Methodology

Atlas node

Flint Atlas

Federation

seed

Civic objects

8

Catalog nodes

5

Layers

5

Scene manifests

1

1. 01Start with public records and fixture-safe summaries that can be shown without exposing private uploads.
2. 02Place each record on a city object such as a parcel, block, street, park, building, or neighborhood.
3. 03Separate what a record says directly from what nearby records only suggest.
4. 04Show residents where the atlas is confident, where it is incomplete, and how to send corrections.
5. 05Keep public contributions separate from publication until receipts and review workflows are ready.

The atlas uses two different confidence numbers, and they answer different questions. The first is about the building's type today. The second is about how sure the atlas is about each part of a historical reconstruction. Both are visible here, in the dossier, and inside the atelier; they are not shown as floating numbers on the map.

• Building typology classifier21,182 buildings in the public package

  0.97 average confidence : The classifier is confident about most residential and commercial buildings. It is much less certain about civic, industrial, and mixed-use buildings, where the public training pool is thin.

• Reconstruction per-part confidence5 Lost Flint reconstructions in the current fixture

  0.60 to 0.95 per part : Each reconstruction stores a separate confidence for mass, facade, roof, and ground-floor use. The atelier and dossier show these individually, so a resident can see exactly which part of the building is well-supported.

Each historical reconstruction stores a separate confidence number for the building's mass, facade, roof, and ground-floor use. The atelier and the dossier panel show these individually, so a resident can tell at a glance which parts are well-supported and which are the model filling a gap.

• 0.90 to 1.00
  High confidence

  Multiple sources agree, or one strong source covers this part directly. The atelier renders the part with full porcelain detail; the atlas hover stays clean.

• 0.60 to 0.90
  Partial confidence

  Some evidence supports this part, but it is incomplete or weakly cited. The atelier renders the part with a softer ghost finish; the dossier shows which sources are partial.

• Below 0.60
  Contested

  The model is filling a gap, or sources disagree. The atelier shows the part faintly with a Contested label; the dossier opens to the disagreement so a resident can challenge it.

1. Stage 01Gather evidence. Pull every record anchored to this parcel and the block around it: maps, photos, directories, assessor rows, standing-building controls.
2. Stage 02Read direct details. Extract footprint, height clues, material, roof shape, storefront use, address, and year from what each record says directly.
3. Stage 03Build the block context. Map the building's neighbors and how they relate: shared walls, shared setbacks, same block, adjacent street.
4. Stage 04Bring in spacetime memory. Use the surrounding block's history to give the building a position in the city's space and time, not a generic city template.
5. Stage 05Predict missing parts from context. Ask the building model to suggest only the pieces still missing after direct evidence and block memory.
6. Stage 06Merge direct and predicted. Combine what records say directly with what the model suggests, and surface every disagreement instead of quietly picking one.
7. Stage 07Queue a renderable asset. Generate the building scene asset only after the merged claim package is complete, with status and review notes attached.
8. Stage 08Run the full pipeline. Execute all seven stages as a single supervised unit, one stage at a time. If any stage fails or produces contested output, the pipeline stops and surfaces the problem rather than continuing with degraded data.

• The public Reconstruction Engine route still uses the Carriage Town fixture while the live GraphQL reconstruction resolver is connected.
• The current fixture has five Lost Flint reconstructions, and none yet meet the three-source pilot bar from the real-reconstruction plan.
• Model scores are support signals. They do not become final decisions about a building, block, or resident contribution.
• Photo, city-directory, and HABS-style ingest lanes need more work before the first production training run.
• Confidence numbers appear here, in the dossier panel, and inside the atelier. They do not appear in atlas hover tooltips. That is intentional: a number floating over a building reads as a verdict, which it is not.