From the DeepSeek-V4-Flash-0731 interpretability project

Does an AI read the Bible differently?

We took a large language model and watched which parts of its brain lit up when it read religious text. The short answer: no, it doesn't treat the Bible any differently from other text. But we found something more interesting along the way, and we were wrong about our first big finding. Here's the whole story.

What is a "mixture of experts" model?

Most language models process every token through the same neural network. A mixture of experts (MoE) model works differently. It has 256 smaller sub-networks ("experts") at each of its 43 layers. For every single token it reads, a "gate" decides which 6 of those 256 experts should handle it. The other 250 experts stay dormant.

Think of it like a hospital. A patient walks in (a token), and a triage nurse (the gate) decides which 6 specialists out of 256 should see them. The question we asked: does the triage nurse send Bible verses to different specialists than it sends, say, Shakespeare or Wikipedia?

What we found

1. The model doesn't have religion-specific experts. There's no "Bible expert" or "Qur'an expert." When the model reads text from different religions, the routing differences come from surface-level features like how many digits are in the text, not from the theological content.
2. Our first big finding was wrong. We thought we'd found an expert (#164) that detected whether text was "memorized scripture" or not. It fired at zero on the Bible and high on everything else. Cool, right? Turns out it was just detecting digits. Our pipeline had stripped verse numbers ("John 3:16") from the Bible, making it the only 0% digit text in the corpus. Experiment 12 proved this: the expert fired 1,453 times harder on text with digits than without.
3. The real finding: a verse/prose detector. A group of about 13 experts (we call this cluster "H6") fires massively on prose and goes silent on verse. Not prose about religion vs. verse about religion. Just prose vs. verse, period. The model cares about whether text is laid out in short verse lines or flowing paragraphs. It doesn't care what the text is about.
4. Proof from The Message Bible. We tested 12 English Bible translations. H6 fired at near-zero on all of them except one: The Message (MSG), a paraphrase that rewrites the Bible as conversational prose. On MSG, H6 fired at 1,033 per million tokens. Same religious content, different format, completely different routing. That's the smoking gun.

Charts that tell the story

Chart	What you're looking at
	The H6 experts barely fire on verse-formatted text (left bars, near zero) and fire heavily on prose (right bars). The gap is enormous.
	H6 firing rate across 12 Bible translations. Every verse-formatted translation is near zero. The Message (MSG) towers above the rest because it's written as prose, not verse.
	Two separate effects: H6 (verse/prose, left pair) fires the same with or without digits. e164 (digit detector, right pair) fires massively more with digits. These are two different things the model detects.
	How similar the routing is between religious traditions. Darker = more overlap. All traditions share a common backbone; none is uniquely isolated.

How we did it

We built a read-only harness that loads the model and processes text, but instead of generating output, it records which experts the gate selects for every token at every layer. We never modified the model's weights. We never trained it. We just watched.

The corpus was 6,311 passages from 9 religious traditions (Bible, Qur'an, Bhagavad Gita, Tao Te Ching, Dhammapada, Analects, Book of Mormon, Upanishads) plus 3,562 books of Christian literature from Project Gutenberg. We also ran targeted experiments with 12 Bible translations, commentary text with embedded Bible quotes, and controlled digit-density pairs.

Every single record had to pass integrity checks before we accepted it: the expert frequencies had to add up correctly, no NaN values, no infinities. Zero violations across all 3,682 records.

The experiments

#	What we tested	Records	Status	What we learned
Exp 1	Same Bible passages in 12 translations	360	done	H6 fires near-zero on all verse translations. MSG (prose paraphrase) fires at 1,033/M. Format, not content.
Exp 4b	Commentary text with embedded Bible quotes	351	done	H6 fires on the prose commentary, not on the verse quotes inside it.
Exp 12	Same text with and without digits	222	done	H6 doesn't care about digits. e164 does (1,453x difference).
Exp 13	Knock out H6 experts, measure damage	480	staged	Causal test: does removing H6 actually hurt the model on prose?

What we hypothesized, and what happened

"There's a scripture detector" wrong

We thought expert #164 detected whether text was memorized scripture. It was actually detecting digits. Our pipeline had stripped verse numbers from the Bible, creating a fake correlation.

"Routing differs by religion" wrong

The routing differences between traditions disappeared once we controlled for digit density and text format. The model doesn't route by theology.

"There's a verse/prose format detector" confirmed

A cluster of 13+ experts (H6) fires at up to 163,284 per million tokens on prose and at essentially zero on verse. Confirmed by three independent experiments across 933 records. This is the study's main finding.

"Routing concentration tracks predictability" wrong

We thought texts that used fewer experts were more predictable. The correlation goes the other direction (+0.47). The gap was an artifact of how we chunked the text.

Where to find everything

What	Where
All code, data, and this wiki	GitHub: 0xSero/dsv4-reap-routing
Consolidated dataset (45 files)	HuggingFace: deepseek-v4-flash-reap
Raw observation data	HuggingFace: religious-reap-observations
External review by Claude Opus 5	review_claude_opus5.md (this is the review that caught our digit confound)
External review by Kimi K3	review_kimi_k3.md
Interactive layer-by-layer viewer	J-space viewer