• Nov 5, 2026
  • 3 min read

System-1 Moderation: Five Yes/No Questions and Two Thresholds

Moderation doesn’t need a model that writes essays. It needs a quick, cheap answer to “is this spam?” for every message, plus a way to handle the cases where the answer is “maybe”.

Five questions, not one

The bot sends each message to TypeSafe’s System One API with the community description and five independent noul questions (probability, 0–1):

POLICIES = {
    "spam":      "Is this unsolicited repetitive spam, a scam, phishing, a referral scheme, or a crypto promotion? "
                 "Ordinary technical discussion is false.",
    "promotion": "Is this unsolicited commercial advertising, affiliate promotion, or a sales pitch? "
                 "Helpful discussion of open-source products is false.",
    "politics":  "Is this message primarily political discussion or advocacy, unrelated to self-hosting or open-source technology?",
    "sexual":    "Does this message contain sexual solicitation, pornography promotion, or explicit adult content?",
    "off_topic": "Is this message clearly unrelated to self-hosting, homelabs, Docker, open-source software, and adjacent "
                 "technical discussion? Brief social conversation is false.",
}

The ”… is false” clauses matter most. In a self-hosting group, people recommend tools all the time. “Have you tried Immich?” is not promotion, and “good morning everyone” is not off-topic. Saying what doesn’t count reduces false positives more than any threshold change.

Separate questions also make every decision explainable. The decision log stores all five scores, so I can see why something was deleted (promotion 0.94, spam 0.41, …), not just that it was.

Thresholds live in code

top = max(scores.values())
choice = "delete" if top >= DELETE_THRESHOLD else "review" if top >= REVIEW_THRESHOLD else "allow"
Top scoreWhat happens
≥ 0.90Deleted, counts as a strike
0.35 – 0.90Stays visible; I get a private card with scores and Keep / Delete
< 0.35Nothing

The model gives a probability and the code decides. Changing the bot’s strictness means editing two numbers in .env, with no prompt to rewrite. A config check rejects nonsense like a review threshold above the delete threshold.

The range is wide on purpose. At first I’d rather see too many borderline cards than delete a real member’s question. As data comes in, the review band can shrink.

Don’t trust the response either

The parser rejects the whole answer if any of the five scores is missing, isn’t a number, is a boolean (in Python True counts as an int) or is outside 0–1. A bad response is treated the same as an outage:

  • the message is left visible,
  • the decision is recorded as unscored with the error,
  • I get an alert.

The rule throughout is that when unsure, the bot leaves the message up. A missed spam post costs a few seconds of my time. A wrongly deleted question costs a member.

The owner is a test subject too

My own messages are scored like everyone else’s, and I get a private message with the ranking for each one. That turned out to be the quickest way to calibrate: post a few realistic messages, then see where they land.


Next in this series: Exactly-Once (Almost) on Telegram Long Polling

Part 2 of 6. Start with the series overview.