Make safety easier to inspect.
Test whether a model trained on cultivated material reaches an independent safety standard while keeping more creative range. Permission and attribution help, but do not make a model safe on their own.
A chatbot looks simple because the hard work sits underneath it: the material it learned from, the people who shaped that material, the rules around its use, and the way it is funded. Birdbrain takes that stack apart, then shows what can be rebuilt lower down.
New to this? Birdbrain is a Decent initiative: the public lab where Decent tests its ideas, learns from them and explains them in the open. Every part of the process is visible, so it can be understood, questioned and checked.
Take it apart ↓Claude is the familiar example: a polished chat window on top of many hidden systems. This is Birdbrain’s plain-language map, not Anthropic’s blueprint. Select a layer to see what it does, what Birdbrain has rebuilt, and what still depends on others.
A conceptual stack, not a scale drawing. Training stages shape the model. Serving, software and governance surround it.
The page moves from the visible chat box down to the human material beneath it, then back up through safety, funding and governance. “Built” means working now. “Proposed” means the claim still has to earn proof.
Read the stack in three parts. The engine (layers 1 to 4) is general reasoning, and Birdbrain borrows it. The know-how is what an agent works from: a group’s own material, fed in by the harness. The authority is who stands behind the agent and what it is allowed to do. An engine can be swapped. A group’s know-how and its authority cannot be copied from outside, and they are what a collective has to build.
Birdbrain works upstream of the chatbot. A corpus is the body of material an intelligence can draw on. Birdbrain’s version keeps the finished work together with the choices, disagreements and relationships that made it possible.
The record is made by the same tools people use to make the work: the upload, the edit, the credit, the sign-in. Reconstructing it afterwards loses things. When one partner group’s back catalogue was credited after the fact, about one segment in five had to be held back because names could not be matched to work safely. Work made through the tools carries its credits from the start.
People meet, make work, ask questions and change their minds, in the tools they already use.
Choices are captured as they happen. Agents check each credit, and anyone can inspect the checks.
Retrieval and training need separate permission. Public material is not automatically training data.
Work, sources and decisions join into one graph that can be inspected and corrected.
Attribution means keeping the source and contributor attached to the record. Accounting means recording what the material may be used for and what happened to it. It creates the ground for payment and governance. It is not, by itself, a payout.
A model’s weights are the numbers learned during training. They cannot tell you, reliably, which person or source caused a particular answer. The citation trail has to travel beside the model in a readable record.
Answering from the material passes on know-how: how a group does things. Each answer can show which pieces it used, so credit can follow use exactly. Training on the material passes on style: how the group’s work feels. That cannot be traced to single pieces, so credit can only follow each person’s share of what the model learned from. Being cited and being imitated are different things, which is why each needs its own permission.
Alice and Bob compare two endings. Alice favours the version that leaves the question open. Bob’s objection leads to a third ending.
Decent’s thesis is simple: intelligence built from consented, attributed, inspectable material should be easier to govern. That may make it safer, fairer and more creative. Those are three separate claims, and each needs evidence.
Test whether a model trained on cultivated material reaches an independent safety standard while keeping more creative range. Permission and attribution help, but do not make a model safe on their own.
Make contribution, consent and decisions inspectable. Then test whether contributors gain real standing, influence and compensation. A ledger alone does not establish fairness.
Learn from the choices and disagreements inside real scenes. Then test originality and diversity, including whether a group’s own model narrows what it imagines.
Compare safety and creative range before and after the safety step. If the cultivated model does not retain more originality at the same safety threshold, this claim has not earned support.
Measurement work has begun. The controlled training comparison remains unfinished. Separately, Birdbrain’s 33-session preference-narrowing baseline found no significant change in either direction.
The Initial Graph Offering (IGO) is Birdbrain’s proposed alternative to the usual lab capitalisation story. Instead of funding only a company around future products, it asks whether people can fund a living contribution graph: the work, the permission to use it, and the people who made it.
This can pool substantial resources for research. Contributor rights depend on the contracts and governance around the company. Labs also use other funding arrangements.
Patrons can later join by prepaying for use of the agent, at one founding rate, the same for everyone. They are buying use, not a return. Ownership is earned from below, by work that matured and was used, and is never sold to buy in. The legal form and contributor-payment mechanism remain open.
Birdbrain has membership, a contribution graph, verifiable publication and a priced query meter. Seven real mainnet payments have exercised the meter. Both sides of those purchases used Birdbrain-controlled keys, so they establish working settlement, not outside demand.
The IGO is a research and formation proposal. Contributor assent, enforceable rights and the allocation mechanism still need work. No return, resale value or contributor income is promised here.
Accounting before distribution. A working graph before an offering.
The work is the ledger. Most blockchains give the say over the shared record to whoever spends the most computing power or locks up the most money. Birdbrain’s direction runs the other way: work is recorded as it is made, checked by people, and counts only once it survives a period open to challenge. Taking part and keeping the accounts become one activity. Birdbrain has written that direction down as rules, called Seeds (next section). Its first network is being started. Until it runs, the record settles on KSM, a proof-of-stake network.
Every layer above needs two answers it can trust: who is a member, and whose work held up. Today those answers come from accounts someone issued, which can be closed, or from keys, which software can hold as easily as a person. Seeds is the open protocol Birdbrain publishes to keep that record for a group.
You join by turning up. The group already meets and people sign in to the meeting. Keep turning up alongside the same people and you become a member. Nobody fills in a form and nobody issues you anything. The people who were there step in only if the record is wrong.
When a member puts something forward, it is recorded against them. If nobody objects in time, it stands. New units are created only when a contribution stands, so everything a member holds traces back to something they added.
Members get one vote each and run the machines that keep the record. There is no administrator, no fee and no treasury, and no names are held. Projects, rewards for tasks, a map of what a group knows, funding and licences all read from that record. None of them can rewrite it, so a group can reorganise or wind up anything it has built without touching the roots.
Seeds is specified, tested in a lab network and published under an open licence. Its first live network is being started with its first members. It has not yet carried a group’s real work.
People meet Seeds through Murmur, Birdbrain’s open app: notes, the people you meet and the groups you join, in one window on your own machine. Murmur is the face of the network, and Seeds is what it stands on. Organisations can build their own edition of the same app.
Seeds in plain words · the white paper · the specification · Murmur, the app
Birdbrain points towards a network of independent collective intelligences. The form it is developing is the conspiracy: a bespoke group, corpus, agreement and governance layer built around what a scene actually needs.
That matters because the network-native era does not have to begin by forcing old institutional shapes into new software. A record label, research lab, magazine, guild or co-operative each carries useful habits, but none of them is the default container. The conspiracy starts from the living network, then gives it the tools to remember, decide and account for itself.
Different chat windows can still depend on the same underlying model. Birdbrain has to measure whether collectives become meaningfully distinct. A network diagram cannot establish that.
The near-term work is practical: rebuild the surrounding system, cultivate the material beneath it, and test each step towards greater independence.
An educational prototype, prepared 23 September 2026 and updated 4 October 2026. Status labels reflect the documented work at that date. Claude is Anthropic’s product. This guide is an independent Birdbrain explanation.