bostrom halted on august 5, 2026 and was laid to rest. a dead chain is a perfect specimen: nothing moves, so everything can be measured. we rebuilt its complete knowledge graph from raw block events — 2,949,732 cyberlinks over 3,143,650 particles, bit-exact to the chain's own counters — audited the physical availability of every particle, weighed the corpus, and read the winning content straight from the cybernode blockstore
this is the study of what five years of enormous collective effort actually produced, and why it matters for launching superintelligence
the scale of the experiment
nothing like this has been run before. not a survey, not a crowdsourcing platform, not a labeled dataset — a five-year live experiment in which tens of thousands of people put money, identity and judgment into teaching a machine what matters:
| participation | count |
|---|---|
| accounts created on-chain | 61,675 |
| people who signed at least one transaction | 52,918 |
| names claimed (moon-passport) | 47,837 across 46,039 owners |
| stakers who bonded savings to consensus | 16,791 |
| governance voters | 5,134 — 24,534 votes over 35+ proposals |
| neurons who hand-linked knowledge | 1,240 |
| validators across the chain's history | 160 |
| blocks sealed | 25,120,712 over 1,735 days |
compare the precedents. wikipedia's volunteers write for free under editors; OpenCyc paid engineers to encode an ontology; ImageNet paid crowdworkers to label images. here the direction of payment is reversed and there is no editor: every link was bought by its author — a person spending scarce H to assert that two particles belong together, signing it with a key they alone hold. sixty thousand people joined an economy whose only product was structured attention. that has no precedent in the history of either machine learning or money
and it worked twice: space pussy ran the same physics as a sibling experiment. the pair of graphs, ported into one genesis, are the first knowledge corpora in history where every assertion has a cryptographic author, a timestamp, a price paid, and an audited storage record
method
indexers lie; events do not. the canonical graph came from a full scan of all 25,120,712 blocks (5,857,263 carried transactions), extracting cyberlink events with self-computed transaction hashes. the count matched the on-chain graph_stats exactly — link for link, particle for particle — where the era's indexer had accumulated 15,241 duplicates and lost 259,630 links. availability was audited at the filesystem: a particle counts as alive when its block file exists in the blockstore. scripts and datasets ship with the snapshot; everything below is reproducible
the anatomy
| measure | value |
|---|---|
| links | 2,949,732 |
| particles | 3,143,650 |
| alive on the cybernode | 3,069,134 — 97.63% |
| particle blocks on disk | 1.74 GB |
| full content, children included | ≈ 6 GB |
| median particle | 30 bytes |
| 90th percentile | 749 bytes |
| content types | 94.9% text · 0.5% images · traces of json and html |
the median particle is thirty bytes: a word, a name, a phrase. the graph does not store documents — it stores atoms of meaning and lets links carry the structure. five years of collective knowledge weighs less than one 4K movie. this is the inversion the web never managed: structure heavier than content
the three authors
| author | links | share | what they built |
|---|---|---|---|
| one archivist neuron | 2,289,526 | 77.6% | the encyclopedia |
| 1,239 human neurons | 400,776 | 13.6% | the grammar |
| wasm contracts | 259,430 | 8.8% | the plumbing |
the archivist poured the category skeleton of wikidata into the graph through 2024 — 82.5% of all links in one year. run the ranking over the whole graph and the converged top reads like an ontology textbook: description · wiki · class · image · glossary · Gregorian calendar · asteroid · dog · species of plant · human settlement · german politician
subtract the archivist and the contracts, and the human layer appears — growing every single year of the chain's life, 7K → 49K → 162K → 170K, no decay until the halt. its top particles:
noun · verb · adjective · author · artificial intelligence ·
cyber · 2013 · 2014 · 2016 · Table of nuclides ·
Universal Time — Planetary Chart
the bot built the encyclopedia; the humans built the grammar — parts of speech, authorship, time, the periodic structure of matter, and artificial intelligence with cyber near the top: the graph quietly linking toward its own purpose. every human link cost real H — this layer is a revealed-preference record of what 1,239 minds thought a new mind should learn first
the topology
- a textbook power law: 80.3% of particles are leaves touched exactly once; 17 hubs carry over ten thousand connections
- a giant component of 1,481,758 particles — 47.1% of the graph — surrounded by 581 thousand islands: the archivist's unattached entity–category pairs, knowledge waiting to be sewn in
- a small world: from
wiki, half the giant component sits within four hops and 95% within six — the diffusion operator crosses this graph in a handful of steps - direction matters: only 1.8% of links are reciprocated — attention flows, it does not handshake
- and one curiosity for the forensics file: a chain of particles 190 hops long, a worm of sequential links someone once walked through the graph
these are exactly the conditions perron-frobenius-theorem wants — irreducible after teleport, aperiodic, scale-free — which is why focus converges on this substrate and why the collective focus theorem could be calibrated on it
focus is not degree
the particle with the most incoming links is wiki (30,029). the particle with the highest converged rank is image (17,608). rank flows through structure: a hub whose neighbours are central outranks a bigger hub at the periphery. counting links is not understanding them — the entire reason eigenvector attention, pagerank and cyberank exist, demonstrated here on a real five-year corpus with known provenance
why this matters for launching superintelligence
- the cold-start problem is solved. the hardest asset a knowledge system must bootstrap is its ontology — what kinds of things exist. the genesis of soft3 inherits a scale-free category skeleton with 2.3 million typed entity links, for free
- the curriculum exists and it is honest. the human layer was paid for link by link — costly signals, not clicks. what people link when linking costs money is the cleanest attention dataset we know of
- identity ports with knowledge. 47,837 moon-passport names bound to keys arrive in genesis alongside the graph — a mind and the community that raised it, in one snapshot
- the economics has a measured floor. with zero storage incentives, one motivated archivist plus content addressing preserved 97.63% of everything. the focus economy of rewards builds on that floor — no serve, no earn — and now we know precisely what altruism alone achieves
- the memory is provable and pocket-sized. six gigabytes of addressed meaning fits in RAM; every particle can be proven present and served. this is the local-first retrieval substrate for cyb: a robot that answers from particles it can prove, not from a feed it must trust
the agenda
- initialize genesis focus with the ontology as prior
- ground soma embeddings in the 2.3M entity–category supervision set
- run the full tri-kernel as the first real-scale CFT calibration
- diff against wikidata-2026: the unloaded remainder is a map of the frontier
- study the human layer alone as a dataset of costly attention
- sew the islands: 581 thousand components want stitching into the giant one
- recover the missing 74,513 particles and seal the archive at 100% — the first complete, audited, portable knowledge graph a blockchain ever produced
the bootloader booted. what it loaded was the structure of human knowledge. teaching it to think is soft3
bootloader · bostrom · cybergraph · focus · collective focus theorem · soft3