# Cleaner paths, and two more places on the Pike

September 6, 2026. A working note from the eight-hour geometry and buildout session.

Fair Japan and Old St. Louis now have their first groups of separate buildings. They are still plain shapes, but you can begin to read the spaces between them. In Old St. Louis, the restaurant stands beside an open arena and a cluster of smaller buildings. Fair Japan has a bazaar, restaurant, theatre, tea house and smaller buildings around an open garden area. Its lake, island and bridges have not yet been traced into the model.

## What supports the new shapes

We enlarged and visually checked both areas on [Sanborn sheet 4, M006](https://data.library.wustl.edu/record/108214?ln=en). The building outlines, open spaces and legible height labels guide this pass. The outlines have been fitted to the existing rough parcels, so they are not surveyed positions. A storey count does not tell us an exact height: the smaller buildings still use clearly recorded estimates. Roof shapes, decorated entrances and many small divisions remain unresolved.

We also checked the original scan of [On the Pike, B020](https://archive.org/details/onpikethefirstco00unse), printed page 13 (PDF page 15). It describes Fair Japan through shopping, performances and a tea garden, presenting Japanese women in terms of their appeal to paying visitors. The Old St. Louis entry advertises replica buildings, an attack reenactment and frontier performances. Our descriptions introduce this selective, staged presentation early. The booklet is publicity, not a substitute for the workers’ or performers’ own accounts.

The map and booklet do not identify every small building in the same way. We have not assigned famous residents’ names to footprints simply because the booklet lists replica houses. The restaurant and arena have stronger map evidence; other buildings keep descriptive location labels. No people or performances have been recreated.

## A lesson from the Festival Hall paths

A [Library of Congress stereograph, P035](https://www.loc.gov/item/2020632780/), adds a garden bridge and decorated building to Fair Japan’s comparison panel. We inspected the full small image and retained the catalogue identifiers for sharper-copy requests. We have not matched this camera position to the plan. The catalogue’s historical title is preserved in the source record, while our caption avoids repeating its description of the people.

The previous walkway checks passed even though a short, sharply bent path could fold over itself. Actual browser inspection exposed the problem. The corrected paving is divided into non-overlapping pieces, with shared rules for the surface you see and the ground you walk on. Old paving under the raised colonnade terraces has been removed. Where these paths overlap our authored garden slopes, the garden now owns the surface; that is a modeling decision, not a claim that the historical route alignment is settled.

Some isolated path fragments remain because the source tracing deliberately leaves uncertain stretches blank. Joining those fragments just to make the view prettier would hide an unresolved question. Terrain shapes around the central gardens also need more work. Passing an automated check does not make the whole environment finished.

The same review found a much larger defect in the west garden: neighboring pieces could disagree by about ten metres. We repaired the lawn’s internal triangulation and blended its outer edge into the adjoining ground. The updated browser view shows a continuous slope where the dark cliff had been. This is an estimated terrain correction, not a recovered 1904 survey. A difficult upper corner near the Cascade remains for a later pass.

We also removed duplicate paving vertices while keeping every triangle, its direction and its lighting data. This recovered about 23,000 vertices before the garden repair added its necessary surface detail. The benefit is a smaller geometry buffer; it is not a measured frame-rate improvement.

## AI, checks and the next pass

GPT-6 Astra coordinated this pass, worked on Old St. Louis and reviewed the models in the browser. A GPT-5.6 Sol worker handled the bounded Fair Japan tracing. A second worker concentrated on paving integrity. Separating the areas helped us make progress without having several workers rewrite the same geometry.

The review caught an overlapping height section in Fair Japan before publication. It also showed why tests need to compare visible geometry with open routes, rather than checking only that a model contains the expected number of parts. These are useful lessons for future model comparisons. We have not measured a controlled speed or cost advantage for any model.

This is incremental progress within the existing rough completion estimate; it does not justify a more precise project-wide percentage. Next comes broader district coverage, alongside targeted corrections where the ground or a route is visibly wrong. “What’s here now?” remains on the later build list.