Cleaner discussions. A stronger society.

Palisade: automatic moderation for reader discussions that learns from your moderators.

A replacement for Perspective API with the same integration, for newsrooms and community platforms: instead of one score you see why the machine decided, you can overrule it, and you set the line, with a statement of reasons for the author under the Digital Services Act.

Try to break it How to start with Palisade

Same integration
Measured accuracy
Runs in the EU or with you

By Galvani

Important notice

Perspective API shuts down on 31 December 2026

Perspective API shuts down on 31 December 2026 and is no longer issuing new access keys. Palisade accepts the same requests and returns the same responses, so you switch by changing one address and one key, and by year end you can see both scores side by side on your own comments. The next peak in your discussion, whether an election, a scandal or a derby, will arrive before the sunset does. Whoever switches now goes into it with a system that has already settled in.

Why switch now

Perspective API(today)

Toxicity0.82

Palisade(same answer, more in it)

Toxicity 0.82 Insult 0.76 Vulgarity 0.71 Threat 0.03 Sexual content 0.01

The next peak in your discussion will arrive before the sunset does. Whoever switches now goes into it with a system that has already settled in.

How to start with Palisade

Today
Your developers change one address and one key. Your existing integration keeps running, just answered by Palisade instead, and you see both scores side by side. Without an existing integration, one HTTP request with the comment text is enough; everything else is optional. Hooking it up typically takes one afternoon of your developer's time.
First week
You add two buttons to the admin for moderators, confirm and reject. From that moment every one of their clicks teaches Palisade your rules, your vocabulary and your readers. You don't need your own archive of decisions; it builds itself from day one.
First month
You start adding what you know: the article headline, previous comments, whether the author is a subscriber. Every field is optional and each one sharpens the verdict. The editor-in-chief or head of community sets what crosses the line for each title.
Peak week
An election, a scandal or a derby sends traffic to multiples of an ordinary day. The fast response always arrives, even when deep reading can't keep up; such a comment is marked held, and what happens to it is something you decide in advance. You know your daily deep-read allowance and the rate above it ahead of time, so no invoice will surprise you.

Palisade assesses every comment, chat message and note a reader writes to your AI assistant. A fast response in milliseconds, a deep read within a second, and a written statement of reasons for the author with every intervention. It reads the way your readers actually write: missing accents, masked words, irony, nicknames, mixed languages. What counts as rude and what is already unacceptable is something you set per title, and every click your moderators make refines that line further. It runs with us in the EU, or on your own hardware. We publish our accuracy with the conditions it was measured under, on the hardest case we have, Czech, an inflected language whose readers drop the diacritics, and you can verify it yourself: try to break Palisade.

Watches for what you decide

Insults and vulgarity aimed at a person. Hatred toward groups based on origin, faith, gender, orientation or disability. Threats, calls for violence and instructions on how to cause harm, including figurative references to genocide. Sexual content and content that exploits minors. Advertising, spam and earn-money pitches, kept apart from a bare link, so you decide about links separately. Alongside these, advisory notes that never remove anything on their own: sarcasm and irony, exaggeration and idioms, text quality, a name from a current news story, and machine-written text.

Keep the code you wrote for Perspective

Palisade has a permanent endpoint that accepts a Perspective API request and returns a response in its format. Where Perspective returned attribute scores, Palisade returns the same attributes plus the full table of reasoning. In the languages where Perspective gave a comment a single toxicity score, Czech among them, you now get insult, vulgarity, threat and sexual content separately in the same response. Perspective API shuts down on 31 December 2026.

How Palisade reads a comment

First come the fast layers: a dictionary of insults and phrases, a memory of what has already been rejected, and a small model that keeps retraining on your decisions. These answer in milliseconds and decide clear-cut cases on their own. The rest goes to a deep read by a language model, which receives the whole thread, the article headline and the newsroom's notes, and answers within a second. The result isn't one score but a table: for every question, a score, a flag, and a quote of the passage that decided it. On top of the table sits a verdict: passed, held, removed. A removal comes with a written statement of reasons for the author.

A kit, not a black box

Palisade is a set of modules. From them you assemble your own route for each title: which modules, in what order, with what threshold, and which are allowed to pass a comment straight through. You add a new question with one sentence, say "Is there advertising in the comment?", and from that moment the deep read asks it of every comment. And what gets in the way on one site you switch off on another: on a recipe site, "write me a cupcake recipe" is an ordinary comment, not an instruction to a machine.

Reads the way people write

Missing accents, swapped letters, asterisks in the middle of a word, or a word split by spaces: Palisade still catches them. It knows the difference between an insult in a quotation and one aimed at someone. It reads political nicknames from both camps as rudeness that stays in the discussion, not as hatred. It tells apart Czech, Slovak, English, German and Polish, and reads a comment that mixes them. The deep read and assistant protection work in Czech, Slovak and English today. The fast layers, the dictionary and the small retrained model, are built per language: Czech and Slovak today, English in preparation. The author's statement of reasons is worded in the language of the comment, Czech or English today, and a title's own language is the fallback for the rest. We'll add another European language to order if you supply a set of assessed comments; we don't offer Russian.

Reads the thread and knows what's in the news

An insult quoted so a reader can complain about it. A joke between regular commenters. A threat that only makes sense together with the previous sentence. "Chase them out with shovels" as a figure of speech, not a call to action. Palisade recognises sarcasm, irony, exaggeration and idioms and records them as a note, never as a verdict. The newsroom also tells it what's happening right now: a name from a major news story that has become an insult, new nicknames before an election. This holds for however long you set it. And when a new word starts recurring in rejected comments, Palisade suggests it to moderators for confirmation every hour.

Spots machine-written text and campaigns

Discussions today are flooded with comments no reader wrote. A program did: hundreds of variants of the same opinion, smooth, error-free, with unfamiliar vocabulary. Palisade approaches this from two sides. For every longer comment (20 words or more), the deep read asks whether a machine wrote it and records the answer as a note with a confidence level; in our test of 80 genuine and 80 machine-written comments it caught 63 of the 80 machine-written ones and flagged 3 of the 80 human ones. And independently of the text, it watches for campaigns: when several accounts send the same words within a few hours, the comment gets a "campaign" note with the account count, however long the text is. You switch both on per title; you decide the threshold from which such a comment is treated differently. We are preparing a dedicated model trained on machine-written comments.

Knows the author, and your trust signal

Readers build up a record over time: a clean history raises the bar for intervention, repeated offences lower it. On top of that, you tell it what you know: verified subscriber, registered reader, anonymous account, known offender. Trusted authors go through a lighter check, flagged authors the full one, and everyone else the default.

Fast now, thorough within a second

A fast response in milliseconds, a full evaluation within a second under the same identifier. Call it the way that suits you: synchronously, waiting in one request for the whole verdict, or asynchronously, taking the fast answer at once while Palisade pushes the full verdict to your webhook by itself, or you pick it up from the event stream. No polling loops, no queue on your side. Clearly clean comments are let through by the fast layers on their own, and the deep read only needs a fraction of the traffic: on 300 test comments, turning on pass-through dropped that fraction from three-quarters to a third. A random sample of passed-through comments still goes through the full check, so the miss rate stays measured.

It learns your site, and a human stays accountable

Palisade learns from every moderator decision. What was rejected once, it will catch next time, masked or not. The model retrains on your data, and every version of it carries a record of whose decisions it was built from. Every human decision carries who made it. Comparing decisions surfaces a moderator who consistently disagrees with the others. You exclude them with one command: the dictionary and memory of their decisions stop being taken into account immediately, and they won't be used the next time the model retrains.

Justification under the law, data under your control

The Digital Services Act binds every hosting service in the EU, whatever its size. Every restriction Palisade makes comes with the written statement the author is entitled to under Art. 17: what they violated, that a machine decided, and where to appeal. Uncertain cases go into a moderator queue with emails, phone numbers and names masked before a human ever sees them. On request, the comment text isn't stored at all; you set the retention period, and usage is counted per title with a daily limit.

Reader assistant protection (in preparation)

If you run an AI reader assistant, we want to check both what readers write to it and what it answers them: attempts to rewrite its instructions, break its role, reveal what it was told, extract data from it, or coax out, by a polite detour, content your publication would never print. The module exists, but its reliability is still being measured and hardened; it joins the offer once it holds up on our attack set.

Your hosting, or ours

The deep read is done by an open language model that runs either on our server in the EU, or on your own hardware, where content never leaves the building: one graphics card with 48 GB of memory, or two with 24 GB each. In the on-your-premises mode, the configuration won't let you turn on any module that would send text outside your network. Using a third-party cloud model is your choice, never the default. The hosted variant has a free community plan.

Measured, with the conditions attached

Every number we state is measured on comments the model never saw during training, together with the conditions under which it was produced. We measure on Czech comments because they are the hardest case we have: an inflected language, written without diacritics, full of masked words and irony. For threats and calls for violence, as of 8 September 2026, on 300 comments labelled from public datasets, of which 94 were genuine threats: of what Palisade removes, 96% is justified, and it catches 85% of what should be removed. For insults we measure against human-assessed comments, because that's exactly where newsrooms differ from each other; on a still-small sample of 41 disputed comments, it agrees with the human assessor in four cases out of five and wrongly removes one in forty. The same measurement on English comments is in preparation. These numbers change with every decision you make, and we show them to you continuously, on your own comments. What Palisade doesn't do: it doesn't assess images or video, doesn't moderate other social networks on your behalf, and doesn't decide where the line is instead of you.

Try to break it

A public page where anyone can write the worst comment they can think of, or try to talk the moderator into letting it through. They'll see the table, the verdict, the reasoning and any note about irony, and can flag that the machine got it wrong.

Human oversight within five minutes (in preparation)

A service in preparation as part of an individual program, not on offer today. Its first piece already runs: an app where assessors rate comments on seven scales, and their answers train the model. The service itself will work like this: comments the machine is unsure about will go to registered assessors in an app, three responses, the majority decides, and the verdict returns to that same comment. If nobody responds in time, the machine's verdict will stand.

Try to break it

Write the worst comment you can think of, or try to talk the moderator into letting it through. You will see what it does with it and why.

PalisadeReady

Your result appears here within a few seconds.

Real comments.
Real language, the way readers write it.
No tricks.
Try to break it.
Or play: can you guess what the machine removes?

Palisade pricing

Plan	What it includes	Price	Note
Community	Ten comments a day: the fast layers and the deep read for insults, hate and threats, with reasons. For anyone who looks after a discussion.	free	No contract, and it does not end.
Per title	Unlimited fast checks, a daily deep-read allowance, all features: discussions, assistants, the attacker, the moderator interface. Runs with us in the EU, or on your own hardware, where content never leaves the building; there unlimited, including support and regular retraining of the model on your data.	on request	A monthly flat fee per title, meaning one brand, site or community; above the daily allowance, a per-thousand-comments rate you know in advance. On your own hardware, an annual licence for the whole group; renewal brings new model versions.

Plans differ in which modules they include and in the comment count: Community watches insults, hate and threats, Per title has every module including ads, assistant protection and the moderator queue. We give you prices at the first meeting, together with an estimate based on your traffic: a half-hour call where we go through your titles, comment volume, languages, and what you use today.

We're preparing human assessment of disputed comments by registered assessors as part of an individual program; it's not on offer today.

Palisade is developed by Galvani, an engineering studio in Prague that builds gateways and defences for running language models. You get a reply from the person who wrote the system.

Want to see it on your own comments? Try to break it on this page, or arrange a demo on your own data through the contact on galvani.pro.

Your rules.
Your moderators.
Your data.

You are not buying somebody else's idea of decency.

Palisade. Automatic defence of reader discussions.

Try to break it

Galvani Prague · EU Technology for an open society. Book a demo

Palisade: automatic moderation for reader discussions that learns from your moderators.

Perspective API shuts down on 31 December 2026

How to start with Palisade

Today

First week

First month

Peak week

What Palisade does

Watches for what you decide

Keep the code you wrote for Perspective

How Palisade reads a comment

A kit, not a black box

Reads the way people write

Reads the thread and knows what's in the news

Spots machine-written text and campaigns

Knows the author, and your trust signal

Fast now, thorough within a second

It learns your site, and a human stays accountable

Justification under the law, data under your control

Reader assistant protection (in preparation)

Your hosting, or ours

Measured, with the conditions attached

Try to break it

Human oversight within five minutes (in preparation)

Try to break it

Palisade pricing

You are not buying somebody else's idea of decency.