How we decide who's in the directory and how the map connects them.

What these databases are for

The Organization Directory, Connect and the Affinity Map make the people and organizations working on state capacity visible to each other, and together they make up the Ecosystem pillar. The mission and vision are on the About page.

What state capacity means here

We include organizations that help government become more capable and exclude organizations that do government's work for it. An organization is added if it helps government in the United States expand its own effectiveness or operating capacity.

Data fields

• Organization Name: the name of the organization.
• Primary Segment: primary categorization based on the organization's focus.
• Secondary Segments: additional areas the organization works on.
• Geography: geographic level at which the organization operates.
• Description: a 1–3 sentence summary of what the organization does.
• Funding Model: how the organization is funded.
• Funding Detail: free-text description of funding detail.
• Website: the organization's homepage.
• Problem Area: problem topics classified into 7 buckets: Participatory Democracy, Procurement & Operations, Service Delivery, Talent & Hiring, Technology & Data, Test & Learn, Domains.
• Problem Topic: 36 specific issues nested under the Problem Areas.
• Status: active, or closed when the description says the organization has closed. For now we infer it from the description text.

Segment taxonomy

An organization's segment describes its primary operating model.

• Advocacy: Organizations pushing for policy and structural changes that enable better government.
• Capacity Building: Groups training and strengthening the government workforce.
• Community: Networks and communities of practice connecting government practitioners.
• Digital Services & Consulting: Firms doing hands-on technology and service design work with government.
• Ecosystems: Connectors and conveners building the field itself.
• Fellowships: Programs placing talent inside government agencies.
• GovTech: Companies building software and tools for government customers.
• Government: Public agencies and internal teams leading change from within.
• Investor: Funds and investors backing companies in this space.
• Philanthropy: Foundations and funders supporting the broader ecosystem.
• Research: Organizations studying what works and building the evidence base.

Problems taxonomy

An organization gets a problem tag if it solves that problem directly for government, funds or builds the infrastructure that improves it, or sets the standards that govern it. The problem has to be part of the organization's mission. Side projects and one-time grants do not count.

The directory

The Organization Directory shows the underlying data as a table, with filters, a search box and a detail row for every organization:

• Search accepts an organization name or a natural-language phrase ("procurement reform in NYC", "AI in benefits delivery") and ranks the table by relevance using the same engine as the Search page (see Search below).
• Primary segment, Geography, Problem area and Problem topic are multi-selects; an organization matches if it carries any of the selected values.
• Hide closed organizations removes organizations whose description says they have closed. They stay in the directory as history, marked with a Closed chip.
• The detail row shows the full record, the three most similar organizations from the affinity score, the people and opportunities on Connect from the same organization (or, failing that, on the same problem topics), and a Suggest an edit button that opens the submission form prefilled with the current values.
• Every filter, the search text, the page and the open row live in the URL, so Copy link gives you a view you can send to someone else.

Connect

The Connect page is a directory of individuals and open opportunities: practitioners, researchers, funders, and others who have opted in to be findable, alongside open calls (grants, fellowships, jobs, collaborative projects) indexed by problem and offering type. It is separate from the Organization Directory: organizations belong in the Directory; people and opportunities belong in Connect.

Each entry carries: name, organization affiliation, role, offering, problem area(s), problem topic(s), geography, due-by date, details, and a contact. When a person asks to keep their contact private, the entry shows Request intro instead and the SCE team forwards the request. Entries are submitted through the on-page form and reviewed before going live. An organization name that matches one in the Organization Directory links to that organization's row, and filters, page and the open entry live in the URL for sharing.

Change feed

The directory refreshes daily. Each run compares the new spreadsheet to the last one and records which organizations were added, removed, or edited. The latest changes appear on the Ecosystem page.

Search

One search box ranks organizations and the people and opportunities on Connect together. It lives on the Search page, in the Ecosystem page's search panel and behind the directory's search box. Type a problem, a place, or a name; filter by what to show, segment, problem area or topic, and geography; every state lives in the URL so a result set can be shared.

Search opens on a guided sentence: "I am a/an [role] looking for [anything] working on [any problem] at [any] level." Any one choice is enough. The goal sets a visible filter (funding shows Philanthropy and Investor organizations and Connect entries offering funding or grants; tools and vendors shows GovTech and Digital Services organizations; and so on for nine goals), the problem and level narrow it, and the role only suggests a goal. The result is a starting point: links that open the Organization Directory, the Affinity Map and Connect with those filters applied, plus the top organizations and people, ordered by how many of the chosen topics they carry. Closed organizations are left out. The second tab is the keyword search described next.

How keyword search ranks. At build time, every organization's description, problem-topic tags, segments, and funding-detail text are tokenized into one bag and a TF-IDF vector is stored in a static file of about 190 KB; the people index is built the same way in the browser from Connect entries. At query time the browser tokenizes the query, builds an IDF-weighted vector, and takes the cosine similarity against every organization and person. Then four adjustments: an exact substring match on a name adds 0.5; a query that contains a full topic or area name multiplies tagged organizations and people by 1.75; a place name that appears in an organization's name or description (a state, a major city, NYC) multiplies by 1.5, and a level word (state, federal, local) multiplies by 1.25 for organizations at that level; a query about money (fund, funder, foundation, philanthropy, investor) multiplies Philanthropy and Investor organizations by 1.5. Multiplying rather than adding keeps an irrelevant org from outranking a relevant one on geography alone. Each result says why it matched.

Search is keyword-based: it cannot know that "permits" and "licensing" are the same concept unless the words co-occur in the corpus. The affinity score below uses sentence embeddings for exactly that reason; search stays on TF-IDF because the index has to run in the browser with nothing to download. Two-letter terms that matter here (AI, ML, UX, HR, DC, EV) are kept; other tokens under three letters are dropped.

The affinity score

The map draws a line (an edge) between two organizations when they appear to be working on the same underlying problems. Four signals are computed for every pair, each is converted to a percentile so the signals are on the same scale, and the percentiles are combined with these weights:

Every signal points at one question: what is this organization trying to accomplish, and who else is trying to accomplish the same thing? Topics say what they work on, descriptions say the substance, funders say who believes in the approach, and segments say the mechanism. Percentile scaling matters because the raw signals live on different scales: two organizations that share one of two tags score 0.5 on topics while two 30-word descriptions rarely score above 0.5 on cosine, so without scaling the tag signals would decide nearly every edge regardless of the weights.

Description similarity (40%)

Each organization's name, description and topic list are embedded with the open sentence-embedding model all-MiniLM-L6-v2 at build time, and the signal is the cosine similarity between the two vectors. Embeddings recognize that "permitting reform" and "faster approvals for builders" describe the same work even when no word overlaps. The build records which text signal was used in the data file; if the model is unavailable the build falls back to TF-IDF cosine and says so.

Shared problem topics (30%)

A weighted Jaccard over the two organizations' topic tags. Each of the 36 topics is weighted by how rare it is (the log of organizations in the directory over organizations carrying that topic), so sharing Child Welfare, which a handful of organizations carry, counts for more than sharing AI in Government, which many do. Because the tags are curator-assigned, this is the highest-confidence signal in the dataset. The coarser Problem Area bucket is not used here; it is too broad to separate anyone.

Shared funders (15%)

Funders are pulled from the free-text Funding Detail two ways: a curated list of well-known philanthropies, foundations, and venture funds (Schmidt Futures, Open Philanthropy, Omidyar Network, Ford, MacArthur, Knight, Bloomberg, Arnold Ventures, and others), and a pattern that catches any capitalized name ending in Foundation, Fund, Ventures, Philanthropies, Trust, or Initiative. The signal is the Jaccard similarity of the two funder sets, and it is zero when either organization has no detected funder.

Shared segments (15%)

Plain Jaccard over the union of each organization's primary and secondary segment tags, with no boost for a matching primary segment. The weight is small on purpose: segment overlap is the obvious thing (research shops look like research shops), and the map exists to surface the cross-segment connections.

Which edges are kept

Every organization keeps its six strongest pairs, and the map's edge set is the union of those lists. An edge is mutual when each organization is in the other's top six; the map draws mutual edges solid and one-way edges faint. There is no minimum score and no cap on how many lists an organization can appear in, so a fellowship program that many organizations resemble can hold twenty edges while a niche vendor holds six. Organizations whose description says they have closed are marked closed, drawn hollow, and can be hidden.

The published dataset has 334 nodes and 1,376 edges, with a maximum edge score of 0.97 and a median of 0.74. Of those edges, 628 are mutual and 28% connect organizations in different primary segments. The map opens at the 25th-percentile score so the weakest quarter of edges is hidden until you lower the slider.

Inside the score table

For reference, here is what each signal ranges over before percentile scaling:

The map's tools

• Why connected. Click a line, or an organization listed in another organization's panel, to see the shared topics, shared funders, and shared terms behind it, with the four component scores.
• Only links between different segments. Hides every link whose two ends share a primary segment, leaving the connections the score was built to find.
• Show direct connections and Show connections two steps out. These restrict the map to one organization and the organizations around it.
• Find a path. Names two organizations and shows the shortest chain between them, with each hop explained. This is how an introduction route gets found.
• Funders view. Redraws the map with funders as squares joined to the organizations they back, showing funders that back two or more organizations. It is only as complete as the funder detection above.
• Copy link. Every filter, the threshold, the view and the selected organization live in the URL.

What the graph does and does not show

It shows structural affinity: which organizations work on the same underlying problems, describe them in similar ways, and often share funders. The strongest edges are between organizations that anyone in the field would put in the same cluster, but the value of the graph is in the edges that are not obvious: a research shop and a GovTech vendor on the same problem, or a fellowship program and a capacity-building organization sharing a set of funders.

It does not show documented relationships. The data does not encode partnerships, joint programs, alumni flows, board interlocks, or co-investments. An edge is an inference from public descriptions and tags and does not mean two organizations work together.

Color palette

Primary segments are color-coded consistently across the directory's segment column and the Affinity Map's circles. The mapping:

• Research
• Government
• Philanthropy
• Fellowships
• Community
• GovTech
• Advocacy
• Digital Services & Consulting
• Investor
• Capacity Building
• Ecosystems

Credits

The Organization Directory, Connect and the Affinity Map are built and curated by the State Capacity Ecosystem (SCE) team. The visualization and interaction layer is maintained by Tal Roded.