Changelog · Chazz.ai

A dated record of every change to Chazz.ai a reader would notice: new features, changes, fixes, security work and incidents. Also published as RSS and Atom.

The Chazz.ai changelog is a dated record of every change a reader would notice, newest first, including security fixes and incidents. The most recent entry is 2026-08-23: The site answers to machines properly. It is also published as an RSS and an Atom feed.

Changes a reader would notice, newest first, with a date on every one. Security fixes and incidents are listed alongside features rather than kept somewhere quieter.

This page is also published as a feed: RSS and Atom.

2026

2026-08-23: The site answers to machines properly

New. Every URL now returns a real title, description and readable body without JavaScript, and the site publishes a documentation corpus, a feed, an OpenAPI description and a written policy for automated readers.

This is a client-rendered application, which means that until now a plain fetch of any address returned the same empty shell: one title, one description and an empty container that only a browser could fill. Search engines that run JavaScript eventually saw the real thing. Nothing else did, including every language model that fetches a page to answer a question about it.

Every address now answers with its own title, its own description, its own structured data and a readable body for any client that will not run the application. The complete documentation is published as a single plain-text file so a model can read the whole product in one request rather than crawling forty screens. There is a written policy for automated readers, a machine-readable form of it, a vulnerability disclosure address, and a feed.

A set of new pages exists for the questions nobody could previously answer without signing up: what this is, what it costs, who runs it, what it does with your writing, and how to reach a person.

For AI agents
About
Pricing

2026-08-20: Your persona is a full character card

New. The reader's own persona grew from a short description into a complete character card with a portrait, so you can play as somebody specific rather than as a blank.

A persona used to be a name and a paragraph. It is now the same card a character is: appearance, personality, age, speech, a portrait and the rest, edited in the same editor.

Existing personas were carried across unchanged. Nothing needs re-writing.

Personas

2026-08-19: One classifier call per turn, and a record of every block

Security. Moderation now runs as a single batched classifier call per turn and writes every decision, with the score behind it, to a record that can be reviewed.

Screening previously made several calls per turn and kept nothing. It now makes one, and every block is written down with the score that caused it, so a decision that looks wrong can be found and argued with rather than shrugged at.

The batched path falls back to individual screening rather than skipping any item, which is the behaviour you want when the failure mode of the alternative is content that was never checked.

Safety and moderation

2026-08-19: Credits for a public post

New. A public post about the service that clears fifty likes or upvotes is worth a credit grant, claimed through a support ticket.

Post about it somewhere public, clear fifty likes or upvotes, send the link, collect the credits. The support queue also gained server-side paging in the same pass, which matters once there is a queue worth paging.

Pricing and credits

2026-08-14: Personalisation is real, and AI is disclosed where it is used

Changed. Measurement settings that were stored but never read now take effect, and every surface where an AI wrote something says so.

A handful of privacy and personalisation controls were being saved and displayed without anything reading them. They now bite. Where the answer changed, it changed toward collecting less.

Separately, the chat disclosure was appearing twice in some places and not at all in others. There is now one, in one place, on every surface that needs it.

Privacy

2026-08-14: A new visual system

Changed. The interface moved onto a glass and compact system with a retuned palette and a house icon set drawn for it.

Every page was swept onto the same surface treatment, the palette was retuned for contrast rather than for mood alone, and the borrowed icons were replaced with a set drawn for this product.

2026-08-10: A reply's facts can be corrected after it is written

New. Continuity correction fixes a reply that contradicts established facts without a second model call, and carries the correction into the next turn.

When a reply contradicts something the story has already established, the contradiction is corrected in place rather than left for the reader to argue with, and the correction is carried forward so the next turn does not repeat it. It covers ensemble scenes and multiplayer rooms, not only one-to-one chats.

2026-08-09: Fixed: the living world answered its own narration, and Pause never stopped it

Fixed. Narrated world beats were entering the transcript as the reader's own turn, and the pause control did not pause anything.

Two defects in the same feature. Beats the world narrated were being recorded as though the reader had typed them, so the world was replying to itself. And the pause control was a button that wrote a value nothing read, so a paused world kept running.

Both are fixed. A beat is an event in the world, never the reader's turn, and pause stops it.

Living worlds

New. Any character, scenario, lore pack or story can be shared by link, including a private one, with the token in the URL fragment so it never reaches a server log.

A shared link opens for somebody who has never signed in, which is most of the point of sharing. A private item's link carries its token in the fragment, so the part that grants access is never sent to a server and never appears in a log. Revoking a link makes it answer 410 rather than pretending it never existed.

2026-08-09: Fixed: chat illustrations drew a different scene

Fixed. Images generated from inside a conversation were composed from a key nothing wrote, so they illustrated a scene that was not the one on screen.

The image composer read a character field that no part of the product had ever written to, so every in-chat illustration was built without the character's appearance. It now reads the field that exists, and mature scenes generate as the scene rather than being quietly sanitised into something else.

Image generation

Elsewhere on Chazz.ai

This document is the server-rendered form of /changelog. The interactive page needs JavaScript, and shows the same content.

The complete documentation of this service is available as one plain-text file at /llms-full.txt. A short summary, with the operator's policy for automated readers, is at /llms.txt. Changes are published as a feed at /rss.xml.

Chazz.ai is operated by Cognitive Industries, ABN 62 794 528 747, Australia. Contact: [email protected].