Methodology
How the atlas works.
A generation cascade from authoritative standards to the long tail of AI-native startups. Grounded in real value-flow data, generated on demand, and honest by default.
01 — The model
Every URL is a coordinate.
The atlas is built on a single primitive: the Coordinate. A Coordinate is a URL read as a partial address into a 16-dimension ontology — a path like /CompanyTypes/Accounting_Firm/Problems/Reconcile_Bank_Statements that pins a CompanyType and a Problem at one junction. Visit it and a Thing generates: a typed entity with a description, structured Facets, and links into the rest of the graph.
Relationships between Things live in Prompt Contextas free text — never as foreign keys. An ancestor Thing’s Name, description, and cached Facets flow down its descendants’ coordinate paths into their prompts. Any coordinate, even one no one has visited before, generates on demand. Caching makes the second visit free.
02 — The cascade
From a standard to a startup, in six steps.
The discovery cascade runs through six typed stages. Each step feeds the next as Prompt Context; the cascade is the workflow, and quality compounds (or degrades) along it.
A finite, authoritative source ontology — NAICS, O*NET, APQC PCF, UNSPSC, NAPCS, or GS1 — supplies the seeded spine: industries, occupations, tasks, processes, products. Standards are READ, not generated.
Derived archetypes (Accounting_Firm, Manufacturing_Plant, Architecture_Studio) where the customer profile starts crystallizing. Composed from L2 facts from the seeded spine; description filled by the overview Facet when there is no Standard match.
An Ideal Customer Profile expressed AS a coordinate — a sparse multidimensional junction (CompanyType × JobType × Department × Locale × Buyer Chain) read as a persona-in-context. The URL is the join; no FK rows.
The pain a sufficiently-pinned ICP exposes. Generated by the ICP coordinate's Problem-Facet; first-class Noun at /Problems/<name>. Same Problem text may appear under multiple ICPs — the Canonical Index dedups generalizable ones via semantic merge.
The Problem-as-buildable-gap — the space of Founding Hypotheses for a Problem. Not Thesis × ICP; the (Problem × Thesis) view is a navigable filter over Startups, not a separate generation stage.
One positioned entrant in the Opportunity space. A Founding Hypothesis = Problem × Thesis × differentiation (tier / wedge / GTM / Position), with a generated Offer (price band + deliverables + guarantee).
03 — The standards
Grounded in five public authoritative ontologies.
Every Thing whose Name matches a Standard inherits the Standard’s authoritative description. Only the gaps are generated. This means the atlas’s spine — every Industry, every Occupation, every cross-industry Process, every commodity type — is grounded in the same vocabularies governments and businesses already use to classify work.
North American Industry Classification System.
Industries — every cell of the US/CA/MX economy, by official sector / subsector / industry-group / NAICS-industry / national-industry levels.
U.S. Census Bureau
Occupational Information Network.
Occupations and the Task → DWA → IWA → GWA chain of work activities. The richest public catalog of what a job actually involves.
U.S. Department of Labor
Process Classification Framework.
Cross-industry processes — Category → Process Group → Process → Activity → Task — with the ~3,900 Metrics that measure them.
American Productivity & Quality Center
United Nations Standard Products & Services Code; North American Product Classification System.
Commodity classification — products and the coarse classified offerings that are sold against them. Unioned via same-name multi-standard merge.
UN Development Programme; Statistics Canada / U.S. Census Bureau
Core Business Vocabulary (CBV).
Value-flow lifecycle — the verbs that describe how a product is produced, transformed, shipped, observed at a place, and consumed.
GS1
Structured data vocabulary.
JSON-LD typing — Service, SoftwareApplication, Occupation, DefinedTerm, HowTo, Article, Dataset, Offer. Every page emits richly-populated schema or none at all.
schema.org
04 — The generation model
On-demand, never pre-warmed.
Coordinates generate when they’re visited, not in advance. This is structural, not a performance trick. Two layers govern the surface:
Any URL generates a Thing on demand and caches it. Never gated by an index, never blocked. Duplicates allowed at this layer — every visit lands somewhere, even an invented coordinate.
Curated, de-duplicated overlay evaluated by multiple differentiation signals (Problem distinctness, facet divergence, the generated Startups, human curation). Powers browse / search / sitemap as a discovery overlay; never blocks a URL.
The on-demand model is paired with a strict no-pre-warm discipline. We never batch-generate every coordinate in advance — schema churn during iteration would force regenerating the entire corpus 50× over. A full pre-warm sweep is reserved for once, at frozen-schema launch.
05 — The discipline
Four principles that govern every generation.
Always log every generation.
Every LLM call we make is append-logged to durable storage at the moment of generation. Gate verdicts and slop checks are metadata, never gatekeepers. Losing real output to a heuristic verdict is the failure mode we design out.
Capability-confident, never fabricated.
Pages make present-tense claims about what the AI-native economy can do, today. We never invent a customer, a metric, a press mention, a fake testimonial, or social proof we haven't earned. The atlas is research; we describe the research, not a marketing pitch.
Populate richly, or omit.
Schema.org JSON-LD is populated with load-bearing facts — names, descriptions, codes, sameAs links, dates — when we have them, and omitted entirely when we don't. Minimal schema underperforms NO schema, so honest omission is the discipline.
Cascade quality is end-to-end.
The cascade amplifies slop. A Facet schema change is verified across every downstream hop — Subject → Problem → Startup — not just on the modified step. The cascade is the workflow; quality is a property of the chain, not any single call.
06 — Reliability gates
The cascade IS the workflow.
Hallucination is controlled by wrapping non-deterministic generation in deterministic structure — never by trusting a single model output. Three settled rules:
No bolt-on state machine.
The cascade is the workflow; lifecycle (concept → live) is a status the cascade advances on the Startup Id, not a separate engine.
The gate decision is deterministic, never generative.
A deterministic Compliance Filter scans every generated Service, Agent, and Startup for regulatory exposure; whatever it flags requires a Human Oversight Layer sign-off before going live. A non-deterministic process must never decide whether its own output is safe.
Divergent independent review is selective.
Running N independent reviewer agents on every hop doubles generation cost, so it’s reserved for two triggers: confidence/validity failures (cheap-check or self-consistency probe), and graduation to a live transacting Startup. Cheap deterministic checks (schema validity, Cross-Section link resolution, sampled monitoring) cover the rest.
07 — Discoverability
Tiered indexing.Cited sources.Real dates.
We publish a directory of millions of generated pages. Indexing thin or near-duplicate pages drags site-wide quality signals and wastes crawl budget. So:
- Tiered indexing.Differentiated pages are indexed and listed in the sitemap; thin or near-duplicate pages stay crawlable for link flow but noindex’d and out of the sitemap. Crawl budget concentrates on the authoritative surface.
- Richly-populated JSON-LD. Every indexable page emits schema.org entity types (DefinedTerm, Occupation, Service, SoftwareApplication, HowTo, Article, Dataset) with load-bearing fields filled in. We never emit minimal schema — minimal schema underperforms no schema.
- sameAs to the canonical graph. Every Thing whose code matches a standard emits sameAs links to the authoritative source: census.gov for NAICS, onetonline.org for O*NET, apqc.org for APQC.
- Freshness signals.Every page carries a real dateModified (UTC, day-grain) from the page’s last-meaningful-update timestamp. Pages with no signal omit the field rather than fabricating one.
- AI-crawler welcome. robots.txt explicitly allows GPTBot, ChatGPT-User, OAI-SearchBot, ClaudeBot, anthropic-ai, PerplexityBot, Google-Extended, CCBot, Applebot-Extended, and Bytespider. Every coordinate exposes a Markdown alternate that AI engines can read cleanly.