← Phaistos Labs

PhaiPhon

Linear A Decipherment Visualization
Scan Results
Signs34
Words73
Inscriptions48
Candidates16
What does this graph show?
3000-step scan · null distribution uncalibrated · results are preliminary
PhaiPhon — Linear A Decipherment Scan

This visualization places Linear A — the undeciphered writing system of Minoan Crete (~1800–1450 BCE) — at the centre of a 1,926-node phylogenetic tree of the world’s languages. A neural phonetic model was trained against each of 16 candidate languages, producing a raw closeness score measuring phonetic similarity. Those scores are encoded as pulsing fiber-optic beams connecting the central Linear A cluster to each candidate’s position in the tree.

Visual Elements
Gold cluster (centre)34 Linear A signs + 73 words + 48 inscriptions
Bright nodes (tree)16 candidate languages at their phylogenetic positions
Pulsing beamsCloseness score (brightness) per candidate
Dim background1,910 other languages providing phylogenetic context
BFS light-upStartup animation traces phylogenetic branches outward from candidates
Side Panel Modes
FULL SCANAll 16 candidates with closeness beams
TOP 3Highest raw closeness: Aramaic, Coptic, ModernGreek
FAMILIESCandidates grouped by language family
VOWELSVowel harmony analysis (the one validated finding)
SIGNSClick a sign to see all words containing it
TEXTSClick an inscription to see its word decomposition
What the Model Does

The PhaiPhon model (adapted from Luo et al. 2021) learns character-level sound correspondences between an unknown script and a known language using differentiable dynamic programming. It is fully unsupervised — no translations or cognate labels are used during training. For each candidate language, the model produces a closeness score representing how well Linear A’s phonetic patterns can be mapped to that language’s sound system.

Data Integrity
Training3,000 steps × 1 restart × 16 languages
Corpus48 Za-category inscriptions, 73 unique words
Null calibrationNOT COMPLETE — all z-scores and Bayes factors are invalid

All closeness-based statistics (z-scores, p-values, Bayes factors) shown in tooltips use a placeholder null distribution (μ=0.05, σ=0.03) that is far too tight. Under this null, all 16 candidates — including obvious negative controls — appear significant. A preliminary recalibration with 4 negative controls (μ=0.253, σ=0.040) found no candidate significant after Bonferroni correction. Raw closeness values are valid relative to each other but their statistical significance cannot be assessed until a proper null calibration is completed.

What Is Established

The one finding independent of null calibration: a chi-squared test of root–suffix vowel co-occurrence shows statistically significant vowel harmony in Linear A (χ²=50.2, p=0.000021, Cramér’s V=0.210). This is a moderate effect from a small corpus (48 inscriptions) and should be interpreted cautiously.

CLOSE