The Traversette is the one candidate with a physical case. Between 2011 and 2017 William Mahaney and a large team cored and sectioned a small alluvial mire at 2,580 m in the upper Guil, about 0.9 km west of the col and 370 m below it. They reported a churned, organic-rich bed at 38 to 50 cm depth, which they called a "mass animal deposition" bed, with elevated faecal lipids, a spike of Clostridia bacteria and a radiocarbon age model that "firmly places" the bed in 218 BC (Mahaney and colleagues, Archaeometry 2017, parts I and II; Mediterranean Archaeology and Archaeometry 2014 and 2017). Below the col on the Italian side they had earlier described a two-lobed rockfall which they identified with the landslip of Polybius 3.54.7 (2010; 2014). The authors called it "some of the first direct evidence of the route followed by the Punic Army", and the press called it solid evidence. I read all of the open papers, the 2020, 2024 and 2026 restatements and the critics (Kuhle and Kuhle 2012 and 2015, Hunt 2016, de Galbert 2016, MacDonald 2016), and recalibrated every published date against IntCal20 with my own code. The part that bears on the pass follows.
The dates and the bed are genuine, but they do not date the bed to 218 BC. In the original core the authors have two dates from the 40 cm level, 2070 ± 31 and 2530 ± 90 BP, which they say "bracket the invasion year". Each on its own excludes 218 BC at two sigma, in their own IntCal13 table (176 BC to AD 1 and 820 to 408 BC) and in IntCal20 (169 BC to AD 8 and 814 to 407 BC), and they are not two measurements of one age: the Ward and Wilson test gives T = 23.4 against a critical value of 3.84 (p = 1.3 × 10⁻⁶). Two dates from one depth that disagree by this much mean that the peat at that depth is a mixture of ages (which is what "churned" means), and a mixed bed cannot be dated to a year by any number of dates. The age model does not change this.
Table 1 of part I could be read more harshly. It prints, between the two 40 cm dates, a row "Hannibal: 218 BC, 2168 cal BP, ± 0", an event of known date and zero uncertainty with no agreement index beside it, which looks like the invasion year fed into the model as an input that the model then returned. I do not think the charge holds. A reference date printed in a results table does not show that the date was imposed on the model, and the paper's own plot of the model (Fig. 4 of part I) points the other way. OxCal's multiple plot draws a row for each element of a model as run, unless the user removes it; this plot shows eight radiocarbon dates in three groups ("Pre-invasion Beds", "MAD bed; Hannibal invasion", "Post-invasion bed") between four boundaries, inside a sequence named "Hannibal: Alps G5 and G5A sites", and no row for a 218 BC element. The 218 BC appears as a dashed vertical line and a dashed box drawn round the two 40 cm dates, with a label in another typeface. The simplest reading is that the table row is a reference entry and the line an annotation. I cannot confirm that reading either way. The OxCal input is not printed in part I, in part II or in the four later restatements (Mediterranean Archaeology and Archaeometry 2017 and 2020, Scientific Culture 2024, the 2026 paper), and no supplementary file is deposited with the journal, with either author's university repository or on Zenodo (I checked each). Without the input, neither the charge of circularity nor its denial can be verified.
What the paper prints is enough to decide the narrower question, because nothing in its own model output places the bed in 218 BC. The modelled ranges in Table 1 for the two 40 cm dates are 2464 to 1597 and 3002 to 1804 cal BP at two sigma. The modelled range for the boundary between the churned bed and the beds above it, the table's upper "Transition", is 2356 to 997 cal BP, which is 407 BC to AD 953. "Firmly places" and the caption's "high probability of >95%" rest on the agreement index, Amodel = 97.6, which measures how well the measured dates fit the model's ordering; it is not a probability that any bed dates to any year. Part II and the magnetic-susceptibility paper of 2017 carry the same reading over: "linked to radiocarbon dates precisely at 2168 cal yr BP" and "2168 cal yr BP (i.e., 218 BC with a 95% confidence interval)". The table has a further fault. Its modelled columns cannot be the posteriors of the dates in the model plotted: for UBA-23647 the unmodelled range is 3895 to 3720 cal BP and the modelled range 3509 to 2109, with no overlap, beside an agreement index of 98.3, and an index near 100 requires the posterior to sit on the likelihood. I cannot tell what those columns report. The plot's row labels carry the calibrated means (for example "2040 ± 45" for UBA-24875, whose radiocarbon age is 2070 ± 31) rather than the radiocarbon ages, so plot and table cannot be matched row by row without the input either. The dating evidence does not independently establish a disturbance specifically in 218 BC, much less identify the animals responsible as Hannibal's.
The later sections do not repair this. Section G5B has no date within 500 years of 218 BC and its "218 BC bed" is interpolated between a seventh-century AD date and a ninth-century BC one. G5D's "218 BC bed" is at the level of a sample whose median is 400 BC. G5C has the one measurement in the whole series whose range includes 218 BC, UBA-30330 at 51 cm, 2190 ± 35 BP: it puts 17 per cent of its probability in 230 to 200 BC and 42 per cent in 355 to 282 BC. Nor could any measurement do better. At this point the calibration curve is nearly flat between 350 and 250 BC and falls steeply afterwards, so a sample formed in exactly 218 BC and measured to ± 30 years would calibrate to 380 to 197 BC; a sample of 207 BC, when Hasdrubal took the same road, or of 231 BC, when a Gallic host crossed, would be indistinguishable from it. The claim names a single year; the method, for samples of this age, cannot resolve better than about two centuries. In 2026 Mahaney reread the 2070 BP date as peat that grew back "nearly a century" after the event, with every older date reworked; on that reading there is no date from the event at all, only a Bronze Age floor and a second-century BC ceiling. The dates allow a disturbance in the late third century BC but cannot distinguish it from one at any other time between about 380 and 150 BC.
The other parts of the argument are weaker still. The lipid evidence shows herbivore dung at every depth from 15 to 60 cm, with a peak in the churned bed that is no larger than the peak at 15 to 25 cm which the authors themselves assign to medieval transhumance; by the authors' own criterion, "the presence of deoxycholic acid in the absence of lithocholic acid indicates a ruminant animal source", the bile-acid pattern they report points to cattle, sheep, goats or ibex rather than horses. The Clostridia are a class of bacteria that dominate the gut of every large herbivore and also live in waterlogged peat; the sequencing was done at class level, in one section, with no ancient-DNA authentication, and the genome programme announced in 2016 has not been published. The "horse tapeworm egg" of the 2016 press coverage has become "roundworm eggs" in the 2020 paper and has never appeared with a taxon, an image or a depth in any peer-reviewed text. The mire is a wet meadow of about sixty by forty metres on the French side, below the col: it cannot be the camp where Polybius's army spent two days at the summit, and Mahaney concedes that "the Col de la Traversette itself could barely hold 100 soldiers". The rockfall exists and is older than 218 BC, but its two lobes are dated by lichen cover and soil development, methods whose resolution is a thousand years or more, so they cannot show that the younger lobe fell "recently" before the crossing, and cannot show it did not; and Polybius's words describe a ledge-path broken away along a precipice, which Hannibal "built up", not a debris sheet lying across a slope. No artefact of any kind has been recovered from the mire, the rockfall or the col. The one physical thing that the fieldwork has established is that large herbivores have repeatedly churned and dunged a watering place below the Traversette over the last three thousand years, which is what one would expect of any high pasture with a spring.
Nothing published since 2017 changes this. A further field season in 2017 produced no reported artefacts; Mahaney's 2019 Archaeometry paper (online 2018) moved to sections in the upper Po valley with a similar churned bed at "about 2,200 years"; the 2020, 2024 and 2026 papers are restatements by the lead author; no independent group has re-sampled or re-dated the mire; and the only new item, Berti and Vollrath's 2026 energetics model that finds the Traversette the cheapest route in joules, is a length-and-gradient calculation that does not engage with the mire, the descent or the rockfall. The dating arithmetic above has not, as far as I can find, been published by anyone; the objections to the camp, the species and the uniqueness of the rockfall were made by Hunt, Kuhle and Kuhle and de Galbert in 2012 to 2016.
In the evidence model the Traversette deposit therefore enters as an item whose likelihood is 1.0 for the Traversette and 0.75 to 1.0 for every other candidate, because the deposit is what a crossing would leave and equally what three thousand years of pasturing would leave. A scenario in which Mahaney's reading is credited in full is run for comparison.