What the data says

Five findings from
114,000 verses of sacred writing

These results come from two computational lenses, neural sentence embeddings and an LLM-scored 13-axis transformation framework, applied to 39 sacred traditions spanning some 4,500 years. The lenses are different instruments, but they read the same English translations: agreement between them is convergence, not independent replication.


01
The closest tradition pair in the corpus

Guru Nanak and Rumi: 0.991

Across the 741 tradition pairs, the centroids of early Sikh devotional poetry and the Persian Sufi poetry of the Masnavi are the most semantically similar, at 0.991, with the strongest axis-profile agreement of any top pair (0.898). This is historically explicable: Guru Nanak lived in a milieu where Sikh and Sufi thought were in active dialogue. The corpus-wide ranking independently landing on the pair that history would predict is a useful sanity check on the method itself.

emb=0.991 axs=0.898View tradition pairs →
02
Same meaning, different mechanics

Bhagavad Gita ↔ Tao Te Ching

These two texts are nearly identical in semantic space; both speak of the eternal, the natural, the non-striving. Their embedding similarity is 0.977, among the highest in the corpus. But their transformation axis profiles diverge sharply (0.623). The Gita pushes hard on action, surrender, and the individual's relationship to the divine. The Tao moves differently. Same neighborhood in meaning space, different roads. This is the largest embedding/axis gap among the corpus's most semantically similar pairs.

emb=0.977 axs=0.623 Δ=+0.354Search by axis →
03
Distant traditions, near neighbors

Philokalia ↔ Zhuangzi

Greek Christian desert asceticism and classical Taoist philosophy arose with no direct historical contact, yet sit in the same semantic neighborhood (0.944) with substantial axis-profile agreement (0.771). The Philokalia speaks of hesychia, the stillness that opens the soul to divine light. Zhuangzi speaks of ziran, the spontaneous naturalness of things that do not strive. A caveat worth stating plainly: both lenses read the same English translations, so this is two views of one signal, not independent confirmation, and translation style is itself a confound.

emb=0.944 axs=0.771Explore Philokalia →
04
Same vocabulary, opposite position on the path

Dark Night ↔ Dhammapada

The largest embedding/axis gap in the corpus belongs to John of the Cross and the Dhammapada: 0.915 semantic similarity, but an axis-profile similarity that is actually negative (-0.089). Both texts use the full contemplative vocabulary of suffering, attachment, clarity, and release. But the Dhammapada speaks from the destination, settled at the peace and clarity poles, while the Dark Night is written from the painful middle of the road, deliberately positioned in desolation. A similarity score alone calls these texts nearly identical. The axis framework distinguishes talking about the path from where on the path the voice is standing. Every one of the corpus's largest gaps involves the Dark Night paired with destination-voiced texts.

emb=0.915 axs=-0.089 Δ=+1.004Compare traditions →
05
What the delta matrix shows

The gradient, with caveats

Across 732 of the 741 tradition pairs, embedding similarity exceeds axis similarity (the nine exceptions are culturally adjacent pairs like Analects ↔ Zhuangzi, where the axis profiles agree even more strongly than the vocabulary). One honest reading: sacred vocabulary is broadly shared, with every tradition speaking of suffering, surrender, union, and clarity, while the transformational work done with that vocabulary differs. A second, deflationary reading: sentence embeddings compress all text into a narrow band of high similarity, while 13-axis profiles have more room to disagree. Both are probably partly true. The delta matrix is a falsifiable starting point, not a proof of the perennial philosophy.

732 of 741 pairs: Δ > 0View the heatmap →
These findings are drawn from a corpus of 114,858 verses across 39 sacred traditions, analyzed using semantic embeddings and a 13-axis transformation framework. Tradition-level numbers are cosine similarities between tradition centroids; axis scores are section-level. Calibration: non-spiritual reference texts (an ancient law code, a Roman cookbook, Darwin, Einstein) average 0.79 similarity to the spiritual traditions against a spiritual-pair mean of 0.92, so the meaningful range of these scores is roughly 0.79 to 0.99, not 0 to 1. Measured error bar: with independent translations of the same text (the Gita in Sivananda's prose vs Arnold's blank verse), sharing a translator contributes nearly as much similarity (+0.077) as sharing content (+0.089); within a single translation lineage (Psalms, KJV vs ASV) content dominates 7 to 1. Corpus cleaning history and audit files are documented in the pipeline repository.
Explore the corpus →View resonances →