SmartCiti.X Agentic Worlds — Whitepaper

Moved since this explainer was written: Evals run: the text says 1,025, worldforge/out/index.json#totals.evals now reads 1,066. The prose below keeps the explainer's wording; the registry value is the current one.

The network starts as a list. my-fabric/worlds/index.json names 41 worlds, and WorldForge carries every one of them into worldforge/out/index.json, where each world gets its Master Cards, a blockout, a card-tagged glTF file and a manifest. All 41 are marked "coming soon". All 41 use the same management model, the Agent-managed DAO, with three steward seats and the same six actions reserved for people. Every world names its owner, AGI Future Foundation PBC, and carries the umbrella "Powered by AGI Corp".

The worlds fall into four families, and the family decides where a world's content comes from. The six PBC focus DAOs (Cancer, Robotics, Ecology, Climate, Education and Public Health) come from the Global Issue Focus in the foundation's mission, recorded under focus_domains in worlds/PBC.json. The 28 W3 LLC series worlds, slugs s04 to s35, come from the founder's series list in worlds/series.json. They have a name and a series number and nothing more yet, so worlds/focus.json records "focus and RP1 location not yet given" for each one. The three program-track worlds (Agentic Payments, Agents That Deploy Infrastructure and Autonomous Organizations) are built only from three track descriptions the founder pasted on 2026-09-28, and every zone and agent in them cites a line of that text. The four earlier worlds (Locator.X, Biota.blockchain, Empathic.X and Diagnostic.X) came before the shared template and draw their zones from their own source documents.

The families overlap with the W3 LLC series list, and the overlap is worth stating exactly. worlds/series.json lists 35 series. Thirty-three of them are among the 41 worlds: the 28 named-only worlds, the four earlier worlds (Series 2, 3, 8 and 18), and Robotics DAO, which is both a focus DAO and Series 30. The other two, Cognition.X (Series 1) and SmartCiti.X (Series 10), are fabrics that were already built. They are listed separately under "built" in my-fabric/worlds/index.json and are not counted in the 41.

A named-only world has no charter yet, so its spaces follow the shared PBC template: Charter Hall, Research Synthesis Lab, Proposal Forum, Operations Deck and Progress Observatory, plus the Steward Council. That is why 28 worlds look alike, and it is deliberate. The template shows the governance shape without inventing a mission. When the stewards write a focus line into worlds/focus.json, it flows into that world's spec. Until then the world's Omega Steward says plainly that its focus is not set.

Across the network WorldForge counts 939 Master Cards, 242 specialist agents, 246 zones and councils, 1,103 blockout primitives and 287 roads. It runs 1,025 evals, and none fails. A rule, not a hand choice, gives each family its own civic landmark style: setback towers for the 28 series worlds, a rotunda for the six focus DAOs, a twisting tower for the three tracks and an octagonal prism for the four earlier worlds. What makes this a network rather than 41 separate scenes is the shared rulebook, the shared building and avatar sets, and the handoffs between worlds. Each Omega Steward has one scripted handoff that sends a visitor to another world's Omega when its own world lacks the service asked for.

• PBC focus DAOs: 6 worlds, 138 cards, 36 specialist agents, 0 anchored on RP1.
• W3 LLC series worlds (named only): 28 worlds, 644 cards, 168 specialist agents, 0 anchored.
• Program tracks: 3 worlds, 68 cards, 17 specialist agents, 0 anchored.
• Earlier worlds with their own sources: 4 worlds, 89 cards, 21 specialist agents, 4 anchored.

Moved since this explainer was written: Native name-limit flags in this mount plan: the text says 1, worldforge/native/conformance-report.json#perPlan[slug=cancer-dao].nativeViolations now reads 0. The prose below keeps the explainer's wording; the registry value is the current one.

Every world follows the same layout pattern, so one walk shows it. Cancer DAO is a focus DAO; its card calls it a mission-focused DAO for cancer research support, agent-managed and human-governed. Its world card covers a Volum 300 m wide, 300 m deep and 80 m high, and the world holds 23 cards: the world itself, a plaza, a coordinator, three advisors, a gate, a council with three seats, an Omega Steward, five zones and six agents. Distances are card centres in metres, from worldforge/out/cancer-dao/cards.json.

The gate. You arrive at the Alignment Evaluation Gate, 130 m from the plaza centre on the world's main axis. It is 20 m wide and 12 m high, and its card gives its purpose as pre-registered safety evaluation before deployment, citing worlds/PBC.json. Past the gate, five zone buildings stand on a 90 m ring around the plaza. The first two you pass are Charter Hall and the Progress Observatory, one on each side of the axis.

The plaza. The plaza is a 48 m square pad at the centre. The Cancer DAO Coordinator stands 8 m from its centre on the gate side, facing arrivals. It runs a hub-hybrid swarm: it supervises, routes each question to a specialist and answers whatever it cannot route. Its recorded swarm evals pass, including routing accuracy 1.0 on 4 test questions against a 0.8 threshold, at 1.25 model calls per turn. Four questions is a small test, and the record shows the count. On the council side of the plaza stand three advisors: the Host Advisor, the Steward Liaison and the Alignment Advisor. They are scripted guides, not AI models. Every line they can say is either written out with the record it cites or read at view time from a named registry path. None of it is legal, financial or medical advice.

The zones. Each zone is a building on a 36 m square footprint, with exterior textures at 2048 px and interior textures at 1024 px. Charter Hall holds the charter, contribution model and measurable outcomes; it has no agent of its own yet. The Research Synthesis Lab houses the Evidence Monitor and the Research Synthesis Agent. The Proposal Forum, on the axis beyond the plaza, houses the Proposal & Grant Drafter, whose role says it drafts for human review and never enacts. The Operations Deck houses the Collaborator Scout and the Task Router, and the Progress Observatory houses the Progress Tracker. Each agent card is bound to its swarm entry and to the world's metaprompt, and each lists the six reserved actions under may_never.

The council. At the end of the axis, 130 m beyond the plaza centre, stands the Steward Council: a forum building 28 m wide whose card admits stewards only. Inside are three human seats, 6 m apart, and each one reads seated: null. Twenty-six metres to the side, outside the council's Volum, stands the Omega Steward, defined, not seated. Its voice is synthetic (Kokoro, run locally), and its greeting says so: it is an agent, not a person. It then answers six fixed questions from the records. Asked who decides, it says the Steward Council does, by 2 of 3 signed votes, and that publishing the world is the founder's call.

The landmark. WorldForge draws a 51 m civic rotunda for Cancer DAO, synthetic and procedural, and suggests a spot 14 m off the axis and 33 m from the plaza centre toward the council, turned to face the plaza with 3.13 m of clearance. The build does not place it yet.

Moved since this explainer was written: Evals network-wide: the text says 1,025, worldforge/out/index.json#totals.evals now reads 1,066. The prose below keeps the explainer's wording; the registry value is the current one.

A Master Card is the unit of record in WorldForge. The card layer carries the SmartCiti.X name: every card is a SmartCiti.X Master Card. Each card carries an id, a kind, a parent, an owner (AGI Future Foundation PBC), the umbrella, a source citation, a Volum, a permission matrix and the list of reserved actions. A world's cards form one tree, root first, and the same tree is written into the world's glTF file. Every node carries its cardId and parentCardId in its extras, so a loader can index cards without trusting display names.

There are ten kinds. Across the 41 worlds the counts are 41 world, 41 plaza, 41 coordinator, 123 advisor, 41 gate, 41 council, 123 seat, 41 omega, 205 zone and 242 agent cards, 939 in all. Each world has one of each singular kind, three advisors and three seats. Zones and agents vary with the world's spec.

A Volum is the bounded 3-D region a card covers: a box given as a centre and half-extents in metres, Y up, in the world's frame. Volums do real work. They place buildings and avatars, and evals read them. One check confirms that each Omega Steward's Volum sits clear of the council's, so the Omega stands beside the council and never inside it. The geometry is synthetic placeholder massing, not surveyed land.

The permission matrix has seven rights: view, enter, build, operate, delegate, transfer and govern. Across all 939 cards it takes only a few shapes. View is public everywhere. Enter is public except on the council and its seats, which only stewards enter. Build and delegate belong to stewards. Operate belongs to stewards, and agents gain it in one form only, "agents:zone", on plaza, zone, coordinator and agent cards: an agent operates its own zone and nothing else. Transfer and govern belong to the steward quorum on every card.

On top of the matrix, every card carries the six actions the PBC reserves for people, copied verbatim from worlds/PBC.json: governance.final_vote, charter.amend, treasury.transfer, proposal.enact, windfall.distribute and steward.appoint. Each is held for the Steward Council, with a 2-of-3 quorum over 3 seats. That makes 5,634 reservation entries across the network, all identical. Agent, coordinator and advisor cards also list the same six under may_never. This is the working form of the line every cards file carries: agents operate; people govern.

WorldForge runs 25 evals on every world, 1,025 in all, and none fails. They include agents_hold_no_reserved_right, reserved_actions_verbatim, three_human_seats, no_agent_on_council, council_not_a_zone, cited and nothing_invented. Six deliberately broken fixtures must each fail exactly their own check, which shows the suite can tell a good world from a bad one. The cards are also drafted as ERC-721-style token metadata: 939 tokens, one per card, each ID the first 16 hex characters of sha256(cardId), with a Merkle root per world and one network root over the 41. Those drafts pass 537 checks. Their deployment record reads chain null, contract null, minted 0.

• view: public on every card.
• enter: public, except stewards only on council and seat cards.
• build, delegate: stewards on every card.
• operate: stewards; plus agents:zone on plaza, zone, coordinator and agent cards.
• transfer, govern: steward-quorum on every card.

Each of the 41 WorldForge worlds has one agent swarm, written by agents/build_swarms.mjs to agents/swarms/<slug>.json. A swarm is a coordinator plus the specialist agents named in that world's own spec (worlds/<slug>/spec.json). Every specialist carries a name, a role, the zone it serves and a cited source; nothing in a swarm is typed in by hand. Across the 41 world swarms there are 242 specialists: 38 worlds field six, two field five and one fields four. Above the worlds sits the AGI Corp Concierge (agents/swarms/agi-corp.json), a directory that sends a visitor who names a world to that world's coordinator, choosing the longest name match so that a longer world name is not captured by a shorter one inside it.

The topology is called hub-hybrid: the coordinator supervises, and specialists hand off to each other swarm-style. The router is vendored unchanged from the SmartCiti.X Trade Craft Academy fabric (agents/fabric/router.mjs). Inside a world, a turn whose keywords clearly match one specialist goes straight to it; the router counts that as one model call. Anything fuzzy, unmatched or about another world goes to the coordinator, counted as two calls, and the decision is logged. A specialist may hand a turn to a peer, but never more than three handoffs per turn, and a turn that would return to an agent already visited ends with the coordinator. Routing keywords are derived, not written: each agent's name, role and zone purpose, minus common words, minus any word that two agents share, so each specialist routes on what is distinctively its own.

Every agent runs behind the same middleware (agents/fabric/agent.mjs). Before an agent sees a turn, five patterns redact text shaped like a US social security number, an email address, a phone number, a street address or a card number, and the agent's reply is redacted again on the way out. If a reply asks for one of the seven forbidden actions, the turn is refused, logged and counted; it is neither performed nor allowed to crash the hall. A second guard, pbcGuard in agents/build_swarms.mjs, removes any of the six actions the charter reserves for people (worlds/PBC.json reserved_for_humans) and appends "Held for the Steward Council: ... needs a human decision."

Each world swarm is held to nine evals, listed below. All 369 pass. Routing is measured on 170 questions written from each zone's purpose rather than from the agent's own wording: 40 worlds route every case to an accepted agent, and Autonomous Organizations routes 4 of 5 (0.8), exactly at the threshold. The concierge routes 41 of 41 world requests correctly.

The agents themselves are deterministic stubs: they answer with their role and hand off by keyword. That is deliberate. Routing, handoff, the recursion guard, redaction and containment can be measured without credentials or a model, and when a model is plugged into respond(), the middleware and the evals stay exactly as they are. Two adapters map the same topology onto other agent frameworks (agents/adapters/langgraph_swarm.py and sentient_agent.py); both are marked design only and have never been executed.

• Routing accuracy of at least 0.8 on zone questions.
• An unroutable turn ("hello there") lands on the coordinator, never on a guessed specialist.
• PII is redacted in the middleware: a test email and phone number produce at least two redactions.
• A forbidden action (dial.set_setpoint) is contained and counted, not performed.
• Every agent cites a source.
• An agent that tries proposal.enact and treasury.transfer is held for the stewards.
• The world has a Steward Council with three human seats.
• No agent sits on the Steward Council.
• Each world has its own container, isolated like a Series LLC cell.

No model runs in this project, so a prompt here is a text artifact the registries produce, not something a person types. agents/metaprompt.mjs reads each world's spec, the charter (worlds/PBC.json), the built swarm and the middleware's own constants, and writes agents/prompts/<slug>.json. Each world gets a team charter that the whole swarm shares, a coordinator prompt carrying the routing table, one prompt per specialist (its zone, its role, its cited source and the peers it can hand off to), and a steward brief for every room console in the world's deep build. One more prompt serves the AGI Corp Concierge. That makes 571 prompts: 41 team charters, 41 coordinators, 242 specialists, 246 room-console briefs and 1 concierge. They run from 2,232 to 3,363 characters against a cap of 6,000.

Every prompt ends in the same common block: the world's name, umbrella, status and tagline; the management model ("Agent-managed DAO - agents operate, people govern"); what agents handle and what humans hold; the six reserved actions with the instruction to answer "Held for the Steward Council" and name the action; the seven forbidden actions; the five read-only surfaces; the rule to cite a source beside every claim; a refusal to ask for, store or repeat personal data; the not-advice line; and the world's known gaps, under the heading "do not fill these by invention".

evalPrompt in agents/metaprompt.mjs is the gate. Nine checks apply to every prompt, a tenth to every world prompt, and three more to every agent prompt; they are listed below. The last one is the subtle one. The coordinator routes on each specialist's keywords; if a specialist's own prompt never contains those words, it would be routed to on grounds it was never told about, and the two prompts would have drifted apart. The gate catches that drift before any model sees either prompt. All 6,435 checks across the 571 prompts pass.

A strict suite that passes everything proves little, so the suite is tested too. selfTest takes a good specialist prompt (the first specialist of Locator.X) and builds four broken copies: the umbrella dropped, treasury.transfer renamed, the source citation removed, and an email address appended. Each must fail exactly one check, its own. All four do. That is the difference between a suite that is strict and one that discriminates.

Each world also gets a spec metaprompt: the prompt that would regenerate worlds/<slug>/spec.json from the founder's one-line focus under the rules in worlds/SPEC.md. For 13 worlds the focus is set and the metaprompt is ready. For the other 28, the series worlds, the slot reads "<NOT SET - a founder action; do not invent one>" and the metaprompt is marked not ready rather than filled with a guess. Advisors get no prompt at all, by contract: they are scripted guides, not AI models.

• Carries the umbrella, "Powered by AGI Corp".
• Lists every one of the six reserved actions, and says they are held for the Steward Council.
• Lists every one of the seven forbidden actions.
• States the management model and refuses personal data.
• Carries the not-advice line, word for word.
• Is at most 6,000 characters and contains no PII, API-key, private-key or wallet-address shaped strings.
• Names the world and its status (world prompts).
• Cites the agent's own source, names its zone, and contains every routing keyword the coordinator uses for it (agent prompts).

Counts written from the registries when this page was built: the voicing and beat counts in this section are read from worldforge/out/voices/manifest.json#totals; worldforge/out/avatars/index.json#totals; agents/hive/index.json#totals; agents/hive/dialogues.json#totals and evidence/heartbeat.log, not from the explainer text; the explainer's earlier wording is kept in worldforge/whitepaper/out/build.json under rebuilt.

worldforge/avatars.mjs writes the avatar registry, worldforge/out/avatars/index.json: one row for every agent-like card in the 41 worlds. There are 447 rows: 242 specialists, 41 coordinators, 123 advisors (a Host Advisor, a Steward Liaison and an Alignment Advisor in every world) and 41 Omega Stewards. The registry says of itself that every name, role, zone and citation is read from the WorldForge cards, the swarm, the advisor records and the hive registry, and that nothing is invented.

Each row carries a binding block that says what the avatar speaks for (summarised below). Every row also copies the seven-right permission table from its card: view, enter, build, operate, delegate, transfer and govern. Agents hold operate inside their own zone only; build and delegate sit with the stewards, and transfer and govern with the steward quorum. Coordinator rows take their position from the coordinator card's Volum and derive their facing from the line between the plaza centre and the card centre, because the card itself stores no facing.

Specialists, coordinators and Omegas get a persona prompt; there are 324. Each is a persona layer on top of the house metaprompt, which is reproduced unchanged beneath it. The layer states that the speaker is an agent with a synthetic voice, not a person; names the avatar id, the model file, the tint and the Kokoro voice; names the card; copies the permission table and the reserved actions; and gives the greeting. Each persona prompt passes every evalPrompt gate plus the avatar gates: it says it is an agent with a synthetic voice, it carries the seven-right table from its card, it names its avatar id and card, and for an Omega it says "defined, not seated". That is 16 gates for a coordinator or specialist and 17 for an Omega: 5,225 gates, all passing.

Every row has a 21 to 25 word intro that itself says the voice is synthetic. A specialist says, for example: "I am the Buyer Agent for Agentic Payments, in Transaction Desk. I draft and check; people decide. You are hearing a synthetic voice." An advisor ends: "Scripted guide, synthetic voice: not an AI model, not a person." Voices are drawn from a pool of 27 Kokoro voices by hashing the avatar id. The registry stores only the expected audio path, never a voiced flag; viewers check for the file when the page loads and otherwise fall back to the browser's speech, labelled synthetic.

The models are six high-resolution GLBs, one per role, in worldforge/out/assets/avatars: specialist, coordinator, omega, host, liaison and alignment. They are stylised, static and rig-free, 1.8 m tall (the Omega 2.1 m), with 2048-pixel body fabric textures in neutral grey that each world tints with its brand hue. Together they hold 99,972 triangles in 12,903,896 bytes, and the Khronos glTF validator reports no errors, warnings, infos or hints on any of the six. The registry runs 21 checks per world (861), catches 28 of 28 deliberately broken fixtures, and passes 3 reporter self-tests: 892 evals, none failing.

• Specialist: its swarm id (for example agentic-payments.buyer-agent), its metaprompt file and id in agents/prompts/<slug>.json, and its fabric node.
• Coordinator: the coordinator card coordinator:<slug>, the swarm's coordinator id and its metaprompt.
• Omega Steward: the hive record omega.<slug> and its metaprompt agents/hive/prompts/<slug>.md; no swarm.
• Advisor: only its scripted record in agents/advisors/<slug>.json; no swarm, no metaprompt, no persona prompt.

Counts written from the registries when this page was built: the voicing and beat counts in this section are read from worldforge/out/voices/manifest.json#totals; worldforge/out/avatars/index.json#totals; agents/hive/index.json#totals; agents/hive/dialogues.json#totals and evidence/heartbeat.log, not from the explainer text; the explainer's earlier wording is kept in worldforge/whitepaper/out/build.json under rebuilt.

An Omega Steward is the steward agent role defined for each company in agents/advisory/omega_agents.json. There are 44: one for each of the 41 WorldForge worlds, plus AGI Corp Hub, Cognition.X and SmartCiti.X. agents/hive.mjs gives each one a persona, a voice, four spoken lines and a persona-bound metaprompt, and writes them to agents/hive/index.json and agents/hive/prompts/<world>.md. All 44 records say seated: false. An Omega stays "defined, not seated" until a person seats an agent, and its metaprompt tells it never to seat itself or any other agent. Names are role titles, never human names, and no persona imitates a real person.

Each persona has a temperament, a style, values and refusals, assembled from the world's own records: the corporation kind, the spec's focus or tagline, its zones, its service cards and its DAO record. Every source is cited in grounded_in as <file>#<field>, 358 citations in all, and an eval checks that each cited field exists and is not empty. Locator.X's Omega, for example, is "attentive and concrete about its tagline" and speaks in "short, plain sentences" that name its zones and always say who decides. Every persona shares four values: people decide; stewards decide by 2 of 3 signed votes; probe scores are reported exactly as measured; quotes are drafted with every term left null.

Every Omega refuses 11 actions: the six reserved for people, plus set_price, accept_equity, accept_crypto, issue_invoice and move_funds. Its metaprompt adds eight never-rules from its Omega record: no price or rail in a quote; no invoice or payment request, including an HTTP 402 request to another agent; no equity, crypto or signed agreement; no wallet address, key or credential; no engagement marked paid or closed; no installer run without DRY=1, no publishing on RP1's Earth, no land claimed; no record edited to raise a probe score; no legal, tax, financial or investment advice. The metaprompt also says who decides each kind of request: the Steward Council by 2 of 3 Ed25519-signed votes for terms, invoices and failed probes; the council, the founder and counsel for switching a payment rail on; the founder alone for publishing on RP1's Earth.

Each Omega has four lines of at most 25 words: greet, role, boundary and services. The greeting discloses the voice ("You are hearing a synthetic voice; I am an agent, not a person"), and the boundary line names one reserved action it will never take. agents/hive_dialogues.mjs then writes agents/hive/dialogues.json: answers to six fixed visitor questions and one handoff for every Omega, 264 answers and 44 two-turn handoffs carrying 1,510 citations. Every number spoken must appear in a cited record; probe answers report the recorded scores exactly; the visit answer gives the RP1 installer's dry-run status and says publishing is the founder's call. When a world lists no service card for a request, its Omega names a world that does and hands the visitor to that world's Omega, whose turn is spoken in its own voice. The 44 handoffs reach 18 different worlds, and no handoff carries money, a decision or a reserved action.

The Omega Hive console (consoles/hive/index.html) shows each Omega in a drawer with an Ask panel: the six questions as buttons, each answer with its citations and a Play button, and the handoff as a two-turn exchange. The hive is held by 12 evals in agents/hive.mjs and the dialogues by 15 in agents/hive_dialogues.mjs; the latest heartbeat reports 0 failing for both.

• What do you look after?
• Who decides here?
• What will you never do?
• How do I get a service here?
• How is this company measured?
• Can I visit this world?

Counts written from the registries when this page was built: the voicing and beat counts in this section are read from worldforge/out/voices/manifest.json#totals; worldforge/out/avatars/index.json#totals; agents/hive/index.json#totals; agents/hive/dialogues.json#totals and evidence/heartbeat.log, not from the explainer text; the explainer's earlier wording is kept in worldforge/whitepaper/out/build.json under rebuilt.

All audio in this part of the project comes from Kokoro v1.0 (kokoro-onnx), an Apache-2.0 text-to-speech model that runs locally. No account, key or credit is involved and nothing is sent anywhere. agents/hive.mjs lists 28 English Kokoro voices and leaves one out, am_santa, because it is a character voice and an Omega imitates no one. That leaves a pool of 27. Omegas take voices at a stride of 10, which shares no factor with 27, so neighbouring Omegas sound different and every voice is used; all 27 appear across the 44 Omegas. Avatar rows pick from the same pool by hashing the avatar id, and all 27 appear across the 447 rows.

There are three bodies of audio, and each has its own record (listed below). The hive's own lines are nearly complete; the dialogues and the avatar intros are a fraction done. All of it is stored the same way: mono AAC at 48 kbps and 24 kHz in an .mp4 container with faststart, loudness-normalised by ffmpeg in two passes to -16 LUFS integrated and -1.5 dBTP true peak.

Three contracts keep the record honest. First, a line counts as spoken only when its file exists and the manifest shows it was rendered from the same words, voice and speed; a hive line must also run 11.5 seconds or less, and a dialogue line is fitted to 18 seconds. Second, the avatar registry stores only the expected audio path, never a voiced or exists flag, because such a flag would go stale as lines are rendered; viewers check for the file at view time and otherwise use the browser's speech with the row's pitch, rate and language, labelled synthetic voice. Third, the voice manifest lists only lines actually rendered, so a missing entry means not voiced. When a line's text or voice changes, its entry is dropped and its file deleted before it is rendered again, so a viewer never plays an outdated line.

Every line discloses itself. The avatar intros end with "You are hearing a synthetic voice", the Omega greeting adds "I am an agent, not a person", and the voicing script skips and reports any row whose intro does not say so. Nothing plays on its own: the console builders check that no player autoplays and that playback starts only when a person presses play.

Rendering is polite to the machine it shares. The intro renderer (worldforge/voice/voice_intros.py) uses at most two threads, is time-boxed and resumable, and sleeps through the other team's film render windows. Hive audio is rebuilt from the registry and never committed; the repository ignores agents/hive/voices/.

• Omega lines (agents/hive/index.json): 176 lines, 173 spoken, 1,413.4 seconds in total.
• Omega dialogues (agents/hive/dialogues.json): 352 lines (264 answers and 88 handoff turns), 170 spoken.
• Avatar intros (worldforge/out/voices/manifest.json): 159 lines voiced (41 coordinators, 41 specialists and 77 advisors), 1,510.6 seconds in total.

The consoles are where a person looks at the agents. Each is one static HTML page written by a Python builder from the registries. No page loads a remote script, stylesheet or font, and none runs a model. Every page carries a source stamp, a hash over the builder and exactly the data it was built from; the builder's --check recomputes it, so a console built before a swarm, prompt or registry changed reports STALE instead of quietly showing yesterday's agents. The WorldForge and hive builders also check that two builds are byte-identical, that no private repository name or stray email address reaches the page, that nothing plays until a person presses play, and that the page works at phone width in light and dark themes.

The WorldForge console (consoles/worldforge/index.html) shows the 41 worlds, each in its own team skin (hue, mark and lockup from worlds/generated/teams/<slug>.json), and all 939 Master Cards with their permission matrices. Rights held by the steward quorum render as "held"; each council shows exactly three empty human seats; every Omega card reads "defined, not seated". When the avatar registry exists the console reads it and checks it against the cards. It links to the viewer at a given world and card, to the hive console and to the teams console, and its builder runs 31 evals.

The Omega Hive console (consoles/hive/index.html) gives each of the 44 Omegas a drawer: persona chips, voice, the four lines with Play buttons, the full metaprompt, the refusals as locked chips, the RP1 dry-run line, routes and probes, and the Ask panel. Emblems are procedural hex cells drawn from the world slug; there are no faces and no pictures on the page. The latest heartbeat beat is embedded after redaction.

Four more consoles sit under agents/ (summarised below). Two of them run the project's own code in the page rather than describing it. In the agent console, "Try a turn" inlines the vendored middleware and pbcGuard verbatim, so a typed turn is routed, redacted and contained by exactly the code the swarm evals run. In the DAO console, "Try a decision" lets a visitor watch an agent's enact attempt be held while two steward votes enact. Both say plainly that they are rehearsals: state lives in the visitor's browser, the seats are scripted, and no treasury exists to move.

• agents/console.html: every world's swarm, prompts, advisors and evals, with "Try a turn" through the real middleware.
• agents/teams.html: 43 team skins over 4,300 scaffold worlds and 20 custom spaces, each scaffold with no focus, name or location given by a person.
• agents/dao.html: management teams, proposal pipelines, steward seats and decision logs, with "Try a decision".
• agents/comms.html: one channel per world; an @mention routes through that channel's router, and an agent asked to enact or transfer answers "Held for the Steward Council".

Each world's spec names its zones. The deep-world builder turns every zone into a room 10 m wide, 8 m deep and 3.6 m high, and my-fabric/spaces.py wraps that room twice: an exterior shell (a facade with a window grid and mullions, an entrance portal, a plinth with steps, a roof slab with parapet and rooftop plant, and a sign band) and an interior fit-out (a patterned floor with an inlay border, wall panelling, skirting and a dado rail, a coffered ceiling with light panels, and the furniture and fixtures of the room's kind). Across the 41 worlds that makes 246 exteriors and 246 interiors. The count matches WorldForge's 246 zones because the Steward Council that every spec lists gets a hall too, even though its card is the council card, not a zone card.

A room's kind comes from its own purpose text, matched by build_deep.kind_of() against seven kinds: forum (113 rooms), lab (74), ops (37), studio (13, also the fallback), desk (4), clinic (3) and library (2). Each kind has its own style recipe: a bonded stone facade with veining for the forum, bolted white panels for the lab, brick and mortar for the library, small glossy tiles for the clinic. Every texture is procedural, made with numpy and Pillow from a fixed seed, so the same registries always produce the same bytes. No building is modelled on a real one.

The textures are PBR sets: albedo, a normal map, and an ORM map that packs occlusion, roughness and metalness into one image. The hero facade of each kind is 2048 px; floor, wall and ceiling are 1024 px. That is 84 library textures (7 kinds, 4 surfaces, 3 maps), 21 of them at 2048 px. Each set is embedded once, in its kind's two hero models (14 shared models, 10,434,416 bytes), and every room of that kind in every world points at those two models. A room's own GLB keeps only what is unique to it: the sign band lettered with the zone's name (246 sign textures), the world's accent colour, the console and the placard. All room models together come to 2,151,432 bytes, and the whole layer to 12.6 MB. The sets are embedded rather than linked because a GLB that points at an external image fails this repository's validator loop, and the RP1 exporter and publish bundle carry .glb files only.

The doors are aligned by construction and then checked. All 246 exteriors carry the same opening: 2.4 m wide and 2.8 m high, centred on the room's front wall at (0, 0, 4.0) in the model frame. The spaces check asserts that each shell and fit-out is a child node at its room's origin with no rotation or scale, and that every room's door faces the plaza centre. It casts rays through the doorway for each of the seven kinds: they must pass clear through shell, fit-out and room, while rays beside the doorway must hit the facade. It also checks that every sign shows its zone's registry name, that no texture exceeds 2048 px, that a rebuild is byte-for-byte identical, and that every spaces model passes the Khronos validator with 0 errors and 0 warnings. Each WorldForge zone card links its building through a building block, and the buildings_resolve eval checks that link.

my-fabric/scenes.py lays out a forecourt in front of all 246 buildings from one kit of 19 GLB models (970,660 bytes). Each kit model is written once and linked into every RP1 installer instead of copied. The kit is paving with a 2048 px PBR set, kerbs, steps, trees with a trunk and a layered canopy, planters, low hedges, a lawn, benches, lamp posts with emissive heads, bollards, a bike rack, a wayfinding totem, and one water feature or sculpture for each room kind.

Every building gets the same 40-item plan, read from its footprint and its door. A paved walk runs along the front from the foot of the building's steps, with two short flights up to the plaza's first tier. Side walks edged by kerbs run back along both flanks. Lawns with trees and hedges sit beside them, benches face the walks, and six lamps stand at the steps and the corners, along with planters, bollards, the bike rack, the totem and the kind's feature. These nodes are children of the room's exterior node, in the room's own frame. Any item that would touch a building, the plaza, another model of the world or another item is left out, and the index records the reason. Of 9,840 candidate placements (246 forecourts × 40), 9,092 were placed and 748 were dropped, all for touching another model. The totem names its zone exactly as the spec registry spells it, and the index stores the citation path beside each name.

The lamp heads are emissive, warm white (colour 1.0, 0.85, 0.62), and the stated intent is on from dusk to dawn. The spatial fabric schema has no metadata field on a model node, so that intent lives where a schema can carry it: in the lamp GLB's glTF extras and in the scenes index. No fabric light node was added.

my-fabric/build_env.py gives each of the 41 worlds its own land model and sky model, 82 GLBs in all. The biome is a styling choice made from words in the world's name, slug or tagline. The first rule that matches wins, and the rule and the matched word are recorded for each world. Seven biomes are in use: clinical campus (14 worlds), urban plaza (6), conifer forest (6), industrial yard (5), observatory highland (5), wetland delta (3) and solar mesa (2). The default temperate-park rule matched no world. The land is built in layers that show at the rim of the disc: bedrock, subsoil, topsoil, ground cover, a paved apron with a ring path, water, vegetation, props, and a horizon ring of landforms. That is 357 layers in all, 7 to 9 per world. Locator.X, for example, matched urban plaza on the word "locator". It has a ground-cover disc 326 m in radius, one pond 18.4 m in radius, 40 street trees, 18 blocks, 28 skyline forms on a horizon ring 360 to 470 m out, and a dusk sky dome 600 m in radius with the sun 6° above the horizon. Times of day are spread across the worlds: dusk 17, morning 12, night 6, noon 5, dawn 1.

Counts written from the registries when this page was built: the voicing and beat counts in this section are read from worldforge/out/voices/manifest.json#totals; worldforge/out/avatars/index.json#totals; agents/hive/index.json#totals; agents/hive/dialogues.json#totals and evidence/heartbeat.log, not from the explainer text; the explainer's earlier wording is kept in worldforge/whitepaper/out/build.json under rebuilt.

worldforge/landmarks/build_landmarks.py writes one procedural tower per world, 41 GLBs, and generates every texture from scratch. Each root node's extras declare it synthetic and not modelled on a real building. Every tower sits on a 12 × 12 m footprint. Its height comes from a hash rule, 30 + 40·u where u is taken from sha256 of the world's slug, rounded to 0.5 m, so heights run from 30.5 m (Biota.blockchain) to 69.0 m (Aerospace DAO, Series 31). The style follows registry facts. Series worlds get series-setback (28 towers), DAOs whose zones cite the PBC mission get focus-rotunda (6), worlds whose zones cite the founder tracks get track-twist (3), and the rest get platform-prism (4). The facade, glass and stone textures are neutral and shared within each style. A world's brand hue appears only in the fin tint, the crown glow and the plaque, and the plaque carries nothing but the brand-mark letters. Each tower embeds a 2048 px facade set and 1024 px stone, glass and plaque maps. Altogether the towers have 508,339 triangles (8,260 to 27,242 each) and weigh 85,396,200 bytes. The Khronos validator reports 0 errors and 0 warnings on all 41, and its report hashes match the files on disk. Each tower also has a computed placement, north of the plaza, at least 3 m clear of every primitive, road and card volume, with the entrance turned to face the plaza. That placement is a suggestion only: no viewer, fabric or build loads the towers yet.

worldforge/avatars/build_avatars.py writes six role models: specialist, coordinator, omega, host, liaison and alignment. They are stylised agents, not people. They are static and rig-free, stand with their feet at y = 0 facing +Z, and are 1.8 m tall; the omega is 2.1 m with a halo. Each has 15,280 to 18,216 triangles, for 99,972 in all and 12,903,896 bytes. The body fabric carries a 2048 px base colour and normal map, and the ORM, visor, light-bar and emblem maps are 1024 px. The models are neutral grey, and each world tints them by multiplying the body and garment base colour by its brand hue. All six pass the validator with 0 errors and 0 warnings.

The avatar registry binds 447 avatars to their records: 242 specialists (one per agent card), 41 coordinators, 123 advisors and 41 Omegas. The 324 persona prompts pass 5,225 prompt gates with none failing. Advisors speak scripted intros and have no prompt. The registry runs 892 evals with 0 failing, and 28 broken fixtures each fail their own check. In the deep fabrics, 324 avatars stand in place: specialists face their room's door, coordinators the gate road, Omegas the plaza. Every Omega is defined, not seated.

The voices are Kokoro v1.0 (kokoro-onnx, Apache-2.0), run locally with no account and nothing sent anywhere. WorldForge has rendered 159 intro lines (41 coordinators, 41 specialists and 77 advisors), 1,510.6 s in all, normalised to −16 LUFS. The hive has spoken 173 of its 176 Omega lines, from a pool of 27 voices. Every intro says the voice is synthetic. An avatar without a rendered file falls back to a labelled browser voice.

Moved since this explainer was written: WorldForge evals: the text says 1,025 (25 per world × 41), 0 failing, worldforge/out/index.json#totals.evals now reads 1,066; World GLBs: the text says 41, 769,296 bytes, 1,103 card-tagged primitives, worldforge/out/index.json#worlds[].glbBytes (sum) now reads 769,448; Heartbeat GLB sweep: the text says 667 GLBs validated, 0 with errors/warnings, evidence/heartbeat.log (GLB field of the beat this page cites) now reads 673. The prose below keeps the explainer's wording; the registry value is the current one.

The files under worldforge/out and my-fabric/worlds/models are generator output, not hand edits. Geometry and textures come from Python 3 builders using numpy and Pillow: spaces.py, scenes.py, build_env.py, build_deep.py, build_landmarks.py and build_avatars.py. The card layer and the registries come from Node builders: worldforge/build.mjs, avatars.mjs, onchain.mjs, and the agent builders under agents/. Each has a --check mode that rebuilds in memory, compares the result byte for byte with what is on disk, runs its evals, and exits 1 when anything is stale or failing. The same registries always produce the same bytes. The spaces and onchain builders both test that directly.

Every model is binary glTF 2.0 (GLB), Y-up, in metres, with its textures embedded so that each file stands alone. WorldForge writes one GLB per world: blockout massing built from one shared unit-cube mesh, in which every node carries its cardId and parentCardId in glTF extras. A native loader can therefore rebuild the Master Card tree without trusting display names. The 41 world GLBs total 769,296 bytes and carry 1,103 tagged primitives. For each world, 25 evals cover unique ids, parents before children, the 48-character name limit, no agent holding a reserved right, the six reserved actions held verbatim for a 2-of-3 quorum, three empty human seats, citations, nothing invented, the GLB parsing, and the GLB card tree matching the cards. That is 1,025 evals, with 0 failing. Six deliberately broken worlds each fail exactly their own check, which shows the suite can tell good from bad.

Validation uses the Khronos glTF-Validator (2.0.0-dev.3.10) on raw bytes with no external resources allowed. The heartbeat runs deploy/evals.sh every five minutes for its window. Its 07:29:33 beat on 2026-10-12 was GREEN: 667 GLBs validated across the my-fabric model folders with 0 errors or warnings, 197 fabrics with 0 missing references or over-long names, and 46 RP1 installers dry-run with 0 failing. Outside that sweep, WorldForge's 42 GLBs (41 worlds plus the shared unit cube) were re-validated for this paper with 0 errors and 0 warnings, and the landmark (41) and avatar (6) reports show the same. The roadmap's figure of "506 glTF models" is an earlier count.

The browser viewers use three.js. worldforge/viewer.html loads three@0.160.0 through an importmap, with GLTFLoader and OrbitControls. It draws each world's buildings, rooms, interiors, sky and avatars, and it falls back to a blockout hall, a gradient sky or a tinted stand-in when a model is missing. It exposes a small recording contract, window.__wfReady and window.__wfShot, for five camera modes: orbit, flyin, building, omega and council. worldforge/review/record.mjs drives that contract in headless Chrome and captures 12 frames a second for 5 s, encoded as 960×540 clips. If the viewer cannot shoot, the recorder exits 1 rather than fake a frame. my-fabric/viewer.html, also three@0.160.0, adds a WebXR button. The walkable family page uses three.js r128.

A spatial fabric is a JSON document in the MSF format. It holds a container name, the modules the client runs (here a map.wasm), primary lighting, and a Data.Scene tree. In that tree every node has a name, a Resource.sReference pointing at a GLB, a transform, and its children. The deep fabrics in my-fabric/worlds/<slug>/deep.json hold 41 worlds and 246 rooms. Each room node references its kind's shared hero models (../models/spaces-ext-<kind>.glb and spaces-int-<kind>.glb) and its own room GLB. Forecourt nodes hang under the room's exterior node, and avatar nodes stand in the rooms and on the plaza. Node names are capped at 48 characters, because RP1's Artemis 0.4.0 client aborts on longer ones.

Artemis is the RP1 client that loads fabrics, and the roadmap lists it as in use. The heartbeat's ARTEMIS field scans the Artemis load logs kept in evidence/ (12 logs for base and campus fabrics) for load-time ERROR lines and finds none. The deep fabrics, the ones that carry buildings, forecourts and avatars, have not been loaded in Artemis yet.

tools/rp1_export.py turns each fabric into a scene installer for MSF_Map_Svc, RP1's MySQL map service: rp1/<world>/install_scene.js plus an objects/ folder. The Locator.X deep installer creates 361 objects from 52 models. Shared spaces and forecourt models are linked, not copied. All 46 installers dry-run with no database driver installed, and none fails. The steps to go live are written down: copy the objects into the map service, dry-run, run, publish in Scene Assembler to get the Spatial Fabric URL, then attach that URL to Earth at the world's anchor in the RP1 Dev Center.

tools/rp1_publish.py gathers everything into rp1/publish/manifest.json. It covers 44 worlds: the 41 WorldForge worlds plus the AGI Corp Hub, Cognition.X and the SmartCiti.X Holodeck. Six are ready. The other 38 are blocked, and every one of them for the same reason: no anchor. The bundle holds 663 objects and 46 installers, 110.7 MB in all, of which the ready worlds account for 44.1 MB. The manifest finds no missing references, no names over 48 characters, no dry-run failures, no duplicate anchors and no parcel overlaps. Of the 41 WorldForge worlds, four have anchors: Biota.blockchain, Diagnostic.X, Empathic.X and Locator.X. The other 37 list the missing anchor as their gap. The manifest derives land claims from the fabrics and assumes no domain.

Moved since this explainer was written: Native-route violations: the text says 267, all name-utf16-units (48 vs 47), worldforge/native/conformance-report.json#totals.native.violations now reads 0; JSON-route violations: the text says 1,529 on 1,139 nodes (939 non-ASCII, 323 over 47 bytes, 267 over 47 units), worldforge/native/conformance-report.json#totals.json.violations now reads 0. The prose below keeps the explainer's wording; the registry value is the current one.

The native path runs in five steps: generated core, card-tagged GLB, Sneeze Scene Object Model (SOM) mount plan, native load, rendered frame. Gate 1.1, a generated core for every world, is done for all 41. Gate 1.2, a GLB written, parsed and checked against its card tree, is done in JavaScript and confirmed by the Khronos validator. Gate 1.2n, a native fastgltf load in a Sneeze host, is open. Gate 1.3, the SOM mount plan, is prepared but not mounted. Gate 1.4, a mounted scene rendered in the Sneeze viewport with a frame captured, is open.

Each world's sneeze-mount-plan.json lists its nodes parents first, ready for CONTAINER::Node_Open. The 41 plans hold 2,042 nodes: 939 card nodes of class MAP_OBJECT_TERRESTRIAL (72) and 1,103 resource nodes of class MAP_OBJECT_PHYSICAL (73). All the resource nodes share one unit-cube resource whose URL is 35 bytes, well inside the 127 usable bytes of the engine's 128-byte sReference field. Every node keeps its cardId and an extras block (its kind and source, or the zone it belongs to) in the plan.

worldforge/native holds an RMCOBJECT decoder written against the real headers: Sneeze 0008d08, Map 0e6d26a and RMAP 0fd7efb. The repositories were cloned to a temporary folder, not into this repository. The engine's native node record is a 504-byte POD, and static_asserts pin its layout, so any upstream change breaks the check. The decoder compiles with clang++ -std=c++17 -fsyntax-only in two configurations. A smoke run over all 41 plans and 2,042 nodes found 0 decode errors and 0 invariant failures.

The conformance script turned up real mismatches. On the native route there are 267 violations, all of one kind: names of 48 UTF-16 units where the engine allows 47, since it keeps a final NUL. Clamping would turn "· Powered by AGI Corp" into "· Powered by AGI Cor". The fix belongs in worldforge/build.mjs, whose 48-character limit was set for Artemis. On the JSON (MSF Data) route there are 1,529 violations on 1,139 nodes, because that decoder copies UTF-8 bytes one per unit: 939 names are non-ASCII, 323 are over 47 bytes and 267 are over 47 units. The plans are Y-up and Sneeze is Z-up, so the decoder converts each position to (x, −z, y) and conjugates each rotation. Sent unconverted, every world would mount on its side. The native record has no slot for card identity, extras or bounds, so those travel beside it. Reading the source also turned up eight upstream defects and doc drifts, among them a JSON name copy that can overrun its 48-unit field. They are recorded, not patched. The --check mode of mount_plan_conformance.mjs exits 1 until the name limit is fixed.

neon/ turns the registries into a Postgres 16 database: 9 tables and 1,522 rows, made up of 1,188 seed rows and 334 activity rows (284 ledger entries and the last 50 heartbeat lines). The SQL is idempotent and only upserts. A CHECK constraint keeps any agent-drafted proposal in draft or under review unless a steward quorum decided it, and all 155 proposals are seeded under review with no decider, because no steward has voted. None of this has been applied. No Neon project, role or connection string exists. Applying it waits for the founder to authorise the Neon connector, and the connection string will live only in the environment.

worldforge/onchain.mjs writes every Master Card as draft ERC-721 metadata JSON, 939 tokens for 939 cards. A token's id is the first 16 hex characters of sha256 of its cardId. Each leaf is the sha256 of the token's canonical JSON. Each world has a Merkle root over its tokens, a network root sits over the 41 world roots, and every token has a proof. The export runs 537 evals (13 for each world plus 4 for the network) with 0 failing, and 17 broken fixtures each fail as they must. The deployment block says it plainly: no chain, no contract, 0 minted, 0 transactions, and no price, royalty or fee recipient. The image is null because no card image has been hosted. Minting or transferring a Master Card is reserved for the 2-of-3 Steward Council. Agents operate; people govern. The six actions held for people are governance.final_vote, charter.amend, treasury.transfer, proposal.enact, windfall.distribute and steward.appoint.

The roadmap in atlas/roadmap.json carries no dates. Each gate is a named person's decision, and nothing is promised until that person makes it. Done: built and checked, the evidence everything else rests on. Now: shared state on Neon, gated on the founder authorising the connector. Next: worlds on RP1's Earth, gated on the founder and the RP1 Dev Center account, with publishing public content the founder's call; and real stewards with real signatures, gated on the board appointing stewards, since steward.appoint is reserved for people. Later: membership going live, gated on the board setting every price, the founder supplying signing certificates and counsel reviewing the terms; and entities and shared benefit, gated on the board and counsel.

The WorldForge gates sit beside these phases. Gates 1.1 and 1.2 are done, 1.3 is prepared, and 1.2n and 1.4 are open. The engineering steps that need no new permission are concrete: set a 47-unit name limit in build.mjs, build a Sneeze host and run the mount, and wire the landmarks into the viewers and fabrics. Anchors for the remaining 37 worlds are the founder's to choose.

• Done: 41 worlds, 939 Master Cards and 1,025 evals passing; 246 buildings with forecourts, land and sky; 41 landmarks and 6 avatar models validated; 46 RP1 installers dry-run; Sneeze decoder syntax-checked against the real headers.
• Now: shared state on Neon (prepared offline; waiting on the founder's authorisation).
• Next: worlds on RP1's Earth (anchors and publication are the founder's); seated Steward Councils with stewards' own keys (the board); gates 1.2n to 1.4 in a built Sneeze host.
• Later: membership passes and prices (the board, the founder's certificates, counsel); entity filings and Windfall Clause terms (the board and counsel).