The Antikythera mechanism was retrieved from a shipwreck off the Greek coast in 1901. ³ Only about one-third of the original device survives, now divided into 82 fragments that include 30 corroded bronze gearwheels. ² Its functions have long been described as controversial precisely because both the gears and the inscriptions on its faces are only fragmentary. ¹

Understanding of the object has advanced in imaging steps. A 2005 investigation used high-power micro-focusing X-ray tomography built by X-Tek Systems, with results published in Nature in 2006. ¹ That work used surface imaging and high-resolution tomography to reconstruct gear function and to double the number of deciphered inscriptions. ¹ Subsequent studies — a 2018 PLoS ONE reconstruction, a 2021 Scientific Reports model, and a 2025 engineering analysis — have progressively refined, and in places challenged, what can be said about the device.

The evidence base has an important structural feature the reader should carry throughout: two-thirds of the device is missing, roughly 80% of its original text is not recovered, and corrosion since retrieval has altered the surviving metal. Every claim about capability or operability is made against that backdrop of absent evidence. ⁴

Direct examination establishes real, decodable sophistication. The 2005 micro-focus X-ray CT decoded the rear structure of the mechanism, although the front face remained largely unresolved due to loss of physical evidence. ² The 2006 Nature work reconstructed gear function and doubled the deciphered inscriptions. ¹ According to Freeth et al., a 2016 discovery of the numbers 462 (Venus) and 442 (Saturn) in the Front Cover Inscription revealed accurate period relations that were factorizable and mechanizable — that is, expressible as gear ratios. ¹ Wikipedia further reports the mechanism appears to presuppose eccentrics and epicycles, the mathematical apparatus of Hellenistic astronomy. ¹

The decipherment has advanced but remains partial. PLoS ONE reported that about 3,000 characters had been transcribed out of an estimated 15,000 originally present, each averaging 1.6 mm high. ¹ A 2023 Springer conference proceeding reported approximately 3,500 letters and symbols deciphered. ¹ These are not conflicting counts of the same quantity: the analyst treats the 2018 (~3,000) and 2023 (~3,500) figures as sequential snapshots of progressive decipherment over five years, a convergent trajectory rather than a contradiction. ² Even so, the University of Athens notes the instruction manual has not been fully read because parts of the inscribed plates are missing. ¹

Gear inventories illustrate why 'how many gears' is harder than it sounds. Nature's 2021 paper and Britannica both report 30 surviving gearwheels, and the analyst's triangulation classes these as convergent on roughly 30. ³ The 2021 model reconstructs a maximal 34 gears in front of the b1 wheel plus 35 extant gears behind it, totalling 69 — arithmetic the analyst verified (34 + 35 = 69). ¹ Against that maximal reconstruction, the Springer proceeding reports at least 39 gears: 29 identified in calcified fragments plus ten inferred astronomically. ¹ The analyst classes the 69-versus-39 gap as divergent but explicable — Freeth's 69 is a hypothetical full reconstruction, the 39 a conservative minimum — with both acknowledging extensive missing evidence. ²

On the surviving construction itself, single sources document fine detail: a doctoral thesis found both back-face spirals were designed as Half Circles Spirals drawn from two centres; the University of Athens reported a copper-tin alloy with tin at 1–14% and lead up to about 3% acting as a lubricant; and gear teeth were equilateral triangles with an average circular pitch of 1.6 mm, hand-made and not all even. ³ The analyst flags that these specifications each rest on a single source, and that alloy figures across sources use different precision and framing rather than confirming one another. ¹

Open: Do the 462/442 Front Cover numbers and the 69-gear breakdown, both single-sourced to Freeth 2021, hold under independent examination of the paper's full data? [C-003][C-004]; How many gears does Fragment A actually contain — 27 (Nature 2021, Heritage) or 28 (Springer 2023)? No source explains the single-gear discrepancy [C-001][C-029].

Physical reconstruction is central to the operability question because a working replica demonstrates that a design can, in principle, run. Within this source set, that evidence is real but narrowly sourced. A YouTube channel reports that Michael Wright, a former London museum curator, built a handmade working reconstruction incorporating all then-known features, including pointers for the Sun, Moon and five planets, a moon-phase display, a Metonic calendar spiral of 235 months in 19 years, and a Saros eclipse dial of 223 months. ¹ Both claims rest on a single enthusiast channel, the weakest source tier in the set; the analyst notes Wright's reconstruction is widely documented elsewhere but not within these graded sources. ¹

The Aristotle University team's reconstructions come entirely from one 2021 MDPI Heritage article. It reports the 2011 replica reproduced predictions matching NASA data for eclipses including 15 June 2011, and that five accurate 1:1 functional models were built for museums including the National Archaeological Museum of Athens and the Musée des Arts et Métiers. ¹ Crucially, the same article reports a countervailing result: a 2011 model built using mean data was non-functional, with some gears blocked or not in contact due to roughly 1 mm tolerance issues. ¹

The analyst classes all three Aristotle University claims as single-source: they distinguish a non-functional mean-data model from five successful models, but no independent source corroborates any of these specific results. ¹ This matters for interpretation. A functional replica built with modern tools and original design specifications demonstrates the design concept is sound; it does not by itself establish that the ancient artifact, at ancient tolerances, ran — a distinction the same source's own mean-data failure makes concrete. ¹

Open: Can the 2011 AUTH replica's NASA-eclipse match and its five functional museum models be independently confirmed beyond the single Heritage article? [C-031][C-033]; Were any reconstructions built to reproduce ancient manufacturing tolerances rather than modern precision, and did those run? [C-032]

This is the inconvenient question, and the evidence pulls in two directions. On one side, 2025 computational work concludes the original could not operate reliably. The Jerusalem Post reports that computer simulations by researchers at the National University of Mar del Plata indicate the mechanism could get stuck or have teeth disengage due to poor gear tooth spacing, and that a study published on arXiv found manufacturing errors exceeded allowable limits for stable operation. ¹ New Scientist reports the same research suggesting the gears may not have been able to turn due to manufacturing tolerances and corrosion-induced dimensional changes. ¹ Wikipedia reports that in 2025 one research team concluded manufacture error was too great for the mechanism to have worked — while explicitly noting the scans may overstate imperfections. ¹

These four reports appear multi-sourced, and the analyst's triangulation classes them as convergent — but with a critical caveat: all four trace to the same underlying Mar del Plata arXiv preprint reported by different outlets. This is consistent cross-outlet coverage, not four independent replications. ³ One further echo, a Facebook page stating that geometric tooth issues could make the mechanism prone to jamming, is graded unverified: its sole source is an anonymous page with no editorial oversight and suspected content injection. ¹

On the other side sits a confounder that undercuts confidence in the tolerance measurements themselves. Multiple independent sources — PLoS ONE, New Scientist and the Jerusalem Post — confirm that the bronze turned into atacamite, which cracked and shrank when the fragments were brought up from the shipwreck, changing the dimensions of the pieces. ³ The 2005 highest-resolution scan of Fragment A had about 27 missing projections out of 2,957 acquired — a failure rate the analyst computes at roughly 0.9% — causing degraded reconstructions. ¹ If measurements are taken from corroded, dimensionally altered fragments, then measured 'manufacturing error' may partly be corrosion error — which is exactly the caveat the 2025 team itself flags. ⁴

There is also a documented design feature that complicates the naive 'poor spacing' reading. The University of Athens reports that each pair of meshing gears has slightly unequal teeth to leave a gap, precisely so they do not break from friction and torque — meaning some tooth-spacing irregularity may have been intentional tolerance, not defect. ¹ The analyst cautions that the various tolerance claims reference different dimensions and methods (tooth spacing versus gear engagement versus overall tolerance) and are not directly comparable without common units. ² Finally, the fragmentary state that opened this investigation reasserts itself here: functions remained controversial because the gears and inscriptions are only fragmentary, and the 2021 model itself added the Line of Nodes 'Dragon Hand' as a hypothetical element for which there is no direct physical evidence. ²

Open: What are the Mar del Plata preprint's actual simulation parameters and tolerance measurements, and can its simulations be independently replicated? [C-013][C-014]; Can metallurgical modelling separate corrosion-induced dimensional change from original manufacturing error on the surviving fragments? [C-010][C-020]

On comparative sophistication the sources are unusually aligned. Wikipedia reports that machines of comparable complexity did not reappear until the fourteenth century, with early examples being the astronomical clocks of Richard of Wallingford and Giovanni de' Dondi. ¹ Britannica reports the device contained 30 gear wheels, a level of complexity unmatched until medieval cathedral clocks were built a millennium later. ¹ The analyst's triangulation classes these — together with Nature's 2006 statement that the device was more complex than any known for at least a millennium thereafter — as a convergent consensus of independent sources on a gap of roughly one thousand years. ²

The comparison holds against the nearest surviving rivals. Wikipedia reports that a Byzantine geared calendar-sundial from the fifth or sixth century AD and a thirteenth-century Islamic geared astrolabe were mechanically much simpler than the Antikythera mechanism. ¹ This is the strongest, best-corroborated part of the sophistication case: whatever remains unresolved about operation, the device's integration of gearing and astronomical cycles stands well ahead of anything documented for centuries around it. ²

Open: The relative simplicity of the Byzantine sundial and Islamic astrolabe rests on a single tertiary source; primary scholarship on those devices is not in the record [C-018].

Four explanations compete, and the evidence discriminates among them unevenly.

H1 — the mechanism was a fully functional astronomical calculator — is plausible. If true, we would expect decoded, mechanizable astronomy (present: doubled inscriptions, Venus/Saturn period relations) and working reconstructions (present, but single-sourced). ³ It is directly contradicted by the 2025 tolerance findings that the gears may jam, disengage, or fail to turn. ⁴ The discriminating evidence — independent controlled testing of replicas built to ancient tolerances, or measurement of original fragments — does not yet exist in the record.

H2 — it could not function reliably due to manufacturing tolerances — is equally plausible. Its support is the convergent 2025 reporting; its contradiction is the reported functional replicas. ⁴ The decisive weakness is that H2's support, though it appears in four outlets, traces to a single arXiv preprint, so its apparent independence is illusory. The Mar del Plata preprint's full parameters and an independent replication would discriminate. ¹

H3 — corrosion and retrieval damage obscure the original functionality, so current measurements may overstate imperfection — is the best-supported of the four, and notably has no contradicting claims in the record. It rests on the multiply-confirmed atacamite corrosion mechanism and on the 2025 team's own admission that the scans may overstate imperfections. ⁴ This hypothesis is not an alternative to H1 and H2 so much as a reason neither can currently be settled from the physical evidence: if the fragments are dimensionally altered, both 'it worked' and 'it couldn't work' inferences drawn from them inherit that uncertainty. ³

H4 — the mechanism was a symbolic or display artifact rather than a working tool — is weak. Its only support is the researchers' own proposal, reported by the Jerusalem Post, that Esteban Szigety and Gustavo Arenas suggest it might have been a symbolic gift or toy rather than a functioning device — a claim graded speculative, which the researchers themselves acknowledge as such. ¹ It is contradicted by the decoded astronomical functionality and the reported working replicas. ² This is a downstream inference from H2's tolerance findings, not independent evidence of original intent, and the archaeological cargo context that could test it is absent from the record.

Open: Does the shipwreck's cargo context indicate a scholarly instrument or a luxury prestige object, which would bear on H4? [C-015]

The evidence forces one firm conclusion and withholds another. On sophistication, the record is strong and convergent: 30 surviving gearwheels, decoded rear gearing, doubled inscriptions, mechanizable Venus and Saturn period relations, and a complexity unmatched until fourteenth-century clocks. ⁴ That the device was an extraordinary piece of ancient astronomical engineering is not seriously in dispute within these sources, even among those questioning its operation. ¹

Whether the original artifact ran as designed is genuinely unresolved. The strongest affirmative evidence — working replicas, including one reportedly matching NASA eclipse data — is real but single-sourced to one MDPI Heritage article, and even that article records a companion model that failed at roughly 1 mm tolerances. ¹ The strongest negative evidence — a 2025 conclusion that manufacturing error was too great to work — is convergently reported but traces to one arXiv preprint, and its authors concede corrosion-era scans may overstate the imperfections. ³ Between them stands the best-supported and uncontradicted finding: corrosion to atacamite altered the fragments' dimensions, so measurements taken from them cannot cleanly separate ancient error from post-recovery damage. ³ This is not a tie to be broken but a stalemate the physical evidence cannot currently resolve.

One distinction should anchor any reader's takeaway, and it is supported rather than speculative: a modern reconstruction that runs proves the design concept is viable; it does not establish that the ancient object, at ancient tolerances, ran. ¹ The 2025 simulations address the latter question, the replicas largely the former, and the two are not in direct contradiction so much as answering different questions.

SPECULATIVE — offered as labelled reasoning, not finding: the proposal that the mechanism was a symbolic gift or toy is the weakest explanation in the record, resting solely on a speculative claim its own authors flag as such and contradicted by the decoded astronomy. ² On current evidence it is best read as an inference downstream of the tolerance debate rather than a positively supported account of the maker's intent. The question the brief poses — how sophisticated, and could it run — thus resolves asymmetrically: demonstrably sophisticated, operability open pending independent replication and metallurgical work that separates corrosion from craftsmanship.