Naven research · 2026-09-12

Screening a message before it sends, and what that costs

Almost every messenger moderates the same way: the message arrives, somebody reports it, and a decision is made afterwards. Naven checks the message first. That is possible for one reason — Naven is not end-to-end encrypted — and that reason is the whole trade.

The mechanism

A message written in Naven goes to the server before it goes to the recipient. It is checked against 10 published categories: scams, malware, harassment, bullying, hate speech, sexual solicitation, illegal transactions, doxxing, encouraging self-harm, and graphic violence. If it matches, it is not delivered, and the person who wrote it is told which rule it met. If it does not, it is written to the conversation.

Three layers do that work in order: deterministic patterns, then a hosted safety classifier, then a general model asked the same ten questions in words. The patterns are the only layer that is fast and certain; the models are the only layers that understand a sentence. A refusal from a model is only honoured when the message itself corroborates it — a classifier that answers "privacy" because it saw the word email is reacting to a topic, not to an exposure.

A second review runs overnight with the surrounding conversation as context, and can retract a message after delivery. That is the layer that catches what a single message in isolation cannot show.

The trade, stated plainly

Screening before delivery requires the server to read the message. That is incompatible with end-to-end encryption, and Naven does not claim both. Anyone whose first requirement is that no server can ever read their messages should use Signal — it is the better product for that requirement, and saying so costs Naven nothing it should want to keep.

What it misses

The screen is software written by one person. It misses things, including obvious things. It is materially weaker in languages other than English, because the rule-based half is written in English. Calls are not screened at all. Insults, arguments and swearing are deliberately delivered — an app that will not carry a disagreement is not protecting anybody, it is editing a private conversation. Somebody saying they are struggling is never blocked; the message is delivered and help offered alongside it.

What the numbers actually look like

Measured on 2026-09-12, across 150 accounts: 8 recorded refusals in total, against 7 accounts. That is the honest shape of moderation on a small platform — the screen runs on every message and almost never fires, which is what it looks like when the rules match what people were going to do anyway.

Over the same period Naven's newsroom published 1202 briefings across 24 channels, each one requiring five distinct established publishers before it would run. Both numbers are read from the production database rather than estimated.

Why publish the limits

Because the claim is only worth anything if the limits are attached to it. "Safe for every age" is a sentence a small messenger cannot keep, and a parent who reads it and later finds something that got through has been told something untrue by the product. A mechanism that can be checked — every message goes through the screen, here are the categories, here is what it misses — is the version that survives being tested.