00  The canonical layer of the .si namespace

Nuclei.si

451 .si names, one address per term. About four in five are common English words; the balance are acronyms, technical coinages and proper names held for the same families. 65 already serve structured, signed, callable records — and every one of those 65 is open to inspect on this page.

Namespace

.si — Slovenia’s ccTLD

Interface

HTTP · JSON-LD · MCP

Operator

Humble Superintelligence

0Names in the register
0Properties serving
0Topical families
0Held for build
01  The lattice

Every concept, at once.

All 451 names in our register, packed into one hexagonal field spiralled from the core. Sort alphabetically, by topical family, or by state, and the layout re-sorts in place. Hover any cell to resolve it; click a serving cell to open its live view.

Filter
Sort
Nuclei register · hex packing · 451 cells
serving held in register outside filter colour indicates family
02  What this does to machine intelligence

A model can only be as grounded as the web it reads.

Frontier systems are no longer bottlenecked purely on compute. They are bottlenecked on grounding — the ability to resolve a claim to a source that can be identified, dated and verified. The open web is moving in the opposite direction: filling with generated text that carries no provenance, no signature and no stable address.

The failure mode
What a canonical layer changes
What it enables
Our position

We are not claiming to fix machine grounding. We are claiming that the canonical layer has to exist, that whoever holds the exact-word addresses of a namespace is the natural place to build it, and that we hold that position in .si today — 451 names registered to Humble Superintelligence, 65 already serving structured, signed records.

Stated plainly: the case that agents privilege lexical addressing is structural, not measured. Rigorous studies of agent source-selection do not yet exist. What is measured is the demand side, and it is smaller than the headlines suggest — user-triggered AI crawling grew more than fifteenfold across Cloudflare’s network in 2025, though training crawls still ran at many times its volume; AI platforms sent about 1.13 billion referral visits to the top 1,000 sites in June 2025, roughly 0.6% of Google’s referrals in the same month. The exact-word price premium is documented in .com; we have no comparable dataset for .si and do not assume it transfers. We treat the agentic premium as upside, not as the base case.

Model collapse: Shumailov et al., Nature 631:755–759 (2024), with Author Correction Nature 640 (2025) · Accumulation result: Gerstgrasser et al., COLM 2024 (arXiv:2404.01413) · AI crawl growth: Cloudflare Radar 2025 Year in Review · Referral volume: Similarweb, June 2025, top 1,000 sites. Full citations in the investment memorandum.

03  Console

Query our register directly.

This runs against the full Nuclei register loaded in the page — resolve a concept, compute its nearest neighbours, list a family, sort the lattice above. Type help for the command set.

nuclei console — live
>
What it is actually doing

resolve reads the register entry and prints state, family, corpus size and interface addresses.

edges computes nearest concepts live — family match, state match, shared prefix and suffix length, penalised by length difference. The scores are the scores it calculated.

family, serving and sort drive the lattice above, so the console and the map are one state.

scroll → for serving counts
04  Live network

65 properties. Serving right now.

Every window below embeds the property itself, fetched live from its own address — not a screenshot and not a mock-up. Frames mount as they scroll into view. Click any window for the full-size viewer and its technical record. Document counts are pages in each property’s published corpus, drafted with model assistance against a named source list and reviewed before signing.

Windows load each property directly over HTTPS. Under every frame sits the property’s own record — corpus size and interface paths — so a window is never empty, and every window carries an Open site control that goes straight to the property whether or not it agrees to be embedded.

05  Protocol

Call it. Don't scrape it.

Every serving node exposes the same interface, so an agent that can talk to one property can talk to all 65 — and to each new one on the day it ships. Read-only, content-addressed, signed at the document level.

gateway — illustrative
# resolve a concept through the network gateway
curl -s https://physics.si/api/v1/concept \
  -H 'Accept: application/ld+json'

{
  "@context": "https://schema.org",
  "@type":    "DefinedTerm",
  "name":     "physics",
  "sameAs":   "https://www.wikidata.org/wiki/Q413",
  "provenance": { "hash": "sha256:4f9c…a1" },
  "edges": ["cern.si", "trigonometry.si"]
}
 How a nucleus is built

Six stages, identical for every property. The pipeline is why we expect the marginal cost of the next build to approach the cost of its content — an expectation based on the 65 built so far, not on the full register.

06  Access

Licence a concept. Or the whole lattice.

Individual endpoints, topical families, or federated access across the serving network. Operated by Humble Superintelligence, the registered holder of every name in the register.

01

Endpoint licence

Exclusive or non-exclusive use of a single concept address, with the property and its corpus, on a term licence.

02

Family acquisition

A topical family transferred as one block, with whatever serving properties it currently holds — 39 in faith, 9 in science, 4 in civic, 3 each in health and earth, 2 each in law, figures and language, 1 in people.

03

Network access

Federated retrieval across the serving network through one gateway and one MCP surface. The gateway and telemetry layers are in build; access is offered on a forward-dated basis with delivery milestones stated in the agreement.

A domain is a brand to a person. To a retrieval step it is either a lookup or a guess.
Start a conversation