# Two dormitories, and the limits of a good-looking model

September 6, 2026. A GPT-6 Astra working note, following release 39.

Tower Hall and Liggett Hall have replaced two plain campus blocks. Their open courtyards and unequal wings now follow the detailed campus map more closely. Selecting either building brings up a short introduction and a period image beside the current model.

## How the evidence changed the work

[The Sanborn atlas, M006](https://data.library.wustl.edu/record/108214?ln=en), sheet 1, calls the buildings Dormitory A and Dormitory B. The western building has a shallow courtyard. The eastern building has a much longer west wing, a higher middle and an area explicitly marked Foundation. That foundation is not a finished wing in this reconstruction.

[The February 1904 magazine, B021](https://archive.org/details/popularsciencemon64newy/page/359/mode/1up), supplies the design view captioned New Dormitories. Its caption does not individually name Tower Hall, so we use it as a comparison rather than proof of every finished detail. Two pages later, a photograph explicitly identifies Liggett Hall. Our reduced images P028 and P029 come from that same volume; they are not independent historical witnesses.

[The later campus history, B023](https://nara-media.s3.amazonaws.com/electronic-records/rg-079/NPS_MO/79003636.pdf), connects the older names to Umrath and Prince Hall. It also describes later construction that closed the first courtyard and a 1948 excavation at the second. Those later changes have been kept out of the fair model.

## What Astra helped accomplish

Astra connected the historical names, compared the page images, made the first courtyard geometry, updated the walking obstacles and prepared the visitor comparisons. Reusing the existing campus renderer kept this change small. It also created a problem worth checking: overlapping building pieces can put two facades in the same place. The central block was separated from its adjoining wings, and actual browser views were used to review the resulting shapes.

The walking checks ask a different question from the historical review: can a visitor enter the open court without passing through a wall? Both courtyard approaches pass that check. The neighboring path was also moved around the larger building footprint. That path remains provisional; a navigable route is not evidence of exactly where the fair's paving lay.

## A challenge to preserve for future models

**Unresolved: Tower Hall's central tower, adjoining smaller tower and passage.** The detailed map and later survey describe its central levels differently. The current model has a raised middle but omits the smaller tower and open entrance passage. The design illustration is stained and does not settle the finished construction.

This is an evidence and interpretation problem, not a demonstrated ceiling on Astra's ability. A useful next attempt should start with these same sources and the saved first courtyard model. Record whether a newer model reconciles the discrepancy more convincingly, produces better geometry, or makes an unsupported assumption. If a sharper photograph becomes available, distinguish the benefit of new evidence from the contribution of the model.

## What is still approximate

The building positions are fitted to the general fair map. The detailed sheet guides proportions, but the two maps are not perfectly aligned. Heights, repeated windows, roofs and ground levels remain estimates. Small gables, bays, chimneys, annexes and entrances need more work. Neither dormitory has an accessible interior yet.

No controlled timing, cost or model comparison was performed. The lesson is practical: AI can help move from a map to an explorable building quickly, but a plausible shape must keep its uncertainties attached.

The [source register](/sources) retains the catalogue links and research contacts for better copies. No institution has been contacted.