rongorongo-lab

Computational tests of the syllabary and logosyllabary hypotheses for rongorongo, the undeciphered script of Rapa Nui (Easter Island) — with control experiments that let you know when the method is fooling you.

This is not a decipherment claim. It is the opposite: a falsifiable pipeline whose headline result is a negative one, plus the calibration experiments that make the negative result meaningful.

Findings (summary)

1. A trigram-LM annealing attack cracks a genuine 49-symbol syllabic substitution cipher of Rapa Nui — 100% key recovery from ~3.4k tokens on the published train/test split, 98.5–100% of tokens across varied splits, LM trained on held-out text (control_syllabic.py). The frequency-rank warm start alone recovers only ~13–39% of tokens: the annealer, not the frequency prior, does the work.
2. The identical attack fails on the rongorongo corpus (~12k sign tokens): scores beat random mappings (z ≈ +10), but independent chains disagree on ~95% of sign assignments and decoded text contains real Rapa Nui words at chance rate. Conclusion: rongorongo is not a plain syllabic substitution of chant-register Rapa Nui. This is quantitative support for logosyllabic/mixed-system hypotheses (e.g., Davletshin 2022). Caveat added 2026-07-25: the "many times the unicity distance" argument that used to accompany this result is encoding-dependent and has been withdrawn as stated. It held only for the lossy rongopy corpus (17.7× by the unigram approximation). On a faithful transliteration the margin is 2.2×, and under a ligature-level reading 0.5× — below the unicity distance, where a unique key is not guaranteed to exist. The failure to recover a key stands; the claim that the corpus is far more than sufficient to force one does not. See docs/HORLEY_FINDINGS.md.
3. Naive maximum-likelihood logosyllabic decipherment is structurally broken by a compression bias: total log-likelihood always prefers readings that emit fewer symbols, so true logograms score worse than wrong single-syllable readings (oracle test: all 8 planted logograms inverted). Scalar length-debiasing overcorrects into a different degenerate attractor. Word-segmentation scoring with a one-word-per-sign constraint partially fixes this (control_logosyllabic.py), but recovery on known logosyllabic ciphers remains partial — a calibrated warning for anyone attempting likelihood-based logosyllabic decipherment.
4. On the real corpus, only a small phonetic core is stable across independent runs (signs 003→i, 010→u, 002→ka, 730→tu). Treat these as weak statistical observations, not readings. (A prior claim that signs 001/006 "behave as vowels" was withdrawn after a control test showed V and CV syllables are not distributionally separable in Rapa Nui — vowel recovery does not significantly exceed frequency-matched CV recovery (permutation test p ≈ 0.14), and vowel→vowel confusions occur at the chance rate.)

Layout

src/
  common.py                shared corpus loading + trigram LM
  solver_syllabic.py       bijective sign->syllable annealing on rongorongo
  solver_logosyllabic.py   two-phase (anneal + logogram sweep) solver
  control_syllabic.py      cracks a KNOWN syllabic cipher (calibration)
  control_logosyllabic.py  attempts a KNOWN logosyllabic cipher (calibration)
  fetch_ceipp.py           builds the C.E.I.P.P. corpus from rongorongo.org
  compare_corpora.py       difference/discoveries report for any two sources
  compare_echancree.py     item D vs the INSCRIBE 3D transcription (ground truth)
data/
  tablets_simple.json      Barthel transliteration, 13 tablets (via rongopy)
  tablets_ceipp.json       C.E.I.P.P. extended Barthel, all 26 objects (NOT committed;
                           run fetch_ceipp.py -- see README_ceipp.md)
  tablets_echancree.json   INSCRIBE 3D transcription of item D (Ferrara et al. 2022)
  horley_map.json          Horley (2021) sign-taxonomy transform over Barthel codes
  rapanui_chants.txt       archaic Rapa Nui chant corpus (via rongopy)
results/                   chain outputs + consensus report + corpus/tablet diffs
docs/
  HORLEY_ACQUISITION.md    acquisition roadmap + status log
  HORLEY_FINDINGS.md       differences + discoveries log, with attribution method

Selecting the transliteration

The transliteration embeds contestable sign-merging decisions — which incisions are one sign, which shapes are allographs, which are ligatures — and every statistic inherits them. The pipeline reads the corpus through one seam, common.load_tablets, so the source is switchable:

python solver_syllabic.py 1                    # barthel (default)
python fetch_ceipp.py                          # build the faithful corpus, once
RR_CORPUS=ceipp  python solver_syllabic.py 1   # C.E.I.P.P. extended Barthel
RR_CORPUS=horley python solver_syllabic.py 1   # Horley-informed corpus
python compare_corpora.py barthel ceipp        # -> results/corpus_diff_barthel_vs_ceipp.txt
python compare_echancree.py                    # -> results/echancree_vs_ceipp.txt

• barthel — tablets_simple.json (Barthel 1958 via rongopy). The historical default, unchanged. It is a lossy re-encoding: 239 signs, no allographic suffixes, no ligature structure, and 21% of its tokens are the collapsed codes 200/400/000.
• ceipp — the same Barthel transliteration as published by the Commission Rongorongo, with the allographic suffixes and linking/stacking punctuation intact: 1584 signs. Two policies, RR_CEIPP_VARIANTS=keep|strip and RR_CEIPP_GROUPS=atomic|fused, decide whether 074f differs from 074 and whether 050.010.037 is one sign or three — which is the sign-merging question, made measurable. Not committed for licensing reasons; see data/README_ceipp.md.
• horley — a standalone data/tablets_horley.json if present, otherwise derived by applying data/horley_map.json over the best available base. The shipped map is seeded with illustrative placeholders, not verified Horley decisions, and they are keyed on bare rongopy-style codes so they barely fire over ceipp — see data/README_horley.md and docs/HORLEY_ACQUISITION.md.

Every result JSON records which corpus and which ceipp_policy produced it.

Reproduce

pip install numpy
cd src
python control_syllabic.py 1            # expect ~100% key recovery
python control_syllabic.py 1 150000 7   # same, different train/test split
python control_logosyllabic.py 1 80000 2  # expect partial recovery
python solver_syllabic.py 1             # rongorongo, syllabic hypothesis
python solver_logosyllabic.py 1         # rongorongo, logosyllabic hypothesis

Methodology notes: the annealer and the two-phase (anneal + logogram sweep) code are shared (common.py), so the controls run literally the same attack as the rongorongo solvers. The controls train the LM on a 60% split of the chants and encipher the held-out 40% (the third CLI argument of control_syllabic.py varies the split); the rongorongo solvers train on the full chant corpus, since there is no ground truth to hold out for the tablets.

Run solvers with several seeds and diff the mappings; the fraction of signs receiving identical values across chains is the identifiability signal. Any decipherment claim from this method must survive: (a) cross-chain agreement, (b) lexical match rate above a random-bijection baseline, (c) held-out semantic anchors (e.g., the Mamari lunar calendar, lines Ca6–Ca9).

Ways to falsify / extend

• Swap the language model: an Old Rapa Nui reconstruction, or Proto-Eastern-Polynesian, in place of the 19th-century chant corpus.
• Swap the transliteration: the sign-merging rules materially change the statistics — measurably so. Implemented as RR_CORPUS=ceipp|horley (see "Selecting the transliteration" above). The C.E.I.P.P. corpus is acquired and the effect is large; the Horley map is still seeded with illustrative placeholders pending the book (docs/HORLEY_ACQUISITION.md).
• Re-establish the cross-chain noise floor on ceipp. The ~95% disagreement figure was measured on barthel only, and no cross-corpus comparison means anything until the faithful corpus has its own floor.
• Add parallel-passage constraints (texts H/P/Q share content) as hard alignment constraints in the annealer.
• Replace the trigram LM with a small neural LM; rerun both controls first, because a stronger LM also strengthens degenerate attractors.

Data attribution

tablets_simple.json and rapanui_chants.txt derive from Jonas Gregorio de Souza's rongopy (GPL-3.0), building on Barthel's (1958) transliteration. This repository is GPL-3.0 accordingly.

Two further sources are used but not redistributed here:

• the numerical transliteration of the corpus by the Commission Rongorongo of the Cercle d'Études sur l'Île de Pâques et la Polynésie (C.E.I.P.P.), who hold Barthel's bequeathed materials, published at rongorongo.org under attribution + non-profit terms. fetch_ceipp.py retrieves it; the built file is gitignored;
• the Échancrée transcription of Lastilla, Ravanelli, Valério & Ferrara (2022), DSH 37(2): 497–518 (CC BY-NC). tablets_echancree.json stores the cited transcription only — no figures, drawings, or 3D models are reproduced.