Vitki Runor

Disclosure: Vitki is developed by Viktor Kvant, the author of Analysverket. It is an open, free research platform for analysing runestones, available at vitkirunor.se — and it applies the same discipline this site argues for in web analytics (measured data, transparent methods, cross-validated models, full provenance) to a very different dataset: inscriptions carved a thousand years ago.

The core is quantitative 3D groove analysis. Upload an STL, OBJ or PLY scan and Vitki measures V-angle, asymmetry, depth, width, bottom radius and surface roughness across cross-sections — manually, with one click per section, or fully automatically. In automatic mode it finds the grooves on the carved face and only measures those recognised as runes, filtering out ornamentation, band edges and natural cracks. Stone averages use the rune as the unit, are reported with spread and confidence intervals, and come with a warning when fewer than ten runes have been measured. Every result carries provenance: the file's SHA-256, the parameters and the method version.

The entire Samnordisk runtextdatabas (Rundata) is built in, searchable by signum, place, text, carver, style group, period and inscription type. A reading can be checked against Rundata word by word without any AI involved. Orthographic stylometry compares spelling, punctuation and bind-runes with known carvers' inscriptions and reports similarity with p-values, while eleven defined linguistic features and the purpose of the inscription — memorial, bridge-building, Christian prayer, journeys abroad and more — add further lines of evidence.

The headline goal is carver attribution: finding previously unknown links between stones, such as evidence that two runestones miles apart were carved by the same unnamed master. The synthesis weighs Rundata, orthography and carving technique by each method's cross-validated accuracy, then tests every candidate against the literature, geography, style groups, dating, bedrock (via the SGU bedrock map) and stone-to-stone measurements. Contradictions and missing evidence are reported; invented probabilities are not.

The statistics layer runs in R: carver territories and distance to water (sf), clustering and MCA (FactoMineR), formula and spelling analysis (tidytext, stringdist), a cross-validated random-forest attribution model (tidymodels), seriation against Gräslund's chronology (ca), formula and kinship networks (igraph), and the landscape around a stone — elevation, viewshed, Viking Age shoreline and least-cost paths (terra, gdistance). Results feed straight into the stone report and can be downloaded as reproducibility packages with R scripts and data.

Its stance on AI is unusually disciplined. AI only writes text or makes explicitly uncalibrated assessments; every number, table and figure is computed. Features that failed validation were removed rather than kept for show: image-based style classification (35% exact style-group accuracy in testing) and blind AI rune reading were both dropped in October 2026. That is the standard every analytics team should hold its own models to.

Output is publication-ready: stone reports in article form following SRI/Futhark conventions, with 3D paradata and figures generated from the scan, plus corpus reports covering several stones and export to Word, LaTeX and Markdown. Researchers can publish their groove measurements to a shared corpus under CC BY 4.0, where others can verify them. The source code is GPL-3.0 licensed and open on GitHub.