the paper — a working draft, not peer-reviewed, not for citation
Conor Reid — working draft of 2026-08-01 — download PDF
The provenance of the Uluburun shipwreck's 105-ingot tin cargo is contested between a Central Asia + Taurus attribution (Powell et al. 2022) and a European counter-reading (Berger et al. 2023; Williams et al. 2025), both argued from visual comparison of isotope biplots. We ask the prior question — is deposit-level attribution statistically identifiable from these data at all? — for the Sn-isotope channel. From a harmonized 11-source cassiterite library we fit per-ingot and batch-level Bayesian mixtures and subject them to posterior predictive checks, fractionation-prior sensitivity, and source-library perturbation. Per-ingot classification fails all three diagnostics: its attributions track whichever source has the smallest variance. A batch-level model survives them and supports exclusions only, with graded strength: Mushiston lies 6–7σ from the 86-ingot cargo bulk and its exclusion as sole batch source withstands fractionation-prior sensitivity, per-batch fractionation, and doubled reference dispersions, whereas Spain's ≈3σ exclusion remains fractionation-conditional, and the five positive-δSn sources (Cornwall, Erzgebirge, Taurus, Serbia, Portugal) remain mutually unidentifiable. Negative δSn does not imply Central Asia: Iberia is the single largest component on the cargo's negative-δSn group. Pre-registered trade-route priors pass through the likelihood nearly unchanged, and a robustness-mixture weight stays at its prior mean: the positive apportionment is prior-driven; the exclusions are not. One isotopic dimension bounds admixture only weakly. The Pb, trace-element, and comparanda channels remain uncomputable here, which is not unidentifiable in principle. Three structural gaps cap all analyses: figure-only Taurus values, no Karnab ore data, and a ~857-analysis compilation never deposited as a dataset.
The draft has absorbed three rounds of adversarial review, each recorded verbatim in the repository together with what was run in response (paper/review_round1.md through review_round3.md). The rule throughout: every criticism gets a computation or a concession, never just a rebuttal. Two review findings cost the paper claims it had made — the admixture bound and a library correction — and both are reported in the current draft as findings, not buried.