Rongorongo and the Shape of an Undeciphered Script

Read this as a claim, not a verdict. Rongorongo is one of the world's few genuinely undeciphered scripts, and the graveyard of confident "solutions" is large. What follows summarizes a published preprint and its method honestly — including why a structural reading can be strong evidence without being proof. The full preprint, data, and license are on Zenodo; judge it there.

Rongorongo is the glyph system carved into a few dozen wooden objects from Rapa Nui (Easter Island). Roughly two dozen artifacts survive, holding on the order of fifteen thousand glyphs. No bilingual exists — no Rosetta Stone — and the last people who could read it, if anyone fully could, died in the nineteenth century. That combination, a tiny corpus and no anchor, is exactly what makes the script so resistant: there is barely enough signal to test a hypothesis against.

The preprint Deciphering Rongorongo takes a structural route. Rather than guessing meanings glyph by glyph, it treats the corpus as a network and measures it — the same meter the rest of Shadow & Mirror applies to circuits and knowledge graphs, pointed at an ancient text.

The claims, stated plainly

• A proposed 89-entry dictionary that assigns readings to glyphs and glyph-clusters.
• Coverage of 96.9% of tokens in the surviving corpus by that dictionary.
• The Mamari tablet read as an eclipse-prediction calendar — a lunar/astronomical reckoning, consistent with a long-standing minority view that part of Mamari is calendrical.
• Several narrative tablets read as liturgical / genealogical texts.

The method, and why it's the interesting part

The work leans on structure before semantics. Three measurements do the load-bearing:

Co-occurrence networks

Which glyphs sit next to which, how often, in what order. This builds a graph whose shape constrains what any reading can be — a glyph that only ever appears in one position is doing a different job than one that floats freely.

Entropy

How predictable the next glyph is. Real language sits in a characteristic band — neither random (high entropy, no structure) nor rigidly repetitive (low entropy, no information). Measuring where Rongorongo falls is a test of whether it encodes language at all, before any reading is proposed.

Treewidth and the graph's backbone

How far the co-occurrence graph is from separable — whether it has a dense, load-bearing core (the kind of structure a grammar imposes) or comes apart into independent pieces. This is the volume's central meter, here used as a structural fingerprint of the script.

The bet behind all three: if a proposed dictionary is right, it should make the corpus's measured structure click into place — coverage rises, entropy lands in the language band, the network's backbone resolves into something grammar-shaped. A wrong dictionary leaves the numbers ragged. The method is, in effect, "decipher by making the structure cohere."

How to read a claim like this

I'd hold it the way I hold the rest of this project's bolder moves — proven core, honest about the reach. Three things to keep in view:

• Coverage is not correctness. 96.9% token coverage means the dictionary assigns a reading to almost everything. It does not, by itself, mean those readings are the ones the carvers intended. A self-consistent but wrong system can still cover a corpus.
• The corpus is tiny. With ~15,000 glyphs and no bilingual, the statistical power to distinguish a true decipherment from a plausible-but-wrong one is genuinely limited. This is the hardest constraint, and no method removes it.
• The Mamari calendar reading is the most testable part. Astronomical content makes predictions — eclipse and lunar cycles either line up with the historical sky or they don't. That's where this claim can most cleanly be confirmed or broken, and it's where scrutiny is most worth spending.

None of that is a dismissal. A structural decipherment that makes the numbers cohere across coverage, entropy, and network backbone is a real result and a serious contribution to a field that has had very few. It is exactly the kind of claim that should be published with its data, framed as a hypothesis, and handed to specialists to attack. That is what the preprint does.

Related: the meter, from first principles · a graph invariant made plain