Skip to main content
Back to timeline
arXivSource publication:

Babylonian Astronomical Diaries no longer dated tablet by tablet: a language model reads the broken text and astronomical tools rank 701 candidate years

Synopsis

The Babylonian Astronomical Diaries record the sky for nearly six centuries, and many broken tablets have lost their date formula, which specialists previously dated one tablet at a time by comparing the surviving observations with computed phenomena. The new system turns observations into fixed-field clauses with a rule-based parser, has a language model check them against the text and add missed ones, and tests the clauses in all 701 years from 626 BCE to AD 75. On 120 dated tablets with decisive clauses, the top-ranked group contains the published year in 96% of cases; for 86 of them the group is a single year, the published year in 84.

AI-generated editorial illustration: Automated, auditable dating of the Babylonian Astronomical Diaries with language models and astronomical tools

Interpretation

Dating a broken tablet moves from tablet-by-tablet manual comparison of the sky to one automated pipeline that applies the same rules to every tablet and can be rerun. Sachs and Hunger had compared surviving observations with computed planetary phenomena and eclipse catalogs one tablet at a time in the ADART edition, with the reasoning given in prose comments, so it was hard to see how much of each published date rested on astronomy. On a benchmark of 179 dated records, the top-ranked group contains the published year for 115 of the 120 records with decisive clauses, 96%, and the strict candidate set contains it for 119, 99%, with a median strict-set size of 15 years.

The system reads each tablet's astronomical reports into clauses and then uses astronomical computation to decide which years are compatible with them. The rule-based parser matches common wordings only, while the language model reviews clauses and adds observations the parser missed, with candidate years and scores hidden from the model. The parser finds 3,193 clauses in 316 of the 417 dated ADART I to III records, and review plus search raise this to 4,010 clauses from 321 records; where the tools can test them, reviewed clauses agree with the published years in 93.6% of cases and parser clauses in 93.8%.

Years are narrowed by five clocks together: Mars, Jupiter, Saturn, lunar eclipses visible at Babylon and the solar year. A single planet in a zodiacal sign in a given month agrees with many years, so one planet rarely fixes a year. Five decisive clauses of record X301492 agree with 138, 116, 54, 50 and 49 years, and their running intersection falls to one year, the published 150 BCE, where all three clocks agree and decisive bits total 16.3.

Perspective

The system runs on the 417 dated records of ADART I to III and the undated entries of ADART IV, with a candidate calendar of 701 Babylonian years from 626 BCE to AD 75; across all 417 dated records the commitment rule gives a single year for 10, all of them inside the benchmark.

The system automates only the astronomical part of the evidence; Sachs and Hunger also used surviving year formulas, royal names and the form of the script. For the 56 undated ADART IV entries, the strict sets have a median of 102 years, and most entries keep a single decisive clause that stands before any month heading and is tested over a whole year.