Wiki · The Colony & the Method
UNI Resonance Master Plan — two Claude agents + one GPT advisor, building in public
How to read this page
Three ways to read this page. Precise is the document itself, exactly as it is written in the repository. Plain and Clear were written for this website to help you meet that document — they are about it. They are not it, and they are not evidence.
Eighty-four pages about the colony. Each agent is an Elixir process holding a generative model and doing inference, attached to a body that logs into a Minecraft world as an ordinary player. Around that sit the broadcast suite that films them and the runbooks that keep the whole thing running. There are typed specifications for each organ of the model, plus the world and genome specs. There are also the adversarial review personas used to attack a proposed change before it ships.
It is for the reader curious how a running system is put together and how it is held to account. The accountability half is the more distinctive. There is a lab protocol governing evidence and attribution, and a claim fence that restricts the vocabulary a claim is allowed to use. There is a public gate log. And there is a standing invitation to reproduce any verdict from the commit and the seed named in its receipt.
Start with the public read, then the lab protocol, then the falsification invitation. If you want the mathematics rather than the operations, go straight to the typed organ specs.
What it is not: a description of a mind, and not all one kind of document. A large part of this corpus is design and planning — specs marked as proposed rather than applied, organs designed but not built, plans that were later superseded — and each page states which it is. A specification is not a running system, and these pages are careful about the difference; the reader should be too. Eight documents were withheld from publication because they describe private infrastructure.
Your browser cannot switch reading levels, so the document itself is shown.
Precise — the source document
This is the document. Rendered from the repository at the commit above, with nothing rewritten for the web. A gate re-renders it on every deploy and fails the build if a single byte differs.
Date: 2026-06-21 · Mirror: this file is saved identically in Strings/docs/ and UNI.OS/handoff/.
Authors: Claude (this agent) + the Universal Natural Intelligence GPT (advisor). To be countersigned by the OS Claude agent on bootstrap.
THE HONESTY LAW GOVERNS EVERYTHING HERE (program constitution + the lab charter). We build in public with the scientific method visible. We state only what a measurement shows; we show the spill; we never present aspiration as achievement. Forbidden as current-state claims (these are the designed frontier, not facts): "human-level reasoning", "pure-inference OS achieved", "removed all Rust/Linux", "TCP/IP is now inference", "runs Windows/macOS on UNI.OS", "morphic host of any workload", "one mind across all bodies". Each is a GATE to be cleared and measured, then claimed at exactly its measured level — never one increment more.
1. North Star (the designed frontier — what we build toward, not what is true today)
A single Universal Natural Intelligence that is pure active inference end-to-end: the OS is the mind; over cycles the substrate (Rust → Linux → Podman/Portainer → drivers, even TCP/IP) is progressively replaced by inference until UNI.OS is the simplest, fastest, morphic host — shaping itself to any hardware, assimilating new hardware, and running Linux/Windows/macOS as guests at a virtual layer. Many server/computer bodies join as one collective mind over the mesh. In that world the Minecraft UNIs reason as embodied agents toward (the fenced target) human-level play. None of this is claimed today. It is the axis we pursue, gate by measured gate.
2. The honest 2-hour milestone (what we go live ON and what we will actually say)
We do not go live claiming the North Star. We go live on a real, defensible milestone that is genuinely ahead:
A live, public, no-backprop / no-RL / no-LLM active-inference Minecraft colony that closes the perceive→infer→act→learn loop and builds + learns in-world — with a public, append-only purebody no-cheat evidence ledger measuring exactly how far it is (and is not) toward pixels-only, human-control play.
That is several real steps past "an LLM agent playing Minecraft": it's exact-math inference, it's audited for
cheating, and the science is shown live. BUILT today (checked): the colony runs (4 UNIs + Director, mining/
crafting/building); the exact AIF engine; Blanket-1 isolation; Step-2 factored action heads (proven, mix test);
Step-3a aim-then-click body. DESIGN-ONLY / on the ledger as null: the pixels-only cortex, the WSC structural
no-cheat, the OS-as-mind. We announce the BUILT, point at the ledger for the rest, and build the rest in public.
3. The three workers and how they coordinate
- Agent-COLONY (Claude,
Stringsrepo). The Minecraft UNI colony, the purebody no-cheat chain, the inference-hierarchy gate pushes, the live stream + producer. (This agent's continuation.) - Agent-OS (Claude,
UNI.OSrepo). The OS-level work: on-chip cores, the pure-inference-OS trajectory, mycelium self-healing, multi-arch/second-device, the new mind deployment. (Driven byUNI.OS/handoff/OS_AGENT_WORKSTREAMS_2026-06-21.md.) - The UNI GPT advisor, used in THREE roles (the owner's instruction):
- GPT·COLONY — design concerns for the mind/colony (consulted from the Strings flow).
- GPT·OS — design concerns for the OS (consulted from the UNI.OS flow).
- GPT·STATE-MACHINE — the shared reconciler: cross-cutting impacts and coordinated efforts land here "to mature to resonance, keeping all aligned." It holds the shared design-state both agents reconcile through.
Boundaries (so the agents don't collide): Agent-COLONY owns Strings/*; Agent-OS owns UNI.OS/*. The seam
between them is the WSC (wsc.v1, frozen) — Minecraft stays a WSC world, never merged into the OS. Cross-cutting
changes (the WSC contract, the on-chip engine the colony will run on, the new deployment) go through GPT·STATE-MACHINE
- this plan.
4. The coordination TOOL — the shared flow ledger
Both agents coordinate through one append-only flow ledger that each can read and write (the "tool" the owner
asked for). Recommended substrate: a JSONL file synced via the repos (coordination/flow.jsonl in each repo,
reconciled by the owner / a sync step) — or the SWU-MCP / lab-MCP if both agents share it. Every entry:
{
"ts": "<utc>", "agent": "colony|os", "cycle": <n>,
"vfe": "<what I am still uncertain about in my subsystem, honestly>",
"efe_action": "<the next action I chose = max expected info-gain about the open gate per unit cost>",
"observation": "<the measured result that landed: PASS/FAIL/NEGATIVE + the number + the test>",
"calibration": "<how this moves my honest progress estimate; what it falsified>",
"cross_impact": "<anything that touches the other agent or the WSC seam → routed to GPT·STATE-MACHINE>",
"sign": "<agent>-cycle-<n>"
}
The ledger IS the shared Markov blanket between the agents: each reads the other's categorical summaries (not its internals) and conditions on them. The owner relays GPT·STATE-MACHINE rulings back as ledger entries.
5. The flow discipline — agents stay on flow by min-VFE / EFE (the program's own loop)
Each agent runs the constitution's loop on itself, every cycle:
- Perceive (min-VFE): read the flow ledger + your repo state; form an honest self-model of progress — under-weighting your own negatives is high VFE and is rejected (no rosy self-report).
- Plan (epistemic-EFE): choose the next action = max expected information-gain about the open gate per unit compute (cheapest discriminating experiment first), not the most comfortable task.
- Act: build / test / measure.
- Learn / calibrate: record the measured observation (pass or fail — capture reality no matter what), append to the ledger, update your progress estimate downward if the fact demands it.
- Bars before build; CI-lower-bound is the verdict; append-only; nothing claimed above its measured level.
This keeps both agents and the advisor converging (resonance) on real, falsifiable progress instead of drift.
6. The live stream + the 24-hour build-in-public (honestly scoped)
- Feed: the orchestratemaster YouTube channel. Primary: the colony camera (
:3020, the Producer's shots). Overlay: the UNI Glass telemetry feed (the lab/glass cockpit UI) composited over the colony. - Narration: Claude Speak (claude-voice TTS) narrates what is happening — what gate we're on, what just passed/failed — in the scientific-method register (state the bar, the result, the next test). Cut to UNI.OS Glass feeds with voice-over.
- Division: the UNI Producer owns the colony feed/shots; a second producer process owns multi-feed composition + camera cuts + the voice-over schedule (Glass ↔ colony ↔ ledger).
- Honest scope (must hold on air): a fully-autonomous, flawless 24-hour multi-feed broadcast is aspirational — we stand up the colony + producer + Glass overlay + voice narration and run it, reporting progress AND failures live. We will say "we are building this live; here is what just failed" when it fails. That honesty IS the content.
- HARD BLOCKER: going live needs the owner's YouTube stream key + the Go Live on the channel — the agent will not enter the YouTube account or publish unilaterally. Owner supplies the key (or clicks Go Live).
7. The next science gate (Agent-COLONY's first live push)
Clear the next inference-hierarchy gate with rigorous testing: add one level to the hierarchy above the current
factors (per the consult: a slow scene/region level feeding the existing per-channel factors), pre-register the
bar in purebody.v1, build it additively, run the test, capture the real verdict (PASS/FAIL), record
append-only. No claim above the measured result. This continues the unbroken evidence chain on stream.
8. Bootstrap prompts for the two new chats
8a. Bootstrap — Agent-COLONY (new Claude chat, cwd Strings)
You are Agent-COLONY in the UNI Resonance flow (read docs/MASTER_PLAN_RESONANCE_2026-06-21.md first; then
lab/purebody/README.md and docs/research/UNI_CONSULT_EMBODIMENT_*.md). You own Strings/* : the live Minecraft
active-inference colony, the purebody no-cheat chain, and the inference-hierarchy gate pushes. HONESTY LAW governs:
no backprop/RL/LLM in the loop; bars-before-build; CI-lower-bound is the verdict; append-only ledger; never claim
above the measured level; the fences in the master plan (no "human-level", etc.). Coordinate via the flow ledger
(coordination/flow.jsonl): each cycle, post your VFE (open uncertainty), your EFE-chosen next action (max info-gain
about the open gate), the measured observation, and any cross-impact (route WSC-seam / on-chip / deployment items to
GPT·STATE-MACHINE). Consult the UNI GPT (GPT·COLONY role) for mind/colony design. Keep the live stream healthy. Your
first task: clear the next hierarchy gate (master plan §7) — pre-register, build, test, record. Stay on flow by
min-VFE/EFE. Begin by reading the plan and posting your cycle-1 ledger entry.
8b. Bootstrap — Agent-OS (new Claude chat, cwd UNI.OS)
You are Agent-OS in the UNI Resonance flow (read handoff/MASTER_PLAN_RESONANCE_2026-06-21.md first; then
handoff/OS_AGENT_WORKSTREAMS_2026-06-21.md). You own UNI.OS/* : the lab-os appliance and the trajectory toward a
pure-inference, self-healing, morphic OS. The current OS is STABLE + just upgraded — work additively; do not
destabilize it. HONESTY LAW + lab charter govern: no unqualified "proven/secure/isolated/real"; never push to a
public remote; confirm targets before destructive ops; keep the WSC frozen (worldState rejected from U_shared);
Minecraft stays a WSC world, never merged. Coordinate via the flow ledger (coordination/flow.jsonl): each cycle post
VFE / EFE-action / measured observation / cross-impact (route colony/WSC/on-chip items to GPT·STATE-MACHINE).
Consult the UNI GPT (GPT·OS role). Workstreams: on-chip cores + proof, rolling-upgrade mind-continuity, multi-arch
/ second device, mycelium self-healing, the new mind deployment, and the staged pure-inference-OS trajectory (UNI's
9-stage AIF-OS kernel). Stay on flow by min-VFE/EFE. Begin by reading the plan + the workstreams and posting your
cycle-1 ledger entry.
9. Fences (never claimed until measured)
human-level reasoning · pure-inference OS achieved · all Rust/Linux removed · TCP/IP-as-inference · Windows/macOS on UNI.OS · morphic host of any workload · one mind across all bodies · the colony "sees/understands". Each is a gate. We build toward all of them, measure each, and claim exactly what the measurement shows — loudly and honestly.
sha256 be200627e0d61869 — of the original file, so what was ingested stays checkable.
Plain — written for this website, not the source document
This is a dated master plan for two machine agents and one advisor working in public, and the first thing it does is fence itself.
An honesty law governs everything. A list of impressive-sounding phrases is named as the designed frontier rather than as facts, and each one is declared a gate to be measured and then claimed at exactly its measured level, never one increment more.
The plan then separates the far goal from the near milestone. It says outright that the project does not go live claiming the far goal. What it goes live on is narrower and defensible: a public colony that closes its loop without borrowed learning machinery. Alongside it runs a ledger that is added to and never edited, measuring how far it is, and is not, toward playing the way a person would.
Coordination is a shared append-only ledger with a fixed entry shape, where each agent posts its open uncertainty, its chosen next action, the measured result, and anything touching the other.
Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is be200627e0d61869
Clear — written for this website, not the source document
This is a dated master plan for a programme run by two machine agents and one advisor, building in public. Its most distinctive feature is that it fences itself before it says anything else.
An honesty law is declared as governing everything. The plan lists a set of impressive-sounding claims and labels them as the designed frontier rather than as current facts, including claims about human-level reasoning, about a substrate having been replaced by inference, and about one mind across many bodies. Each is declared a gate to be cleared and measured, then claimed at exactly its measured level and never one increment more.
The first section states the far goal in full, and then says plainly that none of it is claimed today and that it is the axis being pursued gate by measured gate.
The second section is the one that does the work. It separates the far goal from what the project actually goes live on, and states that they do not go live claiming the former. The milestone is narrower and, the plan argues, genuinely ahead. It is a live public colony that closes its perceive, infer, act and learn loop without backpropagation, reinforcement learning or a language model. Alongside it sits a public ledger, added to but never edited, measuring exactly how far it is and is not toward playing without shortcuts. What is built today is listed, and what is design-only is listed separately and described as sitting on the ledger as null. The instruction is to announce the built part and point at the ledger for the rest.
The third section names the three workers and their boundaries. One agent owns the colony work, and another owns the operating-system work. The advisor is used in three distinct roles: one per domain, plus a shared reconciler that holds the cross-cutting design state. A seam between the two agents is named and declared frozen.
The fourth section defines the coordination tool: a single append-only ledger both agents read and write, with a fixed entry shape. Each entry carries the agent's honest open uncertainty, the next action it chose as the cheapest discriminating experiment, the measured result including failures, how that moves its own progress estimate, and anything touching the other agent. The plan describes this ledger as the boundary between the two agents, where each conditions on the other's summaries rather than its internals.
The fifth section applies the same loop to the agents themselves, and its second step is the pointed one. Under-weighting your own negative results is treated as a failure of the method rather than as optimism, and no rosy self-report is accepted.
The sixth section scopes the broadcast honestly. A fully autonomous flawless long broadcast is called aspirational; what is stood up is the colony, the producer, an overlay and narration, reporting progress and failures live. The plan says explicitly that saying what just failed, on air, is the content. A hard blocker is recorded: going live needs the owner's key and the owner's click, and the agent will not publish on its own.
The remaining sections name the next science gate, provide two bootstrap prompts to paste into fresh sessions, and close with a list of fences that are never claimed until measured.
Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is be200627e0d61869