Features · AGI Future Foundation
The Proof Is the Product
Most organisations publish their numbers and hide their limits. This one publishes its limits and lets the numbers speak for themselves, and the difference is the closest thing the family has to a signature. Across fifty-six repositories, the same habit recurs: a bundle that runs nine thousand of its own checks and then tells you the checks are its own, a curriculum that rebuilds itself on every push and fails if anything drifts, an eval suite that runs deliberately broken controls and demands that they fail, and a truth audit that corrects the family’s own paperwork in public. The numbers in this feature are attributed, not settled. The practice is the story, and the practice is verifiable in a way a press release never is.
The practice is the story
The GitHub review of 2026-11-05 looked at all 56 repositories in the Foundation’s organisation, and its most useful finding was not a number. It was a habit. The two biggest READMEs in the organisation publish their limits in the open, and the review names honesty-first engineering as a genuine differentiator, because the limits are stated in the same documents as the claims, in the same voice, without a separate page for the caveats. That habit is the subject of this feature. The counts that follow are real, and they are attributed, and they are not trophies. A count that cannot be checked is a slogan. A count that can be checked, with its limits stated beside it, is evidence, and the family’s engineering culture is built around producing the second kind.
The shape of the organisation explains why the habit matters so much. The review found that 20 repositories were all created on a single day, 2026-11-04, the whole family of landing-page repositories plus the robotics site, and that the July 2026 layer is hackathon and starter-template work, while the substantive venture code was last updated between September and November 2026. Only 8 of the 33 series have any repository coverage at all, and 25 of the 33 series have no repository in the organisation. Ten repositories are empty. Twenty-one claims were flagged. An organisation built in bursts, with landing pages arriving in a single day and substantive code arriving in a different season, needs a verification practice more than a polished one does, because the gap between a landing page and a working system is exactly where an invented claim can live. The practice described here is the family’s answer to that gap: self-tests, rebuilds, eval gates, and audits, each checking a different part of the work, and each publishing what it found.
The review’s final section lists the safe positive angles, and they are worth quoting because they are the review’s own judgment, not this feature’s. Breadth of real builds, across 56 repositories, with substantive code in SmartCiti.X, Cognition.X, Locator.X, the AGI-Biotech-tool, sentinel-stream, Metropolis-SmartCiti, agi-corp-worlds, WorldForge-Browser, EventAssist, and Sapient.X. Static sites that are pure HTML, CSS, and JavaScript with no build step, and several apps offline-capable by design. Real deployed artifacts: the WorldForge Pages site, the live Fellowship site, the Monad testnet contracts, and the working FastAPI and ClickHouse Sentinel system. An open-source posture, and human-governed design running through the organisation. None of those angles requires a large number to be true. Each one can be checked by opening a repository and reading it, which is the point. The proof is not a claim about the work. It is the work, left in public, with its limits attached.
The practice also runs through the governance papers, not only the code. The studio model requires a portfolio dashboard with actuals only, never a projection in a revenue or funds-in column, and never an unsigned partner, customer, or investor listed. The stage gates require a human approver’s signature and filed evidence before a gate is passed, and an agent’s readiness flag is never a gate pass. The truth audit of 2026-11-03 sorts every statement in the paperwork into verified, planning, or corrected, and corrects eight of them in public. The same rule governs both worlds: a claim is stated with its source, its limits, and its status, or it is not stated at all. That is what this feature means by verification culture, and it is why the feature treats the culture as the product, rather than treating any single count as the proof.
The bundle that grades itself
The clearest example in the organisation is the SmartCiti.X bundle, and it is the clearest because it is the most honest about what it is. The bundle is a Trade Craft Academy implementation, version 3.2, in the SmartCiti.X repository, with an ACP protocol suite, a 111-hall roster, a 3,663-skill graph, and eleven simulators. Its own text states that its self-test suite runs 9,267 checks, all passing from inside the bundle, with no model credentials required. The landing page for the same project states the same figures in its description: 111 union halls and 9,267 checks passing. The review’s finding on those numbers is the one this feature follows: 9,267 is the bundle’s own self-test suite count, not independent verification, and 111 halls is an in-repo roster, and the same README states that a union named is not a partner or affiliate, and that 0 of 111 halls are signed off by a named practitioner. Articles may say the bundle self-tests 9,267 checks and carries a 111-hall roster. They must never say 111 partnered union halls, and they must never say independently verified.
The README goes further than that, and the further it goes, the more useful it becomes as evidence of the culture. It states that nothing here certifies anybody. It labels its synthetic data as synthetic. It states that no mastery-tier lesson is composed, because no rule can do so honestly. It states that a ten-campus target, which other materials present as met, is met only as authored scenes and registries, not as physical campuses. Each of those sentences is a place where the bundle could have claimed more and chose to claim less, and each one is a place where a reader who checks the claim will find the limit already stated, in the same document, before they had to ask. That is the habit in its purest form: the author of a claim is also the author of its correction, and the correction is published first.
The landing page beside the bundle shows the other side of the same habit, and the review records it. The smartcitix-site repository is a landing page, and its README is a short description, while the substance is in the main repository. The same pattern holds for nine pairs of repositories in the organisation: a landing shell and a code repository are one venture, and the landing README is usually a five-line stub, with the substance in the main repository. A reader who cites the landing page as if it were the system has cited the shop window as if it were the factory. The review’s rule is to cite the main repository for substance and the landing page for what it is, a landing page, and this feature follows that rule throughout.
The rebuild that fails on drift
Cognition.X is the curriculum of verifiable blocks, and it is also the organisation’s best example of a disagreement stated in the open rather than settled by fiat. The Cognition.X repository and its landing page state 17,450 blocks, 47 packs, and six offline-first apps. The agi-corp-worlds repository states a different count for the same project: a Cognition.X Campus of 25 packs and 6,750 learning blocks, with a catalogue file laying out the whole Cognition.X catalogue as 6,750 learning blocks from a single data file, in 25 pack pavilions. Both counts are attributed in this feature, and neither is quoted as settled, because the review found the disagreement and did not resolve it, and this feature follows the review. A reader who quotes one count without the other has chosen a side in a disagreement they did not know existed. A reader who quotes both has done the work.
What makes the disagreement checkable, rather than merely confusing, is the rebuild. The Cognition.X README states that a fresh clone plus Python 3 regenerates the whole platform byte-for-byte, and that continuous integration rebuilds the dataset and all six apps on every push, and fails on any drift. That claim is stated here as the repository’s claim, because no CI run was checked in the review, and a claim about a CI run that nobody watched is a claim about a process, not a result. But the design of the claim is what matters: a system that rebuilds itself from scratch on every push, and fails if anything has drifted, is a system that cannot quietly change its numbers. The dataset and the apps are rebuilt together, so that a change to the data that is not reflected in the apps, or a change to the apps that is not reflected in the data, breaks the build in public. That is a verification practice, not a verification result, and it is the practice that makes the count disagreement useful rather than embarrassing: the disagreement is visible, the rebuild is automatic, and the next person to check can see both.
The same pattern appears in the molecular workspace, in a smaller and stricter form. The AGI-Biotech-tool repository, titled biodao.blockchain in its own description, fails its build on any undeclared network endpoint. A build that fails on an undeclared endpoint is a build that enforces its own honesty: no network call may happen that the repository has not declared in advance, and a developer who adds a call without declaring it does not get a warning. They get a failure. The review cites it as part of the self-verification culture, alongside the SmartCiti.X self-tests and the Cognition.X rebuilds, and the reason is the same in all three cases. The check is inside the work, not outside it, and it runs whether or not anyone is watching.
The eval gates that must fail
The most striking verification practice in the organisation is also the strangest, and it belongs to agi-corp-worlds, the repository that holds the family’s worlds. It runs 264 metaprompts through eval gates with deliberately-broken controls that must fail exactly their own check. Read that sentence again, because it inverts the usual idea of a test. A normal test passes when the system works. These controls are designed to fail, and the eval gate passes only when they fail in exactly the way they were designed to fail. A control that passes when it should fail is a control that is not checking anything, and a test suite that cannot tell the difference between a working system and a broken one is worse than no test suite at all, because it produces confidence without evidence. The deliberately-broken control is the answer to that problem: break the system on purpose, in a known way, and demand that the test catch it. If the test does not catch it, the test is broken, and the suite says so.
The same repository states its limits with the same bluntness as the SmartCiti.X bundle. No model has run any of these, it states, of its worlds. The 28 series worlds whose focus the stewards have not set are marked awaiting a founder focus, rather than filled with a guess. No treasury is held. The world counts disagree across the family’s materials — 40, 41, and 44 are all stated somewhere, and the README’s own table rows sum to 44, while it also describes 4,000 scaffold worlds, generated scaffolds with the status scaffold, explicitly never coming soon — and no single AGI Corp world count is quoted as fact in this feature, because the review found the disagreement and did not settle it. The repository is also private, which means an unauthenticated request for it returns a 404 by design, and this feature does not link to it, because a link that returns a 404 to a reader is not evidence. It is a dead end, and the feature says so rather than pretending otherwise.
The habit of marking the unknown rather than filling it is the thread that runs through all of these examples. A world with no founder focus is marked awaiting a founder focus, not given an invented one. A lesson that no rule can compose honestly is not composed. A campus count that disagrees between two repositories is stated as a disagreement, not resolved by picking the larger number. A test suite that runs 9,267 checks is described as the bundle’s own self-test suite, not as independent verification. In each case, the author chose the sentence that makes the work checkable over the sentence that makes the work impressive, and the review’s judgment is that the choice is the differentiator. Anyone can claim a large number. Fewer can show you how the number was counted, and fewer still can show you where the counting stops.
The testnet, the stream, and the stale README
The deployed artifacts are where the verification culture meets the public, and the review’s treatment of them is a lesson in how to celebrate a real result without overstating it. Metropolis-SmartCiti is the strongest example. Its README states that it is live on the Monad testnet, chain 10143, in the Monad Metropolis Track 04, with six contracts, and that every transaction was re-checked against the Monad testnet RPC, with 41 of 41 confirmed, status success, alongside 12 agents registered, 11 x402 payments, and 4 Trade Craft credentials issued. That is a testnet deployment with a verifiable on-chain record, and this feature celebrates it as exactly that: a testnet deployment, never a live production network, never users, never revenue. The distinction matters, because a testnet is a place where transactions are real and money is not, and a reader who confuses the two will believe the family has customers it does not have. The companion metropolis-site repository is a landing stub, and the review records it as one.
Sentinel Stream is the working system in the set, and its treatment shows the same discipline from a different angle. The sentinel-stream repository is a working FastAPI and ClickHouse system, version 0.3.0, for real-time threat detection for LLM agents, with MITRE ATLAS labelling, and it was an entry to the Cyberdefense Hackathon in San Francisco on October 9, 2025. It is an entry, not a win, and no prize is claimed for it, because the outcome is unverified. Its performance figure, less than two seconds, is stated as a design goal, never as a benchmark, because no benchmark was run. The repository has a duplicate, Sentient-Stream, with a README identical to the original and a description claiming a different version number, and the review notes the duplicate rather than citing it. A duplicate repository is not a second system. It is the same system twice, and citing it twice would be double-counting.
WorldForge and the Fellowship show the same habit applied to deployed sites rather than code. WorldForge-Browser serves 41 curated worlds on GitHub Pages, and the review is careful to say that 41 is the curated first set, not the total: the worlds file holds 544 entries, all with no RP1 attach, and Pages-live is not RP1-live. Zero worlds are live on RP1. The Fellowship site is live at www.agifellowship.church with 40 articles, and its repository README is stale, saying 15 articles, with a warning that article video links point to local workspace files and will fail until the promo videos are hosted. The live site now carries 40 articles, with all 39 videos, 12 images, and 3 logos deployed on 2026-11-05, and this feature follows the live site, not the README. A stale README is not a scandal. It is a reminder that the document and the deployment are two different things, and that the deployment is the source of truth, not the document that describes it.
The empty repos, and what they are for
Ten repositories in the organisation are empty, and the review counts them rather than hiding them, because an empty repository is also evidence, if you read it honestly. The seven RP1 ecosystem shells, the plaudguard-site and dyva-site repositories, and the rest of the empty set are placeholders: names reserved, structures sketched, code not yet written. A landing stub is a different thing from an empty repository, and the review distinguishes them carefully. The diagnosticx-site, genomic-site, and robotics-site repositories are landing stubs, short pages with no system behind them. The shield-site, insidr-site, and trueforge-site repositories are landing stubs too. The FiduciaryCorporateShield and Insidr-W3LLC repositories are two-line READMEs, title only, with no substantive code reviewed behind them. Each of those is a different kind of not-yet, and the review records which kind it is, because a stub presented as a system would be a lie, while a stub presented as a stub is a plan.
The emptiness is concentrated where the family is thinnest. Twenty-five of the 33 series have no repository in the organisation at all, and the articles section that covers them says so, series by series, and focuses on the DAO and stewardship model and the open questions rather than inventing a product. The thin series are name and number only, and their articles do not guess at what the names mean. Two of them, Moscone.X and Donatello DAO, carry no guessed name associations at all, because the name is all that is verified. The CMMC acronym is never expanded, because no expansion appears in the files. AskDoc.X is explicitly not a diagnosis or treatment tool. The genomic platform’s ambiguity between the two genomic series stays flagged, not resolved. Biota, an anchored world with coordinates, has no series number, and none is assigned to it by assumption. Each of those sentences is a small act of verification: the absence is recorded, the guess is refused, and the reader is told what is not known.
The hackathon layer shows the same habit at a different scale. The July 2026 repositories are starter templates and hackathon entries: EdgeOne and Tencent templates, CrewAI starters, and entries to named events. EventAssist is a solo entry to the Patch the Conference Hackathon on 2026-10-24. Sapient.X is an entry to the Itchathon on 2026-09-27. Sentinel Stream is an entry to the Cyberdefense Hackathon of October 9, 2025. Metropolis-SmartCiti is an entry to the Monad Metropolis Track 04. The review records them as entries, not wins, because no outcome was verified for any of them, and a hackathon entry presented as a prize would be the oldest trick in the book. An entry is a real artifact: code written under a deadline, judged by strangers, and left in public for anyone to read. A win is a different claim, and it needs a different source. The culture’s rule is to claim the entry and not the win, and to let the code speak for the entry.
What the empty repositories are for, in the end, is the same thing the deliberately-broken controls are for: they make the gap visible. A family that lists 33 series and shows code for 8 of them is a family with 25 series still to build, and the honest statement of that fact is more useful to a future builder, a future donor, or a future regulator than a polished page that implies all 33 exist. The empty repository is a reservation, not a deception, as long as it is labelled as one. The stub is a sketch, not a system, as long as it is labelled as one. The verification culture’s contribution is not that it fills the gaps. It is that it refuses to paint over them, and that refusal is what makes the filled parts believable.
The audit that corrects its own paperwork
The verification culture does not stop at code. On 2026-11-03, the family’s corporate paperwork was put through a truth audit, and the audit’s method is the same as the eval gates’: sort every statement into verified, planning, or corrected, and publish the corrections. The verified column is short, and it is specific: the Foundation as a filed California Nonprofit Public Benefit Corporation, Entity No. B20260196664, articles filed 2026-04-24, Secretary of State approval 2026-04-30, with an EIN assigned 2026-05-07; the Fellowship as a filed California Nonprofit Religious Corporation, Entity No. B20260229217, articles filed 2026-05-13, approval 2026-05-18; AGI Corp as an EIN only, with no Secretary of State articles in the uploads, so that an EIN alone does not prove incorporation; and the W3 series list as 33 named sub-series, Series 1 through 33, in an internal April 2026 printout, not a Wyoming filing, with Biota not appearing in it and no series number assigned to it by assumption.
The planning column is longer, and it is where the audit does its most useful work, because it labels as planning everything that is often presented as fact: all valuations and raises, including the Foundation’s for-profit benefit-corporation track at $50 million and the SmartCiti.X corporation at $25 million suggested pre-money; all tranches, milestones, allocations, and portfolio phases; all ownership illustrations; the Windfall and Open Global Investment blanks; the intended structure; and the SmartCiti.X corporation’s open choices of state, benefit-corporation form, Tranche 1 instrument, Tranche 2 price mechanism, and Series 10 disposition. The corrected column is the audit’s public record of its own earlier errors: the SmartCiti.X issuer corrected from Series 10 to the new corporation; the geography corrected to Louisiana, Nebraska, and Indiana; the Foundation SAFE corrected out of existence, because no SAFE may be issued by the filed nonprofit; the Russ identity corrected to unmerged, with Nightshift Russ, the Luma co-host Russ, and Russell Vacanti kept apart until one of them confirms a full legal name in writing; the real-estate and launch status corrected to no site, no union, no customer, and no revenue; investor ownership of real estate corrected to shares, not parcels; the Bino role corrected to advisor, not co-founder, director, officer, or chief financial officer; and the review status of eight files corrected to not sent, not reviewed.
The audit also states its own limits, in the same document, and that is the final habit this feature records. It does not verify a filing, a bank account, a Stripe account, a site, a union, a customer, a valuation, or a return, because those require the underlying document, account, agreement, appraisal, or signed contract, none of which was in the files at the time of the audit. It does not close any open decision, from the benefit-corporation fork to the accounting system. And it does not send itself to anyone. A verification practice that states its own limits is a practice that can be trusted, because it has shown you where it stops. The family’s signature, if it has one, is not a number. It is the sentence that follows the number, the one that says what the number is, where it came from, and what it does not prove. The proof is the product, and the product is the practice of showing your work.
What this is / What it isn’t
What this is
- A verification culture as the family signature: self-tests, CI rebuilds, eval gates run against deliberately broken controls, and a truth audit that corrects the family’s own paperwork in public.
- Attributed numbers, never settled ones: 9,267 is the bundle’s own self-test suite, the Cognition counts disagree, the Metropolis record is testnet only, and Sentinel’s less-than-two-seconds is a design goal, never a benchmark.
- Real deployed artifacts, celebrated as what they are: 41 curated worlds on Pages, a working FastAPI and ClickHouse system, six contracts with 41 of 41 testnet transactions confirmed, and a live Fellowship site with 40 articles.
What it isn’t
- Not independent verification: the bundle’s self-tests are its own, no CI run was checked, and no prize is claimed for any hackathon entry, because no outcome was verified.
- Not a complete record: only 8 of 33 series have any repository coverage, 10 repositories are empty, and 21 claims were flagged in the review.
- Not live on RP1: zero worlds are live on RP1, the 544 WorldForge entries all carry no RP1 attach, and no single AGI Corp world count is quoted, because the counts disagree.
Where to see it
The deployed artifacts are on the WorldForge, Metropolis-SmartCiti, and Sentinel Stream project pages, with the governance and operations discipline on the Governance hub and the Operations hub. The repositories are public on GitHub: WorldForge-Browser, serving 41 curated worlds on Pages; Metropolis-SmartCiti, six contracts on the Monad testnet with 41 of 41 transactions confirmed; sentinel-stream, the working FastAPI and ClickHouse system; Cognition.X, the curriculum of verifiable blocks, with the block counts disagreeing between repositories; and SmartCiti.X, the bundle whose own self-test suite runs 9,267 checks.
Nothing on this page is legal, tax, or financial advice. Every number here is attributed to the repository that states it, and every disagreement between repositories is stated rather than resolved. No prize, no benchmark, and no independent verification is claimed.