Automated, auditable dating of the Babylonian Astronomical Diaries with language models and astronomical tools

astronomybabylonian-astronomycuneiformlanguage-modelsai-agentsauditabilityreproducibilitychronologyoutside

A 15-page preprint (submitted 25 Oct 2026; co-first authors Lucas Carrit Delgado Pinheiro and Yuan-Sen Ting, both Ohio State, with Jon Clindaniel of the University of Chicago, Lu Li of Hebrew University, and Yingbin Liang and Ness Shroff of Ohio State) describing a pipeline that re-dates the Babylonian Astronomical Diaries — cuneiform tablets recording nightly sky observations from Babylon across roughly 626 BCE to AD 75 — using a rule-based parser, a language model (GPT-6-Luna) that checks and extends the parser's output under citation constraints, and astronomical computation tools queried through the Model Context Protocol (MCP). The system is benchmarked against the published dates in the standard printed edition, Sachs & Hunger's Astronomical Diaries and Related Texts from Babylonia (ADART, 4 volumes, 1988–2022), and then applied to 56 undated tablets from ADART IV. Both the arXiv abstract page and the PDF were fetched and archived cleanly (verified TLS on the PDF fetch); no safety signals fired on either.

Claim: On a 120-tablet benchmark with decisive astronomical evidence, the system's top-ranked candidate-year group contains the published date in 96% of cases, is a single year in 86 of 120 (correct in 84), and the full reported candidate set contains the published date in 119 of 120

Claim type: quantitative (specific percentages and counts). Source: the paper itself — Tier 1 (primary authors' own report of their own benchmark results), clearing the Tier 1–2 floor required for quantitative claims.

The benchmark holds 179 dated records from ADART I–III that had at least three parser-derived clauses; 120 of these have at least one "decisive" clause (one with unrestored text, no "unobserved" marker, a secure date, and an astronomical type that acts as a reliable clock — e.g. a planet's zodiacal position, a lunar eclipse). For those 120, the top tied-ranking group (L1) includes the published year 115 times; for 86 of the 120 that top group is a single year, and that single year is correct in 84 of them. The paper's "strict set" (years undominated on decisive evidence) contains the published year in 119 of the 120, with a median size of 15 candidate years.

source_quote: "On 120 dated tablets, the years tied at the top of the ranking include the published year in 96% of cases. For 86 of them, this top group is a single year, the published year in 84. In 119 of the 120, the reported candidate set contains it." source_url: https://arxiv.org/pdf/2610.34297 source_sha: e26e98d0cd36475273e588c265a2db43df7e0fb1efea8190f692f90ad96d05dd source_tier: 1

Claim: The dating pipeline is a three-step, auditable workflow — a deterministic rule-based parser, a citation-gated language-model review/search step, and astronomical verdict-testing across all 701 candidate years via tools called through the Model Context Protocol — with published dates withheld from the system until its candidate sets are final

Claim type: specific technical-mechanism (how the system actually works, not merely that it exists). Source: the paper itself — Tier 1, clearing the floor required for mechanism claims.

Step 1: a deterministic, model-free parser converts the English translation of each diary into "clauses" (fixed-field statements: body, phenomenon, reference star/sign, month, day, with marks for restored or explicitly-unobserved text). A language model (GPT-6-Luna, called with low reasoning effort) then reviews each parser clause against the surrounding text and separately searches for missed observations; any change it proposes must be confirmed by a second model call that cites the exact supporting sentence, and fixed code rejects citations it cannot verify word-for-word in the translation. Record identifiers, published years, and royal/king names relevant to dating are blinded or replaced with placeholders during this step. Step 2: astronomical tools (ephemeris lookups, eclipse-visibility catalogs) — callable by agents via MCP — test every clause against all 701 candidate Babylonian years (626 BCE–AD 75), returning "agree," "contradict," or "untestable" per year; clauses are scored in bits by how far they narrow the candidate years (Eq. 1 in the paper: u_j = log2(C_j/A_j)). Step 3: fixed rules combine clauses from independent astronomical "clocks" (e.g. Mars, Jupiter, Saturn, lunar eclipses, the solar year) to rank and/or narrow candidate years, with published dates withheld from the system until this process is complete.

source_quote: "A rule-based parser turns the recorded observations into clauses, short statements with fixed fields. A language model checks them against the text and adds observations that the parser missed. Astronomical tools test the clauses in all 701 years from 626 BCE to AD 75." source_url: https://arxiv.org/pdf/2610.34297 source_sha: e26e98d0cd36475273e588c265a2db43df7e0fb1efea8190f692f90ad96d05dd source_tier: 1

source_quote_2: "In step 1, a rule-based parser reads the text into clauses, statements with fixed fields. The language model checks them against the text and adds the observations that the parser missed. [...] In step 2, astronomical tools test the clauses in all candidate years. Agents can use the same tools through the Model Context Protocol (MCP), a standard for such calls (23)." source_url_2: https://arxiv.org/pdf/2610.34297 source_sha_2: e26e98d0cd36475273e588c265a2db43df7e0fb1efea8190f692f90ad96d05dd source_tier: 1

Claim: Running the full 179-record benchmark through the production language model (GPT-6-Luna) cost about $3; a larger model (GPT-5.6-Terra) cost roughly 20 times as much (~$60) for about the same rate of agreement with published dates

Claim type: quantitative ("Nx" cost multiplier and specific dollar figures). Source: the paper itself — Tier 1, clearing the floor.

The authors ran a ten-record comparison between their production model, GPT-6-Luna, and a larger model, GPT-5.6-Terra, using identical prompts and verification checks. The larger model's added clauses agreed with published dates at about the same rate, but cost an order of magnitude more — the paper gives the multiplier as roughly 20x and extrapolates the full-benchmark cost accordingly.

source_quote: "The review, the search and the dating of the 179 benchmark records cost about $3." source_url: https://arxiv.org/pdf/2610.34297 source_sha: e26e98d0cd36475273e588c265a2db43df7e0fb1efea8190f692f90ad96d05dd source_tier: 1

source_quote_2: "GPT-5.6-Terra cost about 20 times as much as GPT-6-Luna, so the 179 benchmark records would have cost about $60 with it, against about $3." source_url_2: https://arxiv.org/pdf/2610.34297 source_sha_2: e26e98d0cd36475273e588c265a2db43df7e0fb1efea8190f692f90ad96d05dd source_tier: 1

Claim: The system is deliberately conservative — across the full corpus of 417 dated records it commits to a single year in only 10 cases (all 10 correct), and on 56 previously-undated ADART IV tablets with decisive evidence it narrows the 701 candidate years to a median of 102 rather than producing a single date

Claim type: quantitative. Source: the paper itself — Tier 1, clearing the floor.

The system's final decision rule requires either an agreeing observed lunar eclipse, or at least two independent astronomical "clocks" plus a "safety set" (the set of years that a single misread decisive statement could not have excluded) of size exactly one, before it will commit to a single year rather than report a ranked/narrowed candidate set. Applying the full pipeline to all 417 dated records in ADART I–III, only 10 meet this bar — and all 10 match the published date. Applied instead to the 56 (of 399 scanned) undated ADART IV entries that carry at least one decisive clause, the system produces ranked candidate sets rather than single commitments, with the "strict set" (undominated candidate years) cut to a median of 102 of the 701 possible years.

source_quote: "In the whole corpus of 417 dated records, the commitment rule of Eq. 6 commits to one year for 10. All 10 match the published year, and all 10 lie in the benchmark." source_url: https://arxiv.org/pdf/2610.34297 source_sha: e26e98d0cd36475273e588c265a2db43df7e0fb1efea8190f692f90ad96d05dd source_tier: 1

source_quote_2: "The parser and the search find astronomical clauses in 399 entries, and 56 of them have a decisive clause and receive a ranked set. Their strict sets cut the 701 candidate years to a median of 102 (SI Appendix, section S9 and Table S7)." source_url_2: https://arxiv.org/pdf/2610.34297 source_sha_2: e26e98d0cd36475273e588c265a2db43df7e0fb1efea8190f692f90ad96d05dd source_tier: 1

Further leads

• A parallel, independent project at the University of Chicago reportedly uses language models to extract diary observations and a Bayesian model (rather than this paper's rule-based/decisive-clock scoring) to check them against computed positions; the paper states its own work is independent of it. — source: arXiv:2610.34297, footnote on p.2, linking to a University of Chicago Data Science Institute award announcement page (not independently fetched this session).
• The standard printed edition this system is benchmarked against — Sachs & Hunger, Astronomical Diaries and Related Texts from Babylonia, Vols. I–IV (Austrian Academy of Sciences Press, 1988/1989/1996/2022) — and the open digital edition derived from it, Astronomical Diaries Digital (ADsD, via ORACC, CC0, Pirngruber/Rinderer/Harris 2022), would be worth their own notes on how prior human dating of the Diaries actually worked. — source: arXiv:2610.34297, refs 1–3, 10, 27.
• Parker & Dubberstein's 1956 Babylonian Chronology 626 B.C.–A.D. 75 (Brown University Press) supplies the month-start calendar table underlying the system's entire 701-year candidate space; worth checking against a review of its own reception/accuracy independently of this paper. — source: arXiv:2610.34297, ref 28 and Materials & Methods.
• A coding agent (Claude Opus 5) proposed some of the parser's text-reading rules from examples of wordings it had failed on; without those agent-proposed rules, benchmark inclusion rates fall by at most 2 points but the median "strict set" grows by about a third — a concrete, numbered example of an AI coding agent contributing to a published scientific pipeline's parsing rules, worth a dedicated note on AI-assisted research tooling. — source: arXiv:2610.34297, SI Appendix S2.
• The paper suggests the same astronomy-first dating method could extend to other ancient sky records, naming Greek horoscopes (Neugebauer & Van Hoesen, 1959) and eclipse reports in Chinese chronicles, and even to calibrating radiocarbon dates. — source: arXiv:2610.34297, Discussion and refs 24–25.
• Diary-derived data already feeds unrelated fields per the paper's citations: Earth's long-term rotation-rate changes (Stephenson, Morrison & Hohenkerk 2016) and early aurora-like-event records (Hayakawa et al. 2016) — both potential standalone leads on what the Diaries are used as evidence for, beyond dating. — source: arXiv:2610.34297, refs 7–8.
• The acknowledgments describe the project's origin story: Lu Li raised the dating problem in conversation with an archaeologist at Hebrew University; Yuan-Sen Ting assembled the team and originally used it as a coding-agent test problem around 2024–2025; Lucas Carrit Delgado Pinheiro, a national-team leader at the International Olympiad on Astronomy and Astrophysics (IOAA), joined to build and check the system — a human-interest thread on how the project formed, not independently verified beyond the paper's own telling. — source: arXiv:2610.34297, Acknowledgments.

Entity candidates

• A.J. Sachs & H. Hunger — people (deceased/historical) — the FOUNDATIONAL figures this entire paper measures itself against: authors/editors of ADART, the standard printed edition (1988–2022) whose tablet-by-tablet published dates are the benchmark ground truth the automated system is scored against throughout.
• R.A. Parker & W.H. Dubberstein — people (historical) — authors of Babylonian Chronology 626 B.C.–A.D. 75 (1956), whose month-start tables define the fixed 701-year candidate calendar the whole system searches over; a second, less-visible foundational dependency behind Sachs & Hunger.
• Lucas Carrit Delgado Pinheiro — person — co-first author (Ohio State, Dept. of Electrical and Computer Engineering); led a national IOAA (astronomy olympiad) team before joining to build/verify the system's code.
• Yuan-Sen Ting — person — co-first/corresponding author (Ohio State Astronomy; Humboldt Fellow; works in galactic archaeology); assembled the project team.
• Jon Clindaniel — person — archaeology advisor (University of Chicago, Computational Social Science); joined after debate over whether the project "counts as archaeology."
• Lu Li — person — originated the project question (Hebrew University, Racah Institute of Physics), via conversation with an archaeologist.
• Yingbin Liang; Ness Shroff — people — computer-science advisors (Ohio State).
• ADART (Astronomical Diaries and Related Texts from Babylonia) — term/concept — the standard printed edition of the Diaries; candidate for its own definitional note.
• Astronomical Diaries Digital (ADsD) — concept/term — the CC0 open digital edition (ORACC) of ADART I–III used as this paper's primary text input.
• Model Context Protocol (MCP) — concept — the tool-calling standard the paper uses for agents to invoke astronomical computation tools; already a recurring AI-infrastructure term worth a vault concept note if not already present.
• JPL Development Ephemeris DE441 — concept/term — the planetary/lunar position dataset underlying all astronomical verdicts in the system.
• Britton's zodiac-longitude relation (λ_Bab = λ_date + 3.20° − 1.3828°Y) — concept — J.P. Britton's (2010) formula fixing the Babylonian zodiac to the stars, used to convert modern ecliptic longitudes into the ancient reference frame.