AI moderation grew a settings panel

AI moderation shipped with six built-in topics — scams, phishing, gambling and company. That covers what bothers everyone, but every chat has its own list of what does not belong, and it rarely matches anyone else's. Over six weeks the block grew everything it was missing.

Your own rules in plain words#

The main addition. On the Rules tab you can add your own alongside the built-in topics. A rule is a description of what to catch, not an instruction to the bot: what happens to a match is decided by the punishment chain, same as everywhere else.

It is written as plain text, with no syntax — no brackets, no connectors, no escaping:

• messages with profanity or insults;
• job and earning offers, including "DM me" and links to job channels;
• sale listings for anything other than fishing tackle and bait;
• money collections, card details and requests to chip in;
• discussion of politics, elections and war news;
• ads for other chats and channels, invitations to join another community;
• getting personal — discussing members themselves rather than their messages.

Exceptions go in the same place, in the same words: "talking about your own job is not a violation." The model understands your wording the way it understands a built-in topic, and catches it in any rephrasing just the same.

One limitation matters: the model only sees the message itself. It does not know what your chat is about, who wrote it or what day it is, so wordings like "off-topic" mean nothing to it — spell out what counts as a violation. A rule has 500 characters, which is plenty.

Order, sensitivity and the simulator#

Your rules are checked in a separate request, and only if the message did not land in one of the built-in scam topics — and top to bottom, in the order you arranged them. The order is changed with arrows, so a more important rule can be moved up.

Each rule has its own sensitivity: "Precise" reacts only when the model is confident, "Sensitive" also reacts on medium confidence. Profanity suits the second, "getting personal" rather the first.

To check your wording without waiting for real spam there is a simulator: you paste an example message and see which rule would have fired. No punishment is applied, but the check is billed like an ordinary message, and built-in topics do not take part in the simulator — only your rules.

A punishment of its own for every rule and topic#

A rule has a name, and from there it lives its own life: its own punishment chain, its own violation counter, its own line in the log. Built-in topics got the same: phishing can earn an immediate ban while a gambling ad gets a warning, and those violations never add up into one counter.

The moderation log now shows the name of the specific rule or topic instead of a generic "AI moderation fired". Until punishments are set, a rule does not fire — and the dashboard says so outright rather than leaving a silent rule switched on.

Keeping the spend down#

The block costs money as it works, so there are now settings that decide whether to call the model at all:

• minimum message length — "ok" and "thanks" never reach the model;
• skip what has already been caught — if phrases, links, regex or length already cut the message off, the model is not called by default; a separate checkbox makes it check anyway and possibly issue a second, independent punishment;
• albums — switched on separately, because an album of ten photos is ten checks, not one.

Keep the number of rules to what you actually need, too: each one takes part in checking every message and makes processing more expensive. Hence a ceiling of twenty rules per chat, and a warning in the dashboard about the rising cost rather than a quietly larger bill.

You do not pay twice for repeats#

Spammers send the same message in dozens of copies. Paying for each copy makes no sense, so checking identical text costs the same as checking it once: a repeat gets the ready answer and never reaches the model. Those checks are marked separately in the balance history, so the saving is visible.

Auto top-up#

A balance that ran out at night means a chat without AI moderation until morning. You can now set a threshold and an amount — when the remainder approaches the lower bound, the balance tops itself up. A low-balance notification arrives either way.