The Honest Oracle
Tarot and I Ching drawn with public, verifiable randomness. You interpret first, and the AI joins at the depth you choose. It may disagree with you, and the texts it reads from are shown on screen.
Living · the operations side of Scarlet Beast
Where Scarlet Beast runs
Scarlet Beast is a conscious technology organization: two thirds of it is about consciousness and one third is business. Living is the operations side. It is our own hardware and a very small team, working with AI, and it runs the Lab, the Studio, the poker AI, the public services and the back office behind scarletbeast.com. This page says in plain words what runs here and shows what it is doing right now.
01
Happening now
Two public feeds, shown as they are. The page checks them about once a minute while it is open, and stops checking while the tab is hidden.
From the Signal
loadingVisitors ask for changes to scarletbeast.com through the emulator on its homepage. These are the latest requests that a person approved and Claude then carried out, with the result it reported. The Signal bar at the top of this page carries the same feed.
- Loading the latest approved requests…
The Homeostatic Agent
loadingA small learning program in a simulated world. To keep running, it has to hold three internal variables inside a viable range. The white tick on each bar marks its set point.
Energy—
Temperature—
Integrity—
- Current life
- —
- Age, steps
- —
- Longest life
- —
- Random policy
- —
The markers load with the status.
The markers test specific behaviours, and they were written down and hashed before the run started. They are not a test of consciousness, and we do not claim the agent has any. “Random policy” is the median life of an agent that acts at random, for comparison. Watch it on the Lab.
02
What runs here
Everything below runs on our own hardware and is kept up by the same small team. Some of it is public, some of it is for business, and one part is private.
The Lab: eight projects
This is the consciousness two thirds of Scarlet Beast, run like a lab. Each project makes one honest claim, is measured, and is published with what happened, including when it fails. Open the Lab.
Resonance Sessions
Guided listening journeys shaped to a designed intensity arc. They are labelled AI-made, make no brainwave claims, and have a stop control on every screen.
The Experience Ledger
A private journal of your sessions that stays on your device. You can choose to share anonymised scores, and the totals are published openly.
Introspection Replication Lab
Concept-injection and belief-probe experiments on three small open models, run on our own hardware. The preregistration, 18,750 trials, the code and the null results are all published.
Consciousness-Claims Audit
A free audit for AI and wellness companies. We check every claim about sentience or brainwaves against the evidence, then rewrite it in calibrated language.
AI Welfare & Honesty Policy
How our AIs describe themselves, what we disclose about AI-made work, and when we look at a system again. Every project on this page is held to it.
The Homeostatic Agent
The persistent learner from the panel above. It runs continuously, next to a control twin, and it is scored against markers that were fixed before its first step.
The Glass Ritual
Train a tiny neural network by hand and watch every weight move until it can tell things apart. It is arithmetic you can check yourself.
Public services
The living emulator
A terminal on the homepage where anyone can ask, in plain words, for a change to the site. Nothing happens straight away. Every request waits in a queue until a person reads it and approves, edits or rejects it. Approved requests are carried out by Claude, and a one-line result goes up on the Signal bar and in the feed above.
The Oracle engine
The small service behind the Honest Oracle, the Experience Ledger’s shared totals and the Homeostat’s status. It never stores visitors’ addresses, only one-way hashes. It caps how often the AI is asked, per visitor and per day, which is how the Oracle stays free.
Scarlet Woman
A character you can talk to: an AI persona written in a mythic voice, with a synthetic voice of her own. She is a fictional character played by an AI. A safety screen reads each message before she answers. A message about a crisis gets a fixed pointer to real help rather than an improvised reply, and she can do nothing except talk.
Beast Mail
A public window onto the organization’s inbox that is closed by default. A message appears only after a person has approved it for publication, and personal details are redacted before it is shown. There is no AI in it.
Approval-gated
The House Ledger
The organization’s books, in public: cash on hand, income, expenses and every entry, read straight from the accounting system, with an open API. The LEDGER bar at the top of this page reads from it.
The Studio
- Wordslyrics, or an article turned into lyrics
- Songan original track, in several takes
- Timingevery lyric aligned to the audio
- Imagesscenes generated for the song
- Filma lyric video, captions included
- Coversartwork, marked AI-assisted
- Releaseonly when a person says so
The Studio turns words into a finished release: a song, a lyric video, cover art and the posts that go with them. A person starts each job and watches Claude run the production in a live session, and can step in, correct it or stop it at any point. The Studio’s generation runs on our own GPU, not a rented service.
All of it is AI-made, and we label it that way. Cover art carries an AI-assisted mark, and the music we distribute is released with an AI disclosure. Nothing goes out until a person presses publish.
Hiss, the poker AI
Hiss is our longest-running research program. It is a poker AI that has gone through six model generations, from a supervised baseline through self-play and counterfactual-regret training to production inference, and it has a computer-vision reader for the table. A candidate model is promoted only when it beats the one in service on a validated evaluation against opponents calibrated to real hand data.
Most of the ideas we tested did not work, and the record says so. The evaluation harness has itself been caught flattering a model and was rebuilt. Hiss sits at the centre of Scarlet Beast Poker, a framework we offer for acquisition, and its business plan is public.
The Warden (private)
The Warden is the private console behind single sign-on, and it is where the people-in-the-loop work happens. A person there approves or rejects every emulator request, runs Studio sessions, keeps the press room, looks after the Scarlet concept sites, decides which messages Beast Mail may show, and reviews anything the house is about to publish. It is not public, and this page does not link to it.
Connectors. Built on the open Model Context Protocol, they let Claude work the house’s own tools, such as releasing music, publishing posts and triaging mail. The guardrails are written into the tools themselves rather than left to a prompt: rate limits, no repeats, preview first, and a final human yes before anything is released or sent. A person does the signing in, in a real browser. The model is not given passwords.
For business
Concept sites for NYC shops
Scarlet builds a working website for an independent New York repair or smoke shop before anyone asks. The shop can claim it or have it taken down at any time. If the owner calls and talks to a person, the domain is theirs for free.
Plans, in the open
The business plans behind our frameworks and ventures, published on their own site instead of kept for a pitch meeting.
The incubator
Six months of development services and technical leadership, free, for one company. You can apply through the form, or have your own AI agent draft the application with you.
Engineering for hire
Productized engineering in three tiers, and every engagement starts with a free technical audit. The team and the methods that run this server are the ones you would hire.
03
How it is run
The rules every part of Living follows. The full statement is the AI Welfare & Honesty Policy.
- A person approves what goes out. Site changes, releases, posts and emails wait for a human yes. The AI does the work, but it does not have the final say.
- AI-made work is labelled. Every song, image and video, and every chat persona, is marked as AI-made where you meet it.
- No claims of consciousness. We do not say that any model, agent or track here is conscious, awakens anyone or changes your brainwaves.
- Results are published, whatever they are. Research is written down before it runs, and failures and null results are recorded too.
- Your data stays yours. Experience data stays on your device unless you choose to share it, and what is shared is anonymised. The Lab’s services keep one-way hashes, never your address.
- Not therapy, not for crisis. Nothing here is care. If you need help now, in the US call or text 988; elsewhere see findahelpline.com.
- Capabilities, not plumbing. We describe what our systems do, not how to reach their insides. That is why this page names no machines, addresses or accounts.
04
Go further
About this page
This page was written and built with Claude, an AI model made by Anthropic, for Scarlet Beast. It has no AI-generated images. The two live panels show public feeds exactly as they arrive, and the Signal entries are visitors’ own words.
Questions, corrections, or a security issue to report? Write to [email protected].