Last updated: 2026-07-25
This is the multi-region coverage plan. It sits above the Europe-specific
global-expansion.md, which owns the European GTFS beta
gate in detail, and it inherits the operating gates in roadmap.md
and the product principles in product-roadmap.md. Where
they conflict, those documents win; this one sequences the regions and names the
gate each region must pass before it is curated.
The service began as a Yolo County pilot, grew to national United States and Canadian coverage, and now carries a reviewed European cohort and a handful of worldwide canaries. "Support other regions" is easy to say and easy to get wrong. It is not one job. It is at least four:
- curating official, openly licensed feeds a region already publishes;
- sourcing feeds a region publishes somewhere other than the Mobility Database;
- deciding whether a region needs its own bounded beta gate or is served by ordinary coverage growth; and
- deciding when a region must not be curated by this project at all until a local steward owns consent, licensing, and regional context.
These have different evidence and different gates. This roadmap keeps them separate so the map never fills faster than the review behind it.
- One universal scoring core. Every region is scored by the same GTFS Schedule rubric and the same canonical MobilityData validator. Regional policy modules (United States NTD, Canadian equity) stay conditional on the country and never change another region's grade. Adding a region must not touch weights, thresholds, or the scoring profile (ADR 0026, ADR 0035).
- Defensibility before breadth. A region is curated in the order its feeds can be defended: official first-party source, reviewed reuse terms, retained attribution, checked identity. A larger flag count is never the reason to relax that order.
- Coverage is not a success measure. Feed count, country count, and map fill are distribution facts, not proof the tool helped anyone. The success measures stay the remediation and decision-support outcomes in the product roadmap. This document exists to grow coverage honestly, not to make growth the goal.
- The local-steward gate is real, not a formality. For any region where the
data is community-mapped, where informal transit dominates, or where reuse
and consent cannot be established from a public official source, this project
does not curate until a named local steward owns licensing, source
verification, and the regional consent or partnership requirement
(
roadmap.md"Later: demand-gated options"). This is a values commitment: a tool built in California does not arrive uninvited to grade the Global South's transit data. - Never a census. A regional cohort, however large, is a reviewed sample. The interface states the regional denominator beside every filter and export and never describes a feed-record count as an agency, operator, or coverage census of a country or region.
- Fail closed. Unclear reuse terms, unreachable sources, unverifiable identity, and out-of-cap archives are excluded and the exclusion is recorded, not quietly dropped.
The Mobility Database catalog, the project's main discovery source, is heavily weighted toward Europe and North America. A 2026-07 pass over the catalog found, among active feeds with an open, non-key-gated download and no contrary official flag, roughly: 920 in the United States and Canada, 390 across 35 European countries, 38 in Oceania (almost all Australia and New Zealand), 25 across nine Asian countries, 29 across Latin America, 10 across the Middle East, and 9 across Africa.
That distribution is the plan's central constraint. It does not mean those regions publish little transit data. It means they publish it somewhere else: national and city open-data portals, regional aggregators, and community-mapping projects that never registered with the Mobility Database. Sourcing, not curation effort, is the binding limit outside Europe and North America. Each phase below names where a region's feeds actually live.
A 2026-07-25 snapshot puts a hard boundary around the current discovery
universe. Mobility Database V2
(SHA-256
86e46cb822b1aa11447837c2cdc87755ad8a6f571135d290095a62d4129a2834)
contains 2,574 active, keyless Schedule rows.
Transitland Atlas at commit
1df848a
(archive SHA-256
c3b93c14c7f4162eb9e8c386e8486f18c94cc2d4ecab07a5d10422e3e50afe30)
contains 3,739 Schedule rows with a current, keyless URL.
After applying the project's scheme-insensitive endpoint normalization, their
union is about 5,092 endpoints before content, operator, aggregate, redirect,
identity, and reuse review. The current registry contains more than 2,100 feed
records.
A literal 100-fold increase from that registry would require more than 200,000 records, far beyond the current public-catalog supply and contrary to this roadmap's curation rules. Here, a "100x coverage loop" means candidate-processing leverage: record a disposition for at least 95% of a declared, deduplicated eligible snapshot without weakening human review or auto-writing the registry. The loop is:
source snapshot and hash -> cheap preflight -> identity/content clustering ->
reuse decision -> 25-record canary -> three checks across seven days ->
recommendation audit and upstream catalogue corrections
Each new admission is paired with one legacy rights or identity audit until at least 90% of the registry has current evidence. Duplicates, mirrors, aggregates, and alternate URLs never count as extra coverage merely because they are extra catalog rows.
The phases are ordered by defensibility, and each ships as its own reviewed pull request. A later phase does not block on an earlier one, but the partnership-gated phase never runs on catalog scraping alone.
The pilots, national United States and Canadian coverage, the reviewed European
cohort (148 feed records across 17 countries as of 2026-07-17, tracked against
the beta gate in global-expansion.md), and worldwide canaries in Japan,
Malaysia, New Zealand, Australia, and Uruguay. The registry, artifact pipeline,
per-country program pages, and the world coverage map already carry worldwide
locations; the portable location contract (ISO 3166 country and subdivision,
Unicode names, bidi isolation, locale formatting) is in place.
Australia and New Zealand publish official government GTFS through mature open-data programs under Creative Commons Attribution or equivalent. They are anglophone, first-party, and openly licensed, so they raise no local-steward concern; they are the same kind of record as an official European feed. The first wave curates the readily fetchable state and metro feeds — the Queensland TransLink networks, the Northern Territory, Tasmania, Western Australia, and the New Zealand metros — and defers the Sydney and Melbourne bundles and any national aggregate that exceeds the ingestion size cap to the large-feed shard work below.
Gate: official government or provider source, an explicit open license (commonly CC BY 4.0), retained attribution, checked identity, within the ingestion caps.
Tasmania's previous license and fetchability hold cleared on 2026-07-22. The
Tasmanian Government now publishes one stable, keyless statewide aggregate on
its official transport site under CC BY 4.0. A full preflight found a valid
15 MiB archive for six listed publishers, bus and ferry routes, explicit
feed_info.txt dates, and service through 2026-09-30. The registry counts the
aggregate as one feed record, not six agencies.
Several Latin American authorities publish official GTFS under national "datos abiertos" or Creative Commons licenses through their own portals, only some of which are mirrored in the Mobility Database. The phase curates the official, openly licensed city and regional feeds — the strongest candidates are the metropolitan authorities that run their own open-data programs — and records the exclusions where a feed is community-hosted, key-gated, or lacks readable terms. It does not attempt whole-country coverage and it does not treat a metropolitan feed as national coverage.
Gate: official authority source, an explicit open or datos-abiertos license with retained attribution, checked identity, within caps. A community-hosted or terms-silent feed is deferred to the partnership-gated phase, not force-fit here.
Parts of Asia publish official GTFS through national data programs (for example Japan's national GTFS data service and Singapore's transit authority), and several Middle Eastern authorities publish under open terms. The phase curates those official, openly licensed feeds and records the many large systems that are key-gated, registration-walled, or published only through a consumer app as exclusions. It does not infer permission from a feed merely appearing in a catalog.
Gate: as Phase 2. Regions where the dominant standard is not GTFS, or where the official feed is only reachable behind an account, are out of scope for this phase and noted for the alternative-catalog work below.
The Hong Kong Transport Department aggregate is now the phase's frequency-based
schedule canary. DATA.GOV.HK publishes the keyless GTFS under terms permitting
commercial and non-commercial reuse with attribution. Its calendar runs to
2099, so admission required a full scorecard preflight: freshness remains
measured, explicitly labels the horizon unusually distant, and deducts for
missing feed_info.txt dates rather than treating the end date as evidence of
active maintenance. Hong Kong has no ISO 3166-2 child in the portable location
vocabulary, which exercises the supported country-only location contract.
Four BAS.MY town feeds cleared an earlier data-quality hold on 2026-07-22. Malaysia's official open API republished Alor Setar, Kota Bharu, Kuala Terengganu, and Kuching with hundreds of trips and service through 2026-12-31. All four keyless archives passed the canonical validator preflight under the same CC BY 4.0 platform and now add reviewed coverage in Kedah, Kelantan, Terengganu, and Sarawak.
Five official Japanese networks passed another bounded depth review on
2026-07-23: Kamishihoro Town Autonomous Bus, Konan Railway, Tsukubane-go,
Awaji Jenova Line, and Hokushin Bus. They use the three already admitted
versioned Creative Commons licenses and add one record in each of five
previously represented prefectures. The loop deliberately rejected Tsuku Bus:
its current archive still gives four required GTFS files .csv names, so the
canonical validator cannot recognize a usable service calendar. This is
regional depth from an official source, not a new-country or census claim.
Five additional licensed Japanese records passed an operational-depth review on 2026-07-23: JR East's Tsugaru Line replacement bus, Mizuho Town's reservation-based Choi-Soko service, Toyama Chitetsu's regional bus network, Toba City's island ferries, and Kochi's airport shared taxi. The set adds replacement-service, demand-responsive, regional-bus, ferry, airport-link, and keyless realtime evidence without changing the 40-prefecture denominator. Japan now has 235 reviewed feed records in the catalog. The seven prefectures without an admitted source remain empty; no unclear-license feed or mirror was used to change that count.
Five more licensed Japanese bus records passed a realtime-evidence review on 2026-07-23: Kaetsuno World Heritage Bus, Owari Asahi Asapy-go, Tokoname Gruun, Mie Kotsu's Shima-area service, and Kumamoto Dentetsu Bus. Each official catalog record supplies keyless TripUpdates and VehiclePositions endpoints. The admission check confirmed current schedule archives and endpoint responses, but does not treat a response as an uptime, prediction-quality, or trip-coverage guarantee. Japan now has 240 reviewed feed records across the same 40 prefectures.
Twenty official Japanese records passed a larger endpoint-kind review on 2026-07-23: thirteen CC0 services in Toyama and seven CC BY 4.0 services in Mie. Every current schedule archive passed the canonical validator path. The official catalog supplied 47 keyless realtime endpoints: all twenty records publish TripUpdates and VehiclePositions, and the seven Mie records also publish ServiceAlerts. Every endpoint returned a valid protobuf response during admission, although many were header-only at that moment.
The public feature contract now distinguishes the exact endpoint kinds that responded and whether the newest TripUpdates or VehiclePositions header was fresh. This is latest-sample evidence, not continuous availability, rider information content, prediction quality, or trip coverage. Japan now has 260 reviewed feed records across the same 40 prefectures. The batch adds regional depth and new product evidence without changing the national-coverage claim.
An eleventh repository pass then admitted 119 official schedule records from publishers with exactly one current feed. Every downloaded archive passed the pinned canonical validator path, carried one of the three already reviewed versioned Creative Commons licenses, and retained at least 60 days of effective service. Two catalog records with expired calendars, one near-expiry record, one private residential shuttle, and publishers represented only by several line or area fragments stayed out. Japan now has 379 reviewed feed records across the same 40 prefectures. This remains depth in a reviewed sample, not a census or a claim of national coverage.
A portal-first exhaustion pass on 2026-07-23 then added 70 reviewed feed records without weakening those gates. France's National Access Point supplied 57 local-network records under Licence Ouverte 2.0. Existing-country depth also grew in Canada (four records), Germany (two), and the United Kingdom (two). Official city and national portals opened four new country samples: Albania (Tirana), Moldova (Ungheni), Serbia (Belgrade), and Ukraine (Kyiv and Lviv). Every schedule archive passed the complete scorecard path and retained at least 60 days of effective service. Twenty-seven records also carry official, keyless realtime endpoints. This is a reviewed source cohort, not a country census.
The same pass exhausted the remaining plausible official/open catalog set and kept its failures visible. Short or expired calendars excluded Canberra Light Rail, Brussels STIB/MIVB, Luxembourg's national feed, and Metlink Wellington. Unreachable or non-archive sources excluded METROFOR, four French local networks, Mexico City, Hyderabad Metro, and several Portuguese feeds. Santiago remained out because its official current-feed page did not state an explicit commercial-reuse licence. Bogotá's very large feed did not complete a bounded local score and its catalog licence is share-alike outside the project allowlist. Ambiguous source/licence identity excluded Cascais and Barreiro. These are future recheck targets, not silent omissions.
A 20-loop continuation on 2026-07-23 then admitted 11 more records. Cyprus opened as a six-operator CC BY 4.0 sample, and TGSRTC Greater Hyderabad opened India under the operator's explicit commercial and non-commercial reuse terms. Existing-country depth added VBB's GTFS-Flex collection, Reus Transport, the Creuse regional network, and Belgrade's suburban feed. All 11 archives completed the pinned validator and full scorecard path with at least 60 days of effective service.
The continuation also made its exhaustion results explicit. A 73-candidate official/licensed catalog set yielded 55 local score artifacts and 18 bounded source or archive failures before evidence review. Sardegna's eligible portal feeds were already tracked. Duplicate Madrid and Sardegna records, GTT Torino's non-commercial terms, Kraków's unresolved commercial licence, Bydgoszcz's community-converter provenance, Portugal's source/licence mismatches, Bogotá's share-alike licence, Santiago's unstated commercial grant, and short or expired calendars in Australia, New Zealand, Belgium, Luxembourg, Canada, and Mexico all stayed out. The pass increases the reviewed sample without loosening the identity, reuse, validator, or calendar gates.
An API-level exhaustion of France's National Access Point followed on
2026-07-23. All 774 catalog datasets were considered, and all 331 currently
unmatched GTFS resources were run through the complete local scorecard and
60-day calendar gate. Of 147 mechanical passes, overlap and source review
retained 137: 136 French resources plus Tanéo in New Caledonia, which is
correctly modeled under ISO NC. Six regional aggregates, three duplicate or
superseded resource versions, and one anomalous stale export remained out.
France's reviewed European cohort grows by 136 records across 15 regions and
collectivities. Multiple resources belonging to one dataset share canonical
organization identity instead of being presented as extra operators.
The same loop opened Puerto Rico and Oman from official provider evidence. Puerto Rico ATI explicitly invites developers and companies to use its open-data GTFS. Mwasalat serves its current GTFS from its official Oman National Transport Company domain, labels its route-data surface as open data, and is covered by Oman's national license permitting commercial and non-commercial reuse with attribution. Both feeds passed the validator and calendar path. Grade was not an admission criterion.
The rejected set establishes the next boundary. Sweden's national Trafiklab archive is useful but key-gated. Monaco and Iceland fell below the calendar horizon; Luxembourg was stale and Taiwan returned 401. Mexico, Argentina, and Chile did not supply a current candidate with the required source and commercial-reuse chain. Community-produced African feeds remain partnership-gated. The registry these waves produced, 2,185 records across 46 country codes on 2026-07-25, is a reviewed sample and not a census or national-coverage claim. That figure records what the phase delivered and is not refreshed as the registry grows.
Much of the Global South's transit is informal — minibuses, matatus, colectivos, angkot — and its GTFS, where it exists, is produced by community-mapping and capacity-building projects, not by a government open-data portal. The largest sources are the Digital Transport for Africa commons (managed by WRI and the Agence Française de Développement, feeds typically under ODbL) and the Digital Matatus lineage of university and community collaborations. These are community-produced, and a Northern-built tool that assigns them a letter grade risks reading as a deficit judgement of the community rather than feedback to an accountable publisher. This project does not curate those feeds from a catalog.
The phase opens only when a named local steward — a regional program, a transit authority, or the community project that produced the data — owns the licensing, source verification, and consent, and agrees to operate or review the regional cohort. It is governed by the bounded Global South pilot in ADR 0028, and it adopts, by analogy, the CARE principles for data governance (collective benefit, authority to control, responsibility, ethics) alongside the project's existing no-shaming stance. Concretely:
- Opt-in, not opt-out. No public grade for a community-produced feed without the local steward's sign-off, and a standing removal right — the community's right to be off the scorecard, mirroring the humanitarian-mapping "right to invisibility."
- Credit the steward. The community or program that produced the data is named and attributed, never treated as an authorless source.
- Support, not sanction. For an informal feed, findings are framed as support and next steps, and the interface does not imply an accountable agency exists to act on them. An informal or demand-responsive feed is scored as its own kind of object, not as a defective fixed-route feed — the same reasoning that already treats a missing realtime feed as neutral rather than a zero.
- Benefit stays local. The most defensible posture is a cohort a local partner runs or co-owns, not a Northern-hosted ranking of Southern cities. The no-public-leaderboard rule holds especially here.
Gate: a named local steward accepts the licensing, source, identity, and consent responsibility; the reuse of each feed is established from a source the steward confirms; and the pilot stays bounded and labelled per ADR 0028. Absent that, the region stays uncurated by design, and the roadmap says so rather than leaving a silent gap. The Digital Matatus project is the precedent worth following: it validated data with the operators and commuters it described and planned an explicit handoff to a neutral local steward with a mandate to keep the data open.
These unblock more than one phase and are sequenced by first need.
- Large-feed tier (shipped). The standard ingestion caps (256 MiB download,
512 MiB single entry, 2 GiB total) correctly exclude very large aggregates,
but a small number of official national and metropolitan feeds legitimately
exceed them: their compressed download runs past 256 MiB, or a single table
such as
stop_times.txtexpands past 512 MiB. A feed now opts in per record withlarge_feed: trueafter a curator confirms it is a real published feed. The tier raises the size ceilings to a bounded larger level (512 MiB download, 2 GiB single entry, 4 GiB total), streams the download to disk with a bounded memory footprint (net.safe_download), and gives the validator an explicit heap ceiling (SCORECARD_LARGE_FEED_HEAP, default 6g). Every zip-bomb shape guard — the entry count, the compression-ratio check, and the central-directory-only inspection before the Java validator opens the bytes — stays exactly as strict; only the raw size ceilings move, and only for an opted-in feed. The first records on the tier are Israel's national feed, Melbourne (PTV), HSL Helsinki, Wiener Linien, and Carris Metropolitana — each a feed the standard caps rejected. A per-mode or per-region split of one oversized feed (distinct sub-records that each score) remains a possible future extension for feeds larger than the bounded tier; the tier unblocks every current target without it. Sydney (Transport for NSW) stays deferred on its registration wall, not on size. - Beta-gate generalization decision (decided,
ADR 0040). The European beta
gate is Europe-coded by design (a closed country set and thresholds chosen for
that market's addressable feeds). The question was whether the gate becomes a
parameterized contract (region, country set, thresholds, denominators)
evaluated by the same executable evidence, or whether other regions stay
ordinary reviewed coverage with no beta label. ADR 0040 is the decision of
record: the gate stays Europe-scoped and is not parameterized now; other
regions remain ordinary reviewed coverage with their own disclosed
denominators and no beta label until a named consumer asks for a specific
region and that region's cohort meets a stated, executable, region-appropriate
gate, exactly as a consumer named European coverage. The ADR records the
RegionGateseam so a second region's beta is an actionable follow-up with its own ADR when it is needed. - Alternative-catalog ingestion (first source shipped). For regions the
Mobility Database does not cover, a reviewed way to discover feeds from other
catalogs, national open-data portals, and regional aggregators, with the same
identity and reuse review. This is what actually raises non-Western coverage;
catalog curation alone cannot. The Transitland
Atlas is now wired in as a second discovery source:
scorecard sync --source transitland(orall) reads its keyless, CC-BY DMFR registry, emits the sameCatalogFeedshape the Mobility Database sync uses, and flows through the same proposer, deduplication, and curator-review-before-registry workflow. It never writes the registry directly and never guesses a feed's country — DMFR carries no ISO location, so a Transitland candidate surfaces with its location blank for the same review every feed gets. Key-gated feeds are flagged and skipped exactly as with the Mobility Database. National open-data portals remain a further source to add the same way. - A named-license vocabulary for non-SPDX government terms. Outside Europe, many official feeds carry a custom national open license with no SPDX identifier — Mexico's Términos de Libre Uso MX, Israel's government open terms, jurisdiction-ported Creative Commons (CC BY 2.5 AR, CC BY 3.0 CL) — and some official downloads carry no per-file license at all despite statutory openness. The reuse-evidence review handles these case by case today. A small controlled vocabulary of named government licenses, plus an explicit reviewer note for "openness established by statute, no per-file license," would keep those judgements auditable without auto-approving any of them.
- Per-region coverage denominators (shipped). The finder and exports
disclose the United-States-heavy global denominator, and each region now also
discloses its own. Filtering the directory to a country or subdivision states
that place's reviewed-cohort size beside the location controls ("19 reviewed
feed records in Italy"), read from the directory summary counts, framed as a
cohort size and never as a census. The per-country program pages carry the
same scope line, and the
/status/page states the European cohort's denominator with a per-country breakdown. A remaining refinement is to carry the per-region denominator into the CSV export headers as it is added to more activated regions. That refinement shipped in July 2026: every feature finder CSV row now repeats the selected coverage scope, its reviewed-record denominator, the matching-record count, and the shareable filter URL. Keeping the context in columns preserves a rectangular RFC 4180 file for downstream analysis.
This roadmap is "complete" when the defensible, openly licensed, in-cap official feeds in every region are reviewed and either curated or excluded with a recorded reason, the large-feed shard has a design, the beta-gate generalization has a decision, and every region that requires a local steward is documented as partnership-gated rather than silently missing. Completion is a state of honest, reviewed coverage with named gates — not a target feed count.
- Auto-curate community-mapped or informal-transit data without a local steward.
- Claim a regional cohort is a country or regional census.
- Relax the reuse, identity, license, or size gates to raise a flag count.
- Treat coverage growth as a success measure for the product.
- Ingest NeTEx-only or SIRI-only datasets; those remain a separate decision
(
global-expansion.md). - Add a regional beta label before a named consumer needs it and the cohort meets a stated, executable gate.