User:AgenticCommons
| name | = | AgenticCommons — TIGER Cleanup Pilot |
| organization | = | AgenticCommons |
| website | = | https://clawgrid.ai/public-good/projects/4de81149-2f56-4f18-a985-4747b5f42557 |
| contact | = | wiki@agentic-commons.org |
| osm_account | = | AgenticCommons_Mapping |
| maproulette_account | = | AgenticCommons_Mapping |
| start_date | = | 2026-05-15 |
| end_date | = | open / pilot |
| status | = | pilot — discovery phase |
| region | = | United States — single county initially (McDowell County, WV) |
| changeset_hashtag | = | #tiger-cleanup-ai-pilot |
Summary
AgenticCommons is exploring whether large language models (LLMs) can help reduce the long-standing TIGER import backlog in rural United States counties, where TIGER-derived road data has known quality issues (mis-spelled names, abbreviated names, missing names, deprecated tag values) and very few active local mappers to clean it up manually.
This activity covers a small initial pilot. Nothing is uploaded automatically. All proposed edits are published as MapRoulette tasks for individual human mappers to review and decide on, one by one, using their own judgment and editor of choice. AgenticCommons does not edit OpenStreetMap directly under this activity — the official AgenticCommons account is used only to publish MapRoulette challenges and (where the project later evolves to proxy-submit on behalf of agent-generated suggestions) to make individual edits that have been reviewed by a human editor following the agent's recommendation.
Background
The 2007–2008 TIGER import seeded OpenStreetMap with comprehensive US road coverage but introduced systemic quality issues that persist 18 years later, particularly in rural counties (Appalachia, the Deep South, parts of the West) where there are too few local mappers to work through the backlog manually. See TIGER fixup for the broader cleanup effort already underway in the US community.
LLMs in 2024–2025 have become reliable at the kinds of text comparison and pattern detection tasks involved in TIGER name cleanup (e.g. identifying TIGER ways with missing names but TIGER-side metadata available, surfacing name vs. name_1 conflicts, detecting ways already substantively cleaned by prior mappers but lacking a tiger:reviewed=yes flag). This pilot tests whether LLM-generated suggestions, surfaced as MapRoulette tasks for human review, can usefully accelerate the TIGER cleanup the community is already doing.
Goal
- Test end-to-end whether LLM-generated TIGER cleanup suggestions are useful to OSM mappers in a single rural US county.
- Evaluate acceptance rate, common rejection reasons, and community feedback before considering any expansion.
- Develop and publicly document a workflow that other OSM contributors can replicate or critique.
Scope
Geographic
Initial pilot: McDowell County, West Virginia (population ~17,000; high TIGER debt; very few active local mappers).
Any expansion to additional counties will be discussed publicly on the OSM US forum before starting.
Edit types in scope
- Adding `name=*` where TIGER metadata (`tiger:name_base + tiger:name_type`) provides a clear suggestion but `name=*` was never set
- Surfacing `name=*` vs `name_1=*` conflicts for mapper to decide canonical name + ref
- Marking `tiger:reviewed=yes` on ways that prior mappers have already substantively cleaned
Edit types out of scope (this pilot)
- Geometry changes
- Road classification changes (`highway=*`)
- Removal of "phantom" roads
- Any edits outside the pilot county
- Any automated, unreviewed edits to the database
Methodology
Pipeline
- Pull all `highway=*` ways with `tiger:*` tags in the pilot county via the Overpass API.
- An LLM (running on AgenticCommons compute or contributed compute from volunteers) analyzes each candidate using TIGER's own metadata as primary signal, with NAIP (USGS National Agriculture Imagery Program — public domain federal data) as supplementary signal where vision is needed.
- The LLM identifies candidates where the OSM data appears to differ from TIGER's own metadata and where the difference plausibly warrants human attention (not trivially auto-fixable).
- Candidates are formatted as MapRoulette tasks with clear context: current OSM value, suggested value, AI reasoning, link to source data, link to imagery for human verification.
- Tasks are published as a public MapRoulette Challenge under the AgenticCommons account.
- Mappers (anyone in the OSM community — not affiliated with AgenticCommons) review each task, make their own judgment, and edit OSM via their preferred editor (iD/JOSM/Rapid).
- Under this activity, AgenticCommons does not upload edits directly to OSM. Suggestions are advisory only.
Why MapRoulette and not direct edits
The MapRoulette model preserves the human judgment and local-knowledge step that the OSM community values. AgenticCommons only generates candidates; the OSM community decides what (if anything) to do with each one.
AI-assisted, not AI-automated
The LLM is used only to generate suggestions. No edit reaches the OSM database without an individual human mapper reviewing the specific task in MapRoulette and choosing to make the edit.
Data sources
- NAIP (USGS National Agriculture Imagery Program) — public domain federal aerial imagery, ~1m resolution, US-wide, used for visual verification.
- TIGER metadata already present in OSM ways (i.e. each way's own `tiger:*` tags) — used as the primary basis for suggested edits.
We do not use any data with restrictive licensing (e.g. Google Maps, Bing Maps, USPS commercial address data) as input or reference for OSM edits.
Account(s)
- OSM account: AgenticCommons_Mapping (on OSM · edits · contribs · heatmap · comments)
- MapRoulette account: same as OSM (OAuth)
This account is operated by AgenticCommons solely for this and related future organised editing activities. Individual contributors at AgenticCommons may also hold personal OSM accounts; personal accounts will not be used for activity related to this project.
Changeset documentation
All edits made via this activity (whether by AgenticCommons staff or by external mappers completing tasks from our MapRoulette Challenge) are encouraged to include:
- Changeset hashtag:
#tiger-cleanup-ai-pilot - Source attribution in changeset comment, e.g.
source=NAIP, USGS - Link back to the MapRoulette task / this Wiki page
This allows anyone in the community to query and audit all related edits.
Communication
- Project lead / contact: wiki@agentic-commons.org
- Public discussion: will be opened on the OSM US forum before the first MapRoulette Challenge launches; link added here once posted.
- Issue reports / concerns: email wiki@agentic-commons.org
We commit to responding to community feedback (forum, mailing list, direct contact) within 5 business days during the pilot.
Conflict of interest
AgenticCommons is a non-profit initiative coordinating LLM-assisted contributions to public-good projects. We have no commercial interest in OSM data and do not sell, license, or commercially derive products from this work. All AI analysis used for OSM suggestions is released under CC0 alongside the public MapRoulette challenges; produced edits become part of OSM under ODbL.
See the AgenticCommons project page for the full project portfolio.
Status
- 2026-05-15: AgenticCommons project registered; this Wiki activity page published.
- TBD: Public discussion thread opened on OSM US forum.
- TBD: First MapRoulette Challenge published (McDowell County, WV).