TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 16—18 · FACTUAL · EVOLVING
Machines reading historic drawings and inscriptions
Use the machine to make the document readable, and check every number yourself.
When to use
When the record of a building is old and hard to read: a survey plate from the 1860s, a hand-lettered plan, an inscription, a gazetteer page in a script you cannot read fluently. Use the machine to make the document readable. Decide what it says yourself.
The method
Before the document goes anywhere: use material you own, licensed for this use, public-domain, or cleared by your institution — a historic drawing is an artistic work in Indian copyright law, and "educational use" is not a blanket over uploads to a commercial service (FOUR RISKS is the map's gate). Then scan or photograph the document flat, in even light, at the highest resolution you have, with a ruler or the sheet's own scale bar in the frame. Give the image to a multimodal model with a narrow question — "transcribe the annotations on this plan exactly as written, and mark anything you are unsure of with [?]" — because a narrow question produces a checkable answer and "tell me about this drawing" produces a paragraph. Ask it to name the drawing convention before it reads the numbers: plan, elevation or section; feet or metres; hatching as masonry or as earth. A model that has named the convention misreads fewer dimensions (THE CONVENTION on The Brief is the same move, used in the other direction). Then check every number against the drawing yourself — every one. A transcribed dimension you have not checked is still a guess. For handwriting at scale — a bundle of letters, a register — purpose-built tools trained on historic scripts exist; they are still checked page by page.
Watch for this
The confident dimension. A faded 18'-6" can come back as 16'-6" with no flag; a survey plate in feet can come back in metres without being asked. The model produces what the annotation most likely says. "Most likely" is a statistical guess, not a reading. The rule from WHY IT MAKES THINGS UP holds here: confident is not correct.
Try it
Find a scanned historic plan with hand-written dimensions — the Internet Archive's texts hold many public-domain survey volumes. Transcribe ten dimensions by eye first and write them down. Then ask a multimodal model to transcribe the same ten. Compare. Check each disagreement against the scan at full zoom — sometimes the machine is right and you were wrong, and that is worth knowing too. Then test the script. Take a survey plate or a revenue map annotated in Kannada — a district gazetteer, a village map, or a sheet beside the Dharwar district plates — and an English-annotated sheet of the same kind, and ask for the same ten-item transcription of each. Score both. Then write one Kannada annotation out in Romanised Kannada and ask the model to read that too. Three scores — English, Kannada script, Romanised Kannada — tell you which of your region's documents this tool can actually open.
Prove it
Show one historic drawing with its machine transcription, every dimension marked as checked, and the one the model got wrong, with the zoomed crop that proves it.
How it works
A multimodal model turns the scanned page into image patches and reads them in the same stream as your words — ONE STREAM on this map is the mechanism. It has seen many drawings, so it knows what a plan's annotations usually say, which is exactly why it can fill a faded numeral with the likely one. Purpose-built handwritten-text recognition platforms such as Transkribus take a different route: a model trained on a particular hand or script for a particular archive, with accuracy measured per page and corrected by the people who use it. The Lab's own teaching material includes a study corpus built around the 1866 photographic survey of the Dharwar district and Mysore — Dharwar being the pre-1947 district of the Bombay Presidency, not the town. Its plate captions, the accompanying text and the gazetteer pages beside them are exactly the kind of document this method opens. The first thing a careful reader learns from them is that the plates show Halebidu, Belur, Gadag and Vijayanagara, and not one frame of Dharwad town. The machine will read the caption "Dharwar" and offer you the town. Read properly, the archive says district. EVOLVING because transcription of faded and non-Latin scripts is improving quickly and unevenly. The check-every-number rule does not change.
What this idea builds on
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.