What the evidence shows, where the instruments fail, and the next experiments worth doing.
Twenty-two objects, five visual questions and a dated history. The moai’s ring has a published egg interpretation worth revisiting. Source-informed readings now make the next independent review concrete.
Explore the comparison study →Why the earlier AI locations need checking · Next test design
The saved result reuses 11 pairings. Equal weighting gives 61.8%, but individual pairs differ sharply. The next test should use complete source episodes and close teaching-figure alternatives.
See every pairing →At 115 m, averaging the Baltic survey retains 0.3% of the measured local relief variance. This numerical test identifies a data limit; archaeological AI needs labelled structures and geological controls.
Inspect the actual survey →New descriptive analyses; no new AI judgments. Draft protocols, required evidence and scoring plans.
Opus ranked the pair highly; GPT-5.5 preferred many alternatives. Both results remain in the record. Some responses misread the moai’s back as a frontal figure, while the prompts mixed visible resemblance with shared-origin judgments.
The later recognition trials improve some labels but do not validate feature locations. Compare source-checked regions, equal-quality views and close alternatives before interpreting a similarity ranking.
Original pilots · Saved perception failures · Recognition diagnostic · Local examples
Rendering style changes the rankings. In independent-rendition validation, the combined metric retrieved the correct sign first in 75.4% of queries, below HOG alone at 80.5%; the embedding component reached 11.6%.
Choose a method on development examples, then test held-out signs and modifier families. A visual match does not establish shared sound, meaning or descent.
The sampled broad categories recur widely. They do not test a bird-and-disc composition, and uneven documentation limits negative findings.
Code precise relations against actual passages. Preserve unknowns and variants. Geographic dispersion alone cannot distinguish inheritance, borrowing and convergence.
Claimed partners ranked first in 248 of 397 completed lineups. The summaries were recognisable: Opus identified 96 of 103 regions in a recognition check. Repeated lineups and mismatched roles weaken the original chance comparison.
The V2 audit checks 29 source passages; it did not rerun the rankings. A better pilot would separate generic roles from specific event sequences and show the strongest alternative matches.
Original lineups · New pair-level audit · Source-based redesign
One model’s coding of compiled summaries shows regional patterns. Post-contact recording does not prove borrowing; regional structure does not exclude older inheritance.
Trace witnesses, independently code specific sequences, and separate related versions. Which birds leave, in what order and with what result is more precise than “animals occur”.
Original prompts, protocols and responses remain unchanged. New analysis is labelled separately. Software checks protect the records and the site; they do not certify archaeological interpretations.