From 312,000 Remains to 10,000 Pins

Sweden's heritage board, Riksantikvarieämbetet, keeps a register of about 312,000 archaeological remains. Burial mounds, cairns, rock carvings, charcoal pits, runestones, hearths. Almost all of it is visible above ground and legally protected. It is also, as a list of things to go and see, close to useless, because a charcoal pit and Tanum's rock carvings sit in the same table with the same weight.

I built a map from it, and the map shows 10,000 pins. This post is about the one decision underneath it: which 10,000?

A place is not a record

The first step is to stop counting records. A grave field of forty stone settings sharing one signpost is one destination, and if I score records, that destination gets counted forty times. So before anything is ranked, the register is grouped into places, and the 312,000 records become 251,029 of them.

The grouping rule is deliberately plain. Registered numbers like "Askim 268:1" and "Askim 268:2" are parts of one site, so they are merged, but only within the same parish. The number prefix alone turned out to be unsafe: parish names repeat across the country, and one bare key spanned two parishes more than a thousand kilometres apart. A spatial guard then splits any group whose members are still far apart. Records with no usable number stay on their own. I used to cluster those by distance as well, and the script's own notes record how that went: it made 2,024 clusters averaging 127 m of spread, the worst of them 535 sites over six kilometres. I removed it.

Ranking without a ground truth

There is no list of "places worth visiting" to learn from. What exists are weak labels: Wikipedia articles, and a small hand-labelled set. The pipeline fits a regularised logistic regression on those, with features taken from the register itself (does it have a folk name, is there a polygon, how long is the description, what class is it) and from outside it, like distance to the nearest road.

The most useful thing I did here was to distrust my own labels. In the measurements, 1,185 of the 1,425 "notable" places were runestones, because Swedish Wikipedia has an article for nearly every one. Counted naively, "has a Wikipedia article" is roughly the same as "is a runestone", and the model becomes a runestone detector. With runestones left out, the folk name signal goes from 3.15x to 8.54x, and the two geometry signals flip from looking harmful to looking strong. The confound did not just add noise, it reversed signs.

The same problem came back later in a smaller form. Inside the runestone class, having an article says almost nothing: 66.0% have one, against 1.0% of grave fields. A single global weight handed every runestone credit that the feature had earned elsewhere, and runestones were 1.13% of the eligible pool but 12.32% of the export. The fix was to score documentation relative to the class's own base rate, so a grave field with an article stands out and a runestone with one does not.

Two scores come out. One judges a place only by what it is, and can be honestly validated. The other also credits Wikipedia and photographs, which is better evidence for ranking but circular to validate. The map uses the second, and the code is open about that trade.

The cut, and the thinning

Take the best 10,000 by that score and you still have a problem: a national top list leaves whole regions empty when you zoom out. So every pin gets the lowest zoom level at which it may appear, assigned greedily, best first. Each tile is divided into a 24 by 24 grid, and a pin becomes visible at the first zoom where no better pin already owns its cell. Zooming out thins the map, but what survives is the best thing in each area, not a random sample.

This runs once per class family, because of a failure I would not have guessed. Thinning everything together keeps only the single best place of any kind in each patch of ground. Hide that one with a filter and the patch goes empty, since there is no runner-up in the tile to promote. Per family, every filter combination stays evenly covered.

The tiles come out of Tippecanoe, and several of its defaults had to be switched off. Its automatic thinning once dropped 9,999 of the 10,000 features at low zooms, so the map went blank when zoomed out. Its point clustering merged neighbours and put the result at their average position, which landed one hill fort 595 m from where it is. Now I decide visibility myself, and the tiles never invent a position.

Descriptions that add nothing

The register is written by surveyors, for surveyors: "Stensättning, rund, 7 m diam och 0,3 m h." Often it is a single line. For the pins I generate a short visitor description with a local model running under Ollama, and the main risk is a model that knows a lot about the Bronze Age and wants to tell you about it. The comment at the top of the script puts it directly: on a map of 10,000 places a plausible invention is worse than a dull fact.

So the model gets temperature 0, an output schema, and an instruction to use only what is in the source. Afterwards a mechanical check compares every number in the output against the input. It flags rather than rejects, and it cannot see everything, so there is a separate guard for invented periods. Today 7,111 of the 10,000 pins carry a generated description. The rest show the register's own text, and the README says plainly that this is a long model run away from being fixed.

How the site uses it

On this site the output is just static files: vector tiles served from the public folder, and the map reads them with MapLibre. All the judgement happens earlier, in Python. If you want the code, the pipeline is at github.com/francomay3/fornlamningar. It replaced an earlier, simpler version, Fornkartan, which downloaded the register, diffed it, fetched details, parsed and built tiles, with a handful of hand-set weights (a proper name, excavated, field-confirmed) instead of a fitted model.

What I take from it

Most of the work was not the model. It was finding where my labels were lying, and writing the exceptions down where the next run would find them. Choosing 10,000 pins out of 251,029 places is a statement about what deserves attention, and it is better made with measurements you can read back than with a score you merely trust.