Translating Ancient Astronomy: A Practical Blueprint for AI-Ready Texts

Imagine finding a brilliant 14th-century astronomy argument tucked inside a manuscript scan. The diagrams are there. The star names are there. But the language barrier is total. You can see the structure of the thinking, yet you can’t read the sentences—so the work stays sealed inside its original alphabet.

That’s the gap Ars Astronomica is tackling: producing first English translations of rare Hebrew and Latin astronomy texts, aiming for readable modern English while still preserving the exact meaning, numbers, and technical terminology of the sources. The surprising part is that this isn’t only a translation problem. It’s an engineering problem involving page images, fragile historical typography, astronomical math, and diagram logic.

This post walks through a technical “mental model” of what such a system has to do—because once you understand the moving parts, historical astronomy translations start to feel less like magic and more like a carefully designed pipeline.

Why historical astronomy translation is different from “normal” translation

Regular translation usually assumes a clean text input: OCR (optical character recognition) turns pixels into characters, then a language model rewrites them.

Historical astronomy breaks that assumption in multiple ways:

1) The input is often page images, not text

Manuscripts and early printed books bring degraded ink, unusual fonts, and layout artifacts. Even when a scan looks crisp to the human eye, software often struggles.

OCR (optical character recognition) is the process of converting an image of text into machine-encoded characters. In this project style, OCR may be avoided entirely: the translation system can read directly from page images using a vision-capable model. The technical reason is straightforward: OCR errors in historical scripts (especially Hebrew) can silently corrupt meaning and numbers.

2) Astronomy texts are full of “structures,” not just sentences

When an astronomy author writes, they’re not only explaining. They’re also:

• listing tabular values (degrees/minutes/seconds)
• referencing earlier definitions and technical terms
• embedding geometric diagrams that carry information, not decoration

A translation that “sounds right” but loses a column value is worse than one that is slightly awkward stylistically.

3) The right translation is often a precision task

Astronomy and calendar computation frequently use exact fractional angles (historically recorded in sexagesimal—a base-60 number system).

• Sexagesimal means the number system is built on 60s, so an angle can be recorded as degrees + minutes/seconds, rather than decimal degrees.

So the pipeline has to reproduce numerical content cell-by-cell, not paraphrase it.

The pipeline idea: treat translation like a production workflow

Ars Astronomica describes an automated, AI-assisted pipeline that runs each text through a multi-stage workflow before final collation. The big technical insight is that a translation system for historical science needs more than one “generate English” call. It needs logging, reconciliation, and careful handling of uncertainty.

Here’s the blueprint in plain language.

Step 1: Batch the scanned PDF into manageable chunks

Instead of feeding an entire book to the model at once, the workflow splits the scan into batches. The purpose is practical:

• reduces failure blast radius (a batch can be retried)
• keeps progress measurable
• makes later reconciliation easier

In the showcased pipeline, the orchestrator logs progress after each batch and enforces pacing to avoid session-window issues during long runs.

Step 2: Translate from vision without OCR preprocessing

In the GitHub pipeline showcase, the approach emphasizes “native vision”: Claude reads page images directly, without OCR preprocessing or fine-tuning.

Technically, this matters because Hebrew and Latin typography vary widely:

• Hebrew square script (common in printed texts) may degrade via ink erosion and compositing mistakes.
• Manuscript hands can be readable but approximate, turning translation into a “best working draft” phase.
• Certain Hebrew letter confusions are predictable in real runs.

The pipeline’s prompts can include guidance for known confusion patterns. That’s not a gimmick—it’s the difference between a system that works sometimes and one that produces reliable manuscript-grade output.

Step 3: Preserve technical meaning with a glossary discipline

A subtle but essential rule in historical science translation is: terms must remain consistent.

This is where a glossary discipline helps.

• When a key Latin or Hebrew term of art appears for the first time, the pipeline can attach a modern English equivalent and then keep using that settled rendering.
• Afterward, the system doesn’t “re-decide” terminology on every sentence.

This is also why historical astronomy translations often include inline identifications for named stars or specialized vocabulary.

Step 4: Reproduce tables and math with cell-by-cell fidelity

Calendar computation and spherical astronomy are packed with tabular data.

A workable translation system needs to treat tables like data, not prose. That means:

• degrees/minutes/seconds must be carried over exactly
• multi-page tables must remain aligned
• empty or misread cells must be flagged

If the original shows a value in the middle of a table, the translation needs the same value in the same spot. Not “approximately,” not “similar to.” Precision is the entire point.

A tiny example of the mindset

Suppose the source lists an angle as degrees, then minutes, then seconds. The pipeline can preserve the formatting logic and reproduce the numbers so that later readers can verify the arithmetic against the original.

Even a correct English sentence like “the angle is about…” can be harmful here if it replaces exact values.

Step 5: Handle diagrams by extracting structure, not just captions

The system described by Ars Astronomica notes that public PDFs are currently text-only, while geometric figures are replaced with structured descriptions that identify labels, arcs, points, and relationships.

That’s a nontrivial engineering choice.

A diagram in historical astronomy might contain:

• concentric-sphere charts (cosmological geometry)
• instrument schematics (how a device is used)
• volvelles (rotating circular overlays with many peripheral labels)

Instead of leaving diagrams as “missing images,” the pipeline describes them in a structured way so the translation stays self-contained for readers.

This is also why multi-pass verification matters: small labels around a rim can be easy to miss.

Step 6: Reconciliation—turn translation uncertainty into an editorial queue

Here’s the part that makes the pipeline feel like real software engineering rather than a one-shot generation.

After translation, the system runs a multi-stage reconciliation before final output:

1. In-flight logging: anomalies are captured as candidates for rerun.
2. Categorized reconciliation report: flagged issues are sorted into types.
3. Human selects rerun set for cases where rerunning is likely to help.
4. Reruns execute under verification framing: prompts ask the model to check against the source rather than “apply a fix.”
5. Collation: batches are merged into a complete manuscript-style output, often with discrepancy reports for editorial follow-up.

This categorization is critical. Not every problem should be “rerun.” Some are compositor errors in the source; others are genuine author/source anomalies.

Why “check, don’t apply” matters

When a model thinks it knows the correct output, it may overconfidently rewrite an error.

By framing reruns as verification—“report what the source actually shows”—the pipeline reduces confident corruption.

Marking uncertainty without erasing it

A mature translation workflow needs a way to represent:

• corrections to clear compositor errors
• uncertain readings caused by damage or ambiguous print

The approach described uses bracketed notations to preserve best readings while flagging uncertainty for editorial review.

From a technical perspective, this is huge: it prevents the system from silently polishing away the uncertainty that historians actually care about.

What you get at the end: readable English plus traceability

The end goal isn’t a “modernized” paraphrase. It’s fluent, readable English that still respects the source.

That includes:

• faithful negations (a flipped “not” is catastrophic)
• faithful numbers (sexagesimal fractions stay exact)
• preserved technical examples and analogies
• precise vocabulary mapping

And because the pipeline explicitly supports versioning and regeneration, updates can improve earlier output when the underlying translation process changes.

The quiet lesson: translation is a quality-control system

Historical astronomy translation feels like scholarship, but the machinery underneath resembles a manufacturing line:

• controlled batch processing
• automated verification loops
• structured error tracking
• careful collation
• human-in-the-loop editorial review

So when English translation appears for a text that previously sat locked behind Hebrew or Latin typography, it’s not only a linguistic event. It’s also a data-engineering achievement.

One question that readers may naturally ask is: can AI really preserve the exact math in these old works? In practice, the answer depends less on “AI capability” and more on whether the pipeline treats tables, diagrams, and uncertainty as first-class citizens.

Conclusion

Translating historical Hebrew and Latin astronomy into readable modern English is hard because the text isn’t only language—it’s computation, geometry, and fragile historical printing. A system like the one behind Ars Astronomica succeeds by treating translation as an end-to-end workflow: image-based reading, glossary consistency, cell-precise math preservation, structured diagram descriptions, and reconciliation that turns uncertainties into an editorial queue.

Once you view it that way, the work feels less like a leap of faith and more like a practical engineering method for bringing historical astronomy into shared language—without losing what makes the original text intellectually precise.