Interpret, he who can II – what stood in the gap?

Part II of the series Interpret, he who can. The first part asked whether the inscriptions no one has been able to interpret are language at all. This one asks what stood in the gaps in the inscriptions that can be read – and uses Part I's language test as a check.

Anyone who reads a runic inscription in an edition sooner or later meets three dots. Where a rune has flaked away, where the stone has cracked, or where a seventeenth-century drawing cannot be trusted, the editor puts a dash for each missing rune. And where the runes that remain are not enough to say which word it was, the interpretation just reads “…”. The dots are not carelessness. They are a decision. A runologist has stood in front of the stone, seen what is left, and chosen not to guess.

We have built a probability model that does exactly what the runologist refrained from: it proposes what may have stood there. It is our own, it is new, and no one outside Vitki has yet examined it. So it never stands alone. We do not claim that any of the words below are facts. We show what the model says, how the proposals are tested and where they meet the editions.

Our solution: two models that check each other

A model that fills gaps has a built-in problem: it always finds something. So every proposal has to pass three tests before we bring it forward.

1. The probability is calibrated. The model's percentages are tested on gaps where the answer is already known, so that 80 per cent means it has been right about eight times in ten.
2. A second model asks whether the text is language at all. The language test measures how language-like a sequence of runes is, entirely independently of the first model. If the test says the inscription is not language, the proposal falls. The test can only judge inscriptions that are long enough and uninterpreted.
3. The edition has the last word. If the editor has described traces on the stone that speak against the proposal, the stone takes precedence over the model.

It works like a writer and a proofreader: one writes freely, the other strikes out what does not hold, and the editor – here the edition and the stone – decides. Neither is enough alone. Together they give a short list of proposals worth testing on the stone.

What the model proposes

The inventory went through 685 gaps on Viking-age runestones. 112 of them the editors have left open. In 17 of the open ones, on 15 stones, the model proposes a word from its vocabulary with at least 50 per cent calibrated probability, and in 7 with at least 80 per cent. Calibrated means that the figure has been tested on gaps where the answer is already known: when the model says 80 per cent, it has been right there about eight times out of ten. The model's own, uncalibrated figure is often higher, and it appears as “raw” in the chart below.

A first pattern shows at once: the model is surest where it is needed least. Among the seven proposals of at least 80 per cent, ok appears three times, with hans, ræisa and risti – formula words that any runologist can guess and that the editors still left open, mostly because the runes are not visible – and þikra on Sm 47, to which we will return. That is less exciting than new names. But it is also what an honest model should do: be surest where the pattern is strongest.

What the editions say

We read what the editions write about every gap on the runestones: Sveriges runinskrifter, Fornvännen, Nytt om runer and, for Öland, Magnus Källström's new reading in Futhark. A second reader then went through every claim and tried to refute it. The editions are seldom silent. In most cases the editor has seen the gap, described what is visible, and refrained from filling it.

On U 765 the edition writes of the rune in the gap: “Av r. 12 återstår hst och möjligen spår av en bst t. v. om hst” – a main staff and perhaps traces of a branch. The model fills in n, so that it reads ans, and takes it as hans; it accepts the missing h because carvers sometimes left it out. But the edition's conclusion stands: “Inskriften är tyvärr alltför skadad för att tillåta ett tolkningsförslag” – the inscription is unfortunately too damaged to allow any proposed interpretation. The list says 97 per cent, but the stone analysis, which reads the whole stone at once, only 36. It is a proposal that can be tested on the stone's fragment in the Enköping museum, not an answer.

Enlarge ⤢

On Sö 245 at Tungelsta the model proposes mag, “kinsman by marriage”, with 73 per cent. Rundata leaves the gap open, but the Swedish National Heritage Board's newer database Runor has already entered the same word: ma[g] sinn, with the note “Tolkad av redaktionen” – interpreted by the editors. The model has never seen that reading. It reached the same word on its own, from the runes and from how runestone texts are usually phrased.

Where the second model says no

The second pattern is the most important. Five of the 15 stones have inscriptions that the editions do not read as language at all:

• U 466 at Tibble: “Inskriften ger ingen språklig mening” – the inscription gives no linguistic meaning.
• U 493 at Brunnby, known only from Peringskiöld's drawings and woodcut of 1684: the same words.
• U 522, a round sandstone plate from Hammarby: an inscription “som icke ger någon språklig mening” – which gives no linguistic meaning – and which ends with the rune row.
• U 1180 at Åby: “yet another example of inscriptions that have no linguistic meaning”.
• Sm 47 at Håringe: the runes “give no coherent inscription with any meaning in it”. The stone may also be medieval or later.

Enlarge ⤢

Still the model fills in words there: risti with 85 per cent, þikra with 93 – a rune sequence found only uninterpreted on another stone, Ög 60 – vaR with 75, ok with 73 and a name with 56. It is built to find language, and it finds language even where there is none. It does not know that a carver sometimes cut runes without writing words – as ornament, as a rune row, perhaps as magic.

This is where the second model comes in. The language test, which we described in Interpret, he who can, measures how language-like a rune sequence is: how well it can be predicted by a model trained on interpreted inscriptions, compared with real texts clipped to the same lengths and with the same runes in random order. The rules of the test were locked on 16 October, with a correction of the data on 17 October, before this article was conceived. It has judged all five: U 1180, U 493 and Sm 47 as “not language”, U 466 and U 522 as “rune row”, which means that at least six runes stand in the order of the futhark. Those are exactly the five that the editions call meaningless. Three other stones in the list, U 784, U 246 and Öl ATA4377/56A, the test judges language-like, and there the editions say nothing about the text lacking meaning. Those three, like Sm 47, have only ten readable runes, the minimum the test accepts, so those judgements are weaker.

The test is not infallible. Among all uninterpreted runestones it calls 27 “not language”, and Rundata's own notes call only 9 of them nonsense inscriptions. It has also been able to judge only 8 of the 15 stones here; the others are too short or already partly interpreted. But where it objects, it objects in the right places.

A sixth stone may belong here. U Fv1976;108 stands at Tibble, a few metres from U 466, and its editor Helmer Gustavson wrote in 1976 that the inscription “may never have given any linguistic meaning, but been an imitation of the inscription on a runestone”. The model proposes sinn there, with 52 per cent. The stone has too few readable runes for the language test.

Two models keeping each other in check: one fills in, the other says where there is nothing to fill. That is, we think, the most useful thing we have learned from this work.

Where the stone disagrees

On U 525 at Österlisa, three fragments of the same stone, the model proposes a place-name, hiortstaðum, with 73–76 per cent in the three gaps. That would be a fine find: a farm whose name began with Hjort-. But the carved text begins with au, not hio, so the word fits only because the model allows certain deviations in spelling. And the edition has described what is visible: “Runan 35 har möjligen varit k” and “Runan 38 förefaller att ha varit ett i” – rune 35 may have been k, rune 38 seems to have been i. The model needs other runes in both places. It works from the runes that are missing, not from the traces the editor saw. Here the stone disagrees, and then the stone carries more weight.

On U 246 at Mällösa, a fragment built into a cellar wall, the model proposes sinn with 56 per cent. After the third rune the edition sees “a short vertical stroke, which may be a divider, or possibly the rest of a main staff” – so the word may continue past the gap. An older reader also took the word before eftiR as the end of “stone”. The proposal is not impossible, but the edition gives it no support.

On Öl ATA4377/56A in Hulterstad church the model proposes ok with 83 per cent. No edition interprets the place, but a later archive reading has a different sign where the model puts r, and reads the rune before it as part of a bind-rune. There too the reading of the stone comes first.

On Öl 57 at Källa the model proposes ræisa at the end of the inscription, with 86 per cent. Magnus Källström has pieced the stone's fragments together and reads it as “… lēt rētta stæin þe[nna] …” – the raising formula already stands earlier in the inscription, so a second “raise” at the end would be unexpected. Of the runes known before his new reading, which is just the end, he wrote that “it is of course impossible to interpret”. An older note does suggest that the fragments may come from two stones. But the model reads Rundata's older text and in any case lacks the context the new reading gives.

Other objects

The list also includes inscriptions that are not runestones, and those we count separately: two rock carvings, a grave slab, two rune bones, two metal sheets – one of them an amulet – and a piece of slate. We have not examined them as closely as the runestones, and for the copper sheet Öl Fv1976;96A we have not read the edition at all. But two of the clearest agreements are there, both with reservations.

Enlarge ⤢

On the rock carving at Fällbro, U Fv1946;258, Sven B. F. Jansson wrote in 1946: “I runföljden 68—73 finnas måhända spår av ristarens namn” – in runes 68–73 there may be traces of the carver's name. He attributed the carving to the carver Visäte on other grounds. In exactly those runes the model proposes viseti, with 61 per cent. The agreement is not fully independent, though: Rundata's second reading of the carving already has the name Viseti, and the model weighs in carver names from nearby stones.

On the rock carving Sö 23 on Mörkö, damaged by fire before the seventeenth century, where this passage survives only in Bureus's drawing, the edition reads the end, with reservation, as “Gud hjälpe ande och själ!” – God help spirit and soul. The model fills in and, spirit, with 78 per cent – the same word, but it places it differently from the edition and leaves out hjälpe.

The amulet from Alsnö hus, U NOR1994;26A, shows a different limit. There the model proposes fængit with 62 per cent. But the signs the model fills are not lost runes. They are carved signs that no one has been able to identify, and Rundata calls the inscription “Nonsensinskrift, magisk?” – a nonsense inscription, magical? The model treats an unidentified sign as an empty place. We will change that.

What we have to stand on

This is what we can stand behind:

• The model is tested where the answer is known. On gaps the editors themselves filled, it was right in 16 of 21 in a blind test locked in advance, and all 13 proposals where it said it was at least 95 per cent sure held. The blind test was made with an earlier version of the model.
• Where scholars have ventured a proposal, it often lands on the same word – mag on Sö 245, and with reservations viseti at Fällbro and ande on Sö 23 – without having seen their readings.
• It does not know when a text lacks meaning, but the language test objects. Of the stones the test could judge, it called exactly the five that the editions call meaningless not language or a rune row.

And this we cannot stand behind:

• The open gaps are harder than the gaps it was tested on. The editors left them open for good reasons, often because the stone is damaged just there. The figures may be too high in them.
• The model does not see the stone. Where the edition describes traces that speak against it, as on U 525, the traces come first. None of the fifteen stones is among the open 3D scans available to us.
• It lacks context a runologist has, such as Källström's new reading of Öl 57.

When everything is weighed together, 5 stones remain where nothing in the editions we have been able to read speaks against the proposal: U 765, U 784, Öl ATA4701/43, Sö 245 and Ög 115. The words are fyrir, hans, mag and ok. For U 784 the reading of the gap itself comes from a later dissertation we have not been able to read. It is not a list of new finds. It is a list of where a runologist with the stone in front of them could settle the matter – with a proposal to test and a figure that says how sure it is.

Try it yourself

Under Research gaps in Vitki is the list with every proposal, the calibrated and the raw figure, Rundata's term for the object, and a Swedish draft translation of the model's reading. The draft is word for word, from a glossary the model learned from Rundata's own translations, and words it cannot translate stay in Old Norse. Open a stone, read what the edition writes about the gap, and see whether you agree with the model. Often one does not, and that is exactly why it is worth trying.

Sources

• Scandinavian Runic-text Database, version 3.1 (ODbL), and the Swedish National Heritage Board's runic register Runor (2020).
• Sveriges runinskrifter: the volumes for Öland (1), Östergötland (2), Södermanland (3), Småland (4) and Uppland (6–9). Stockholm 1900–1961.
• Gustavson, H. 1976. Runfynd 1975. Fornvännen 71, pp. 96 ff. (Tibble pp. 108–109).
• Jansson, S. B. F. 1946. Några okända uppländska runinskrifter. Fornvännen 41, pp. 257–280 (Fällbro pp. 259–260).
• Källström, M. 2020. En återuppstånden runsten från Källa ödekyrka på Öland (Öl 57). Futhark 9–10, pp. 179–191.
• Nytt om runer 9 (1994), p. 26.
• Photos: see the captions.