COMMUNITY INTEGRATION

Into MapSwipe.

This page is specifically about how reviewed GeoAI predictions reach MapSwipe volunteers — the fAIrSwipe-style flow, the project type it targets, the export package shape, and the permission path to get there. For the engine-level worker integration (QueueLLM lifecycle, DDaC contracts — not MapSwipe-specific), see Worker Integration.

STATUS: NOTHING BELOW IS BUILT OR SENT. Design-stage proposal, gated on maintainer agreement. Independent project, not affiliated with MapSwipe, HeiGIT, HOT, or the OSMF.

Overview Architecture Worker Integration MapSwipe Integration

THE FAIRSWIPE FLOW UPSTREAM · MAPSWIPE/FAIR

Everything past the export command is MapSwipe/fAIr territory: theirs to build, confirm, and approve. Never assumed deployed.

dac export run_01J... --format mapswipe-review-package   # explicit, human-approved
      → MapSwipe validation project (Validate Footprints or Assess Image — maintainers' call)
      → volunteers (families, classrooms, anyone) swipe: agree / disagree
      → consensus — preserving individual votes, timing, disagreement
      → OSM via existing MapSwipe/Tasking Manager channels
      → reviewed verdicts feed the evaluation kit and better models

The exported package is a stratified sample — high-confidence controls, boundary cases, low confidence, geographic outliers — never a dump of only the model's doubts. Volunteer experience, consensus, and OSM publishing belong to MapSwipe; GeoAI's contribution to volunteers is invisible queue quality.

WHICH PROJECT TYPE UPSTREAM · MAPSWIPE/FAIR

MapSwipe's six public project types, and how an external model's predictions map onto each. Which one actually hosts our export is maintainer question #1 below — this is our best guess, not a decision.

PROJECT TYPE	HUMAN TASK	FIT FOR OUR EXPORT
Find Features	Identify buildings, roads, waterways, or another visible feature	Labels presence/absence, narrows mapping areas
Compare Dates	Compare before/after imagery	Change and damage triage
Validate Footprints	Judge alignment or quality of mapped/AI footprints	Direct model-output quality control — likely fit for our export
Check Completeness	Find missing OSM features under an imagery overlay	Finds model/OSM omissions
Assess Image	Validate whether an image or proposed detection contains a feature	Flexible classification and model validation — likely fit for our export
View Streets	Inspect ground-level imagery	Adds attributes/local context unavailable overhead

THE EXPORT PACKAGE NOT STARTED PROPOSED SHAPE

mapswipe-review-packagedoesn't exist yet — its real shape is agreed with maintainers (agreement A3). This is what we think it should carry, derived from the geoai.prediction/v1alpha1 schema plus a stratification bucket:

{
  "item_id": "sha256:19c8...",                 // content-addressed source asset
  "task": { "id": "road-obstruction", "version": "1.0.0" },
  "observation": {
    "label": "possible_obstruction",
    "evidence": [{ "claim": "...", "geometry": { "type": "Polygon", "..." } }],
    "alternative_explanations": ["seasonal water", "image-registration mismatch"]
  },
  "scores": { "raw_model_score": 0.83, "calibrated_confidence": 0.71 },
  "stratum": "boundary_case",                  // control | boundary_case | low_confidence | outlier
  "receipt_digest": "sha256:4a12...",           // links back to the full geoai receipt
  "review_state": "approved"                    // this analyst's verdict before export
}

Every candidate a volunteer sees is one an analyst already approved for review — the package carries analyst verdicts forward, not raw unreviewed model output.

THE PERMISSION PATH NOT STARTED

No MapSwipe-facing adapter ships before the upstream boundary is agreed with maintainers. “Agreement” means four specific things:

A1 — UPSTREAM BOUNDARY

Where reusable split/evaluation code lives — fAIr, fAIr-utilities, or a MapSwipe analysis repo. No competing repository before this is settled.

Unlocks: M3 adapter (required)

A2 — ETHICAL PROJECT USE

Which live fAIrSwipe project may be used for a validation study, with a predeclared analysis plan reviewed by maintainers.

Unlocks: Any volunteer validation study

A3 — SIGNAL-PRESERVING EXPORTS

Exports that keep individual votes, timing, disagreement, and prediction confidence — plus the agreed shape of the inbound review package.

Unlocks: M3 adapter (required)

A4 — LICENSE CLARITY

An authoritative license per MapSwipe export type before anything is redistributed or publicly benchmarked.

Unlocks: Public benchmark claims

30 / 60 / 90 — NOT STARTED

DAYS 1–30 — JOIN AND REPRODUCE

Volunteer on MapSwipe across project types. Join MapSwipe Slack + monthly technical call, and HOT Slack. Send the intro, present the one-pager, reproduce a fAIr run.

DAYS 31–60 — PRODUCE EVIDENCE

Publish split/provenance design for review. Run the fAIr baseline across three held-out regions. Design a validation project with a predeclared analysis plan.

DAYS 61–90 — UPSTREAM VALUE

Smallest upstream PR (splits, STAC metadata, or evaluation export). Publish results including failures. Ask which second feature class matters most.

WE WILL NEVER

Push AI predictions directly into OSM · build a competing volunteer app or volunteer accounts · expose volunteers to low-confidence-only queues · make public benchmark claims from MapSwipe data before license questions are answered.

QUESTIONS FOR MAINTAINERS NOT STARTED ASKED: — NOT YET

A working subset of the full question list (repo: docs/outreach/maintainer-questions.md). Charlie asks these personally; nothing here has been sent.

MAPSWIPE

Which project type — Validate Footprints or Assess Image — is the right vehicle for validating an external model's predictions, and what input format does project creation require?

Why: Our export adapter targets exactly one ingestion path.

MAPSWIPE

Which exports preserve individual votes, timing, disagreement, and prediction confidence?

Why: Human-system metrics (review seconds, correction rates) need per-vote data.

MAPSWIPE

What do you require of an external prediction source — provenance fields, confidence semantics, license attestations — before a validation project can be created from it?

Why: Our receipts and prediction schema should carry exactly what MapSwipe needs.

MAPSWIPE

What is MapSwipe's age/safeguarding policy for young volunteers, and is a family/classroom mapathon framing welcome?

Why: The young-mapper journey and mapathon kit only proceed with a blessing here.

FAIR (HOT)

Which fAIrSwipe workflow is production today versus awaiting the broader roadmap integration?

Why: Roadmap dates are estimates; we must not design against undeployed functionality.

FAIR (HOT)

Where should reusable split/evaluation code live — fAIr, fAIr-utilities, or a MapSwipe analysis repository?

Why: ADR-0001's standing rule: no competing repository until the boundary is agreed.

← Back to GeoAI Architecture →Worker Integration →