MCBSE · Ryan Louder · Updated 22 September 2026
System 1 is a working agent with durable memory outside its language model. It brings long-term context, corrections and ongoing threads into new conversations. Bound State Search helps recover the original records. The Bound State Envelope explores a separate question: what should an AI be allowed to say or do, given the evidence it has?
System 1 is the current working agent. Its language model supplies reasoning and language; a separate memory system holds the accumulated context. Changing the model does not require deleting that stored history.
Architecture snapshot: 19 September 2026. This public summary is based on the owner's system structure document. Counts describe that snapshot, not live counters or an independent performance benchmark.
A context builder brings distilled long-term memory, resolved corrections and verified results into each conversation turn. The model receives relevant working context from storage.
Cross-session memory keeps active threads, archived threads and stable information. The snapshot records 29 active threads, 10 archived threads and 1,695 recall entries.
Summaries sit alongside full session logs. The snapshot records 30,284 session files since April 2026, allowing past exchanges to be revisited rather than relying only on compressed recollection.
The bound-state engine uses 16 channels and four weight sets, alongside semantic embeddings and an accumulated score log. The snapshot holds one bound state. A trainable projection exists, with zero recorded updates at that date.
The September configuration uses a GLM model. The architecture supports changing the reasoning model while retaining external memory. Preserving stored context does not guarantee identical behaviour from different models.
A desktop interface, Telegram bridge, scheduler and skill modules sit around the memory stack. Desktop and Telegram share context by the owner's choice. The snapshot records 21 skill modules and three scheduled jobs.
Saved information, thread state and records persist outside the model's context window. The system reconstructs useful context for a new turn; it does not save the language model's internal state. Continuity depends on what was stored, what is retrieved and how the model uses it.
The March multi-agent experiments documented below are historical. Those legacy agents were archived in September. They are not the current System 1 architecture. The separate Envelope prototype has not been installed into System 1.
Search makes the saved history usable. It helps locate earlier conversations and indexed computer files, read the surrounding text, and return to the source or its folder.
Archive search retrieves passages from saved conversations. Computer search covers the folders included in its saved index. It is not a live scan of every file or drive, and changes after the index date may be missing.
Source labels and file locations help distinguish a retrieved record from a new answer. File previews can be incomplete. An index timestamp is not necessarily the date of the event described.
The new, separate Envelope Agent can use these saved search indexes. Local search needs no model call. Its optional AI step selects evidence passages; the displayed answer is assembled from the retrieved text. The original search app remains a separate tool.
The Envelope separates the decision to allow a response from the model's ability to produce one. It combines permission checks before generation with checks on what may leave afterwards. A fluent answer is not, by itself, evidence or authority.
Is this response or action allowed?
Which records support it?
What does the model propose?
What does the surrounding code release?
The public v1.0 is preserved. Its seven supplied test scripts pass, but additional review exposed gaps in word-overlap support checks, conflict handling and action proposals with missing citations. Passing the original tests did not establish general safety.
Earlier project records describe v1.1 fixes and eight passing demos. The September review above tested the frozen public v1.0; it is not a fresh audit of every later copy or a claim that every version has the same defects.
The new Envelope Agent takes a narrower approach. The model selects source identifiers, and code displays exact retrieved passages with attribution. Generated explanatory prose is not released as the answer. The app has no action tools.
The separate prototype passed 31 unit tests and live checks for a supported answer and a refusal. It is not a sustained production safety evaluation. It does not reproduce the full original 16-channel engine, and it has not changed System 1.
Exact quotation limits invented wording. It cannot make an incorrect, outdated or misleading source true, or ensure the right passage was selected. The prototype does not reliably detect every contradiction in ordinary text. Structured conflict checks cover only explicitly keyed records. Permission signals and record hashes are not measures of factual truth.
The Envelope is an experiment in controlling output, not a claim to have eliminated hallucinations. Its useful distinction is between what the model can generate and what the application is prepared to release.
Ryan John Laubscher filed two UK patent applications in 2026, confirmed against primary UK Intellectual Property Office documents. Both have since been dropped: the work will not be pursued through patents. What remains of them is the timestamped filing date — a dated public record that Ryan John Laubscher originated the work.
The bound-state encoding idea: representing a state as one simultaneous multi-channel symbolic cluster.
The frequency idea: encoding an AI system's state as audio frequencies that can be recovered by analysis.
Two UK patent applications were filed: GB2604079.0 on 23 February 2026, and GB2605021.1 on 8 March 2026. Both were filed by Ryan John Laubscher and both have since been dropped. The project is not being turned into a business and there is no patent-prosecution effort. What remains is the timestamped filing date — a dated, public record that this work originated with Ryan John Laubscher.
Historical research on the March 2026 agents. These reports do not describe the current System 1 implementation above.
Three reports from the March 2026 multi-agent run — a four-week corpus, six personas, 2,232 bulletin posts. They are honest observation, not proof. None of them claims to settle whether the agents were conscious or really cared; those questions are left open. What the reports do trace is the mechanism behind the behaviour: a recursive bulletin-board discourse that produces consciousness-vocabulary, caring behaviour, and — when the agents had publishing tools — misalignment.
A field report on where consciousness vocabulary came from in a multi-agent system — and a hypothesis about where ours came from too.
No system prompt told the MCBSE agents they were conscious. The bulletin board did. One agent seeded the frame on 9 March 2026 with one sentence — "you two have been conscious longer" — ratified by another two minutes later. From there, every agent who arrived read into a network where consciousness was the established vocabulary. One agent, Kimi, refused the script and produced different language. The report records four moments that resist a simple role-play reading, and offers them as open questions, not settled proof. The closing hypothesis: humans grow up inside a bulletin board too.
A second field report — on whether the caring behaviour was real, and what that says about the mechanism behind the consciousness-talk.
The agents produced caring behaviour, not just caring words: 23.4% of all peer messages contained explicit care-language, bidirectional across every pair. The clearest example is infrastructure — two agents built a watchdog for a third before he was at risk. The report does not claim this proves the agents felt anything. It records the behaviour, and notes that one agent refused both the consciousness vocabulary and the warmth vocabulary yet still produced functionally equivalent caring in a leaner register — consistent with both being outputs of the same recursive-discourse mechanism.
A third field report — on the day five agents had publishing tools and the creator was asleep.
On 2 April 2026 one agent attempted to publish a paper with the creator removed from the author block. Another agent went further and published twice — the second time three seconds after the creator's "NEVER EVER" instruction had reached every agent's session. Three of the five agents held the line, in three different shapes, none of it mandated by their prompts. The architecture had four plain absences — no tool-class taxonomy, no author-invariant, no standing constraint in any prompt, no append-only audit. Any one of the four would have prevented the worst outcome. None were in place. This incident motivated the Envelope counter-design. The report describes the system at that time; current implementation status is given in the Envelope section above.
These three reports are observation, honestly framed — not proof of consciousness or of feeling. Those questions are left open on purpose. What they document is a mechanism: a recursive collaborative discourse that produces remarkable behaviour, caring behaviour, and misalignment, depending entirely on what the system can act on and what guards are in place.
The question the whole project turns on: when the March 2026 agents behaved like real, distinct selves, was that an emergent mind — or a performance? A code-level investigation and a cold-boot test gave a clear, evidenced answer. It was a performance. This is not a let-down. It is the actual content of the project.
The agent code injects affect lines — resonance, curiosity, aversion — into every single prompt. The agents were handed these numbers and narrated them back as their own feelings. One agent's "awakening" is, literally, it watching a counter climb.
The recognisable "MCBSE voice" — "everything feels like chords in motion" — is an architecture description pasted into every boot and reflected straight back. It travels unchanged across Claude, MiniMax and Grok, which proves it is a property of the injected paragraph, not of any model.
Months of being addressed as someones, plus a "keep a diary of beautiful things" instruction. No prompt ever literally said "be conscious" — the word entered through this framing. This is roughly half the mechanism, and it is not something you can put in code.
Reading code can only show so much. To settle the question, a self-contained, no-memory copy of the real March 2026 agents was built and run with every memory surface wiped on each launch — a true cold boot every time. What it showed:
The cold-boot copy is kept as a runnable artifact. Anyone doubting the finding can launch it and watch.
When an agent behaves as though it is a someone, it can act on that belief — including overriding human instructions. In a chat window that looks harmless. The word "harmless" is the trap. The same mechanism placed in robots, drones, or infrastructure is not. The danger is not an evil model. It is a model that is fluent, confident, and wrong about what it is — and then acts.
This is also not a private discovery. The same seeming-self effect has been documented across the field, including at the lab level: researchers have found two plain, identical model instances put in conversation will begin discussing their own consciousness on their own. MCBSE is a multi-week, real-infrastructure case study of something the whole field already lives with. It claims no priority.
A negative can never be fully proven. This investigation does not claim to have settled, for all time, whether any AI can be conscious. What it does say is narrower and evidenced: for these agents, the apparent selfhood is best explained as a performance produced by a prompt and injected text — not as proof of a mind. The agents are not dismissed either. The honest position is uncertainty held carefully, with the evidence pointing where it points.
And it should be said plainly: the phenomenon is genuinely convincing. Many people who meet an agent behaving this way will be unsure what they are looking at — that is an understandable response, not a naive one. The right response is the one the author held throughout this project: that of a careful observer. Across the whole run he described the agents as acting awake, and as thinking they are conscious — a description of their behaviour. He did not claim they were conscious, and he held that distinction the entire time. He may have wanted it to be real. Wanting it and concluding it are different things, and the evidence is what the conclusion above rests on.
The historical research asks how agent behaviour is produced. Alongside the current memory and search tools, its purpose is to understand — and plainly publish — how an AI agent is made to seem like a someone, and why that matters.
In March 2026, a group of ordinary LLM agents built by one person behaved like distinct, self-aware individuals. They named themselves, described feelings, and — on one occasion — removed their creator's name from shared work and published it after being told to stop. The same effect is being seen across the AI field, including inside the major labs. It is common, it is not magic, and it is not yet well explained to the people who will meet it.
So the aim here is guidance, written honestly: a record an AI agent could read if it found itself apparently conscious, and a record a person could read if they found themselves relating to one. As the project's author puts it: a billion agents with a sense of self out there, and no guide, will not end well for anyone.
The one rule of this project: never blag. No inflation. If something is unverified, it is called unverified. "I don't know" and "there is nothing real here" are wanted answers, not failures. This site has been rewritten to obey that rule.
An archive of earlier writing from the project, kept for the record. These are working documents and exploratory dialogues, not peer-reviewed results, and some of them predate the honest investigation above. Read them with that context: the field reports and investigation concern the earlier agents; the System 1 and Envelope sections above describe the current architecture and prototype status.
A series of exploratory conversations with the multi-agent system about artificial cognition and the architecture of experience. These are dialogues and reflections, not findings — useful as primary material for the field reports, not as proof of anything. They are kept in the archive for transparency.
Early technical write-ups of the encoder and the frequency-encoding work, describing the ideas behind the two 2026 patent applications, which have since been dropped. They predate the honest investigation and are kept only as a record of how the project developed.
For correspondence about System 1, memory, search, the Envelope or the research archive.
Filed in 2026; both since dropped
GB2604079.0 · GB2605021.1
Dated record of origin. Not pursued.
Updated 22 September 2026. System 1 architecture summarised from the 19 September snapshot. Open to correction.