# Santiago pilot review protocol

This protocol separates fast AI-assisted discovery from the slower judgment that
OpenStreetMap-quality geometry requires.

## Review states

| State | Meaning | Next action |
|---|---|---|
| `accepted` | A real, current building is visible and the candidate outline matches OAM closely | Keep it, subject to approved import procedure and a fresh OSM conflict check |
| `retrace` | Real building, but candidate outline is merged, offset, or malformed | Draw clean geometry from OAM |
| `rejected` | Not a building, demolished, outside valid imagery, or too ambiguous | Do not map from this candidate |
| `uncertain` | Imagery or classification is inconclusive | Defer; do not upload |

“Accepted” is deliberately not called “approved” or “upload ready.” It means
that one person found the existing outline accurate against OAM. Community
approval, batch QA, and a fresh conflict check are still required.

## Pilot sampling

The deterministic 200-candidate batch is stratified so a good-looking consensus
subset cannot hide failures elsewhere:

1. AI + Microsoft consensus candidates test the fastest likely wins.
2. AI-only candidates measure model discovery precision.
3. Microsoft-only candidates measure DINOv3 omissions and older-footprint risk.

Existing OSM overlaps are removed before sampling.

IGAC and Google Open Buildings are counted in the source comparison but are
excluded entirely from shareable queue geometry, matching, scoring, ranking,
properties, and pilot selection.

## Geometry checklist

For every accepted or retrace candidate:

- confirm the roof is inside valid OAM pixels;
- distinguish one building from attached but structurally separate buildings;
- keep an existing candidate unchanged only when its complete shape aligns with
  OAM; otherwise correct or retrace it;
- square corners only when the visible structure supports it;
- avoid tracing roof overhang as a wall footprint when the difference is clear;
- check for an existing OSM object immediately before editing;
- use conservative tags such as `building=yes` unless a subtype is verifiable;
- record the OAM imagery/date in the changeset source and comment.

## Go/no-go metrics after 200 reviews

Report, by category:

- plausible-building precision;
- fraction requiring retracing rather than simple correction;
- false-positive reasons;
- missed/merged attached-building rate;
- OSM duplicate rate;
- median human review time.

Do not scale if imagery interpretation is inconsistent, false positives cluster
in a way the queue cannot suppress, or conflation with live OSM is unreliable.

## Community handoff

Before systematic editing, publish the area, imagery record and date, source
licenses, candidate-generation method, human-review procedure, proposed tags,
conflation method, changeset sizing, QA results, and rollback plan. Because the
proposal may retain generated geometry, treat it openly as an import, ask OSM
Colombia for feedback, and complete the import review process before uploading.