# Hannibal's pass over the Alps, 218 BC

**Verdict: the Col du Clapier is the leading candidate, at about 40 per cent on the weights used here and on condition that Polybius's Island lies between the Rhône and the Isère. Several passes remain possible; they are ranked, and the order changes when one judgement changes.**

The crossing. In the autumn of 218 BC the Carthaginian general Hannibal took an army of forty-six thousand men and thirty-seven elephants from the Rhône over the Alps into Italy in fifteen days, with the first snow falling on the pass, a broken cliff path on the descent that had to be built up for the animals, and the elephants brought over "in a wretched state from hunger". Two ancient historians tell the story: Polybius, a Greek who wrote within about seventy years of it and walked the Alps himself, and Livy, a Roman who wrote two centuries later from the same tradition.

Why it is a mystery, and the sides. Neither names the pass, and the two place-names that would fix it are damaged in the manuscripts: the river of the Island is the Isère only by a conjecture of 1624, and Livy's "1,000 feet" for the broken ledge is a Renaissance restoration. The argument has run since 1574. If the Island's river is the Isère, the army went up that valley and over the Col du Clapier or the Mont-Cenis into the Susa valley; the Clapier has been the hedged answer of ancient historians writing in English since Perrin (1883). If the Island lay at Orange, as Livy's tribes imply, the army went up the Durance to the Montgenèvre, the Col de Larche or the Col de la Traversette, the high notch under Monte Viso that de Beer, Mahaney's earth-science team and a 2026 energetics study support. The St Bernard passes carry a Roman tradition Livy rejected.

How it was judged. Both texts read in the original with their manuscript variants; twelve routes over eight cols measured on a 90 m terrain model against Polybius's days and stades; each descent profiled for a broken cliff path; the view of the Po plain computed by sight-lines from every col; old snow judged against the snowline and glacier records; the Pleiades' setting computed for 218 BC; every Traversette radiocarbon date recalibrated. Fourteen lines of evidence were given likelihood ranges for eight passes and combined, with the nine Polybian items at half strength because they share one author. The Clapier comes first at about 40 per cent, the Traversette second at about 30 and the Mont-Cenis third at about 15; with the Island at Orange the Traversette leads at about 54. The Traversette's physical evidence is consistent with any large body of animals at the meadow between about 380 and 150 BC and does not test the crossing. Every share is a judgement, conditional on the items chosen, the ranges set, the halving of the Polybian items and the reading of the Island; the sensitivity table in the report shows how the shares move.

Headline numbers come from the evidence model's headline run (Polybian items at half strength because they share one source, everything else at full strength): Clapier 40% (5 to 95 per cent Monte-Carlo band 20 to 60), Traversette 29% (13 to 48), Mont-Cenis 16% (6 to 31), other col 9%, Larche 3%, Petit St Bernard 1.5%, Montgenèvre 1.4%, Grand St Bernard 0.3%. With every item at full strength the Clapier is 70%; with each source family collapsed to one item the table is nearly flat (15 to 18% for the top four).

Report: https://historical-mysteries.netlify.app/hannibal-alps/report ; a copy is `hannibal_alps_report.html` in this folder.

## Label for the site

Evidence label: **Leading candidate**. The Clapier ranks first among several passes that remain possible; with the Island at Orange the Traversette leads instead.

**Confirms** the cautious verdict for the Clapier that Perrin reached in 1883, that Azan (1902) and Colin (1904) developed, that Walbank (1956), Proctor (1971), Lazenby (1978), Lancel (1995), Hunt (2017) and Hornblower (2024) repeated with the same caution, and that Kuhle and Kuhle defended against Mahaney in 2012; and agrees with them that the question remains undecided. **Shifts** the Traversette case from "the first direct evidence of the route" (Mahaney and colleagues, Archaeometry 2017 part II; "solid evidence" in the 2016 press) to a strong second whose physical case has not been tested. Prior view named: no consensus. Peddie's 1997 count of twenty-four modern writers split six Mont-Cenis, six Montgenèvre, five Clapier, four Petit St Bernard, three Traversette; since then the Clapier has been the usual hedged answer of ancient historians writing in English, the Montgenèvre the scholarly alternative (Connolly 1981, Hoyos 2006 and 2015), and the Traversette the choice of de Beer's line, of Mahaney's papers and of Berti and Vollrath (PNAS 2026).

## What is new (as far as could be found)

All of it is new analysis of evidence already in print. No manuscript, artefact, date or measurement is added to the record.

- A computed viewshed of the Po plain for every candidate col, the first 10 km of each descent and the ground within 1.5 km of each col, on a single definition of "plain": only the Traversette (17% of bearings from the col, 22 to 26% from the knolls beside it) and, marginally, the Clapier (a 2-degree sliver from the col, 5% from 1 km down, 8% from the ridge) see any plain from the col or the descent; the Mont-Cenis has a half-percent sliver from three cells on the ridge above it, and the Petit Mont-Cenis, Montgenèvre, Larche, Petit St Bernard, Grand St Bernard and Agnel see none.
- A recalibration of all fourteen published Traversette radiocarbon dates against IntCal20 (seven in Archaeometry 2017 part I and seven in the 2017 magnetic-susceptibility paper, one lab number printed twice); a Ward and Wilson test showing the two "bracketing" dates at 40 cm are mutually inconsistent (T = 23.4, p = 1.3e-6) and both exclude 218 BC; the finding that nothing in the published age model's own output places the bed in 218 BC (its "firmly" and ">95%" rest on the agreement index, Amodel = 97.6, read as a probability); the finding that the model table's modelled columns are inconsistent with its own agreement indices; and the point that at this part of the curve no radiocarbon date can separate 218 BC from 207 BC or from c. 380 to 150 BC.
- The textual point that "Isère" stands in no manuscript of Polybius (Σκάρας/Σκώρας) or Livy (sarar/bisarar/arar) and is Cluverius's 1624 conjecture in both, and that Livy's "1,000 feet" for the landslip is a Renaissance restoration of a corrupt line.
- The demonstration that Polybius's 1,400 stades to the ascent and 1,200 for the crossing cannot both be satisfied by any line over the real terrain, so neither total discriminates; and the observation that 1,200 stades is fifteen days at the 80 stades a day of the preceding stage.
- An explicit evidence model (14 items, 8 candidates, likelihood ranges argued item by item) with a shared-source correction, Monte-Carlo bands, leave-one-out and seven named sceptical scenarios, so the Clapier-versus-Traversette argument can be located in three cells (the Island, the Allobroges, practicability).
- Smaller: the Pleiades' morning setting computed for 218 BC (true 29 October; visible 4 to 9 November) with two independent codes and a check against Columella; Pliny 3.34 and 3.36 used to show that an Island in the Valence plain belongs to the Segovellauni or Cavares with the Allobroges beyond the Isère, which matches Polybius's distinction between the Island's people and the Allobroges.

## Where the work fell short

- Not read in full (abstracts, snippets or catalogue records only): Walbank, Historical Commentary on Polybius I (1957) and JRS 1956; de Beer 1955 and 1967; Prevas 1998; Hunt 2017; Lancel 1995; Hoyte 1960; Seibert 1993; Proctor 1971; Mahaney 2008; Kuhle and Kuhle 2012 and 2015; Mahaney's Wiley papers 2010, 2013, 2019. Their arguments are known through the open literature and each other.
- Livy read in Foster's Loeb text (Wikisource transcription) with the Oxford apparatus from scans; the Polybian apparatus at 3.49.6 known only through Schweighäuser and the Oxford editors.
- Terrain at 90 m resolution: it locates ground of the kind Polybius describes, it cannot show a particular ledge or landslip existed in 218 BC; route lengths are valley-floor lower bounds.
- Snow bands are judgements from the modern snowline and glacier records; no southern-Alps autumn snow-patch series was found.
- No fieldwork; no fresh collation of manuscripts.
- The OxCal input for the Traversette age model (Mahaney and colleagues, Archaeometry 2017 part I, Table 1 and Fig. 4) was not obtained: it is printed in none of the six Mahaney texts held and no supplementary file is deposited with the journal, with either author's university repository or on Zenodo (attempts in `texts/ACCESS_LOG_revision.tsv`). The model's structure is known only from its published plot.
- Every likelihood range is a judgement; the shared-source correction is a rule of thumb, bracketed by the two extreme treatments.

## Figures

All four figures were made for the report; the code that draws each one is named here rather than in its caption.

- Figure 1, the candidate routes over a hillshade: from the Copernicus GLO-90 elevation model, OpenStreetMap river and place data, and the least-cost valley-floor paths; `code/routes.py` computes the paths and `code/figures.py` draws the map (`data/map_report.svg`, trimmed by `code/make_map_report.py`).
- Figure 2, elevation profiles of the seven main routes: from the Copernicus GLO-90 model along the computed paths; `code/routes.py` and `code/figures.py` (`data/profiles_report.svg`, via `code/make_profiles_report.py`).
- Figure 3, what each col sees: by ray-casting over the Copernicus GLO-90 model; `code/viewshed.py` computes the sight-lines and `code/make_viewshed_fig.py` draws the roses (`data/viewshed_report.svg`).
- Figure 4, shares in the headline run: from the output of the evidence model; `code/evidence_model.py` runs the model and `code/make_chart.py` draws the chart (`data/chart.svg`).

## Layout

- `hannibal_alps_report.html`: the report as published (self-contained; figures inline).
- `dossiers/`: the working reports: `sources_polybius_livy.md` (every datum in both narratives with variants and apparatus), `terrain.md` (route model, viewsheds, snow, astronomy), `geoarchaeology.md` (the Traversette papers checked link by link), `scholarship.md` (advocates of each pass, testimonia, state of the debate, access log).
- `code/`: `routes.py`, `viewshed.py`, `pleiades.py`, `figures.py`, `report_tables.py` (terrain); `calibrate.py` (IntCal20); `evidence_table.py` (the argued likelihood table), `evidence_model.py` (the model and every variant); `make_*.py` and `build_report.py` (figures, tables and the page); `frag/` (the report's sections); `report.css`.
- `data/`: `routes_output.csv/.md`, `routes.json`, `profiles.json`, `waypoint_distances.csv`, `viewshed_output.json/.md`, `pleiades_output.json`, `terrain_tables.md`, `radiocarbon_calibration.json/.md`, `intcal20.14c`, `evidence_model_output.json/.md`, the SVG figures.
- `texts/`: Polybius 3 and 34 (Greek), Livy 21 (Latin), Strabo, Servius/Varro, Sallust, Nepos, Pliny, Ammianus, Appian, Zonaras (`testimonia/`), the Mahaney papers and their critics, the climate and glacier papers (`terrain_sources/`), the public-domain scholarship from Whitaker 1794 to Torr 1924 (`scholarship/`), with `ACCESS_LOG.tsv`.
- `notes/`: the model notes.
- `code/factcheck.py` compares the facts (numbers, names, Greek text, quotations, links and ids) of two HTML fragments.

## Re-running

See the "Reproducing the results" section of the report, or `dossiers/terrain.md` (README section) for the terrain chain. Short form: `cd code; python3 evidence_model.py 20000; python3 make_chart.py; python3 make_tables.py; python3 make_extra_tables.py; python3 build_report.py`. To change a judgement, edit `code/evidence_table.py`.