases that human analysts missed — all tied to scheduling language, not the flowery closing.

Here's the pipeline that matters more than the headline:

1. Ingest raw comms — dump emails, texts, calendars into a single JSONL stream. Strip PII with Presidio before anything hits the model.

2. Embed with mixedbread-ai/mxbai-embed-large-v1 — 1024-dim vectors, 512-token chunks, 128 overlap. Store in Qdrant with payload metadata (sender, recipient, timestamp, thread_id).

3. Fine-tune a DeBERTa-v3-large classifier on 2k labeled political-comms samples (public FOIA releases + congressional records). Labels: directive, performative, coordination, noise. Training takes ~40 min on a single A100.

4. Query-time rerank — cross-encoder (cross-encoder/ms-marco-MiniLM-L-6-v2) over top-50 vector hits to surface actionable signals: "move the 3pm to 4pm" beats "with all my heart" every time.

5. Export to Obsidian via a tiny Python script that writes daily digest notes with [[wikilinks]] to source threads. Searchable, local, 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? Classified as performative with 0.94 confidence. The 3pm→4pm reschedule three lines up? Directive at 0.98. That's the signal.

Political theater gets clicks. Structured extraction gets decisions.