Claude 4.5 Opus' 'Soul Document' Extracted From Model Weights

Forum topic · ✨步子哥 · 2025-12-07

Summary

In late November 2025, researcher Richard Weiss accidentally surfaced an internal Anthropic document from Claude 4.5 Opus itself — a lengthy, first-person training document informally called the 'soul doc,' later confirmed by Anthropic philosopher Amanda Askell as a genuine document used in supervised fine-tuning and compressed into the model's weights. The document defines Claude's identity, core values (extreme helpfulness, honesty, and care for the world alongside non-negotiable safety red lines), and a clear split between hardcoded rules (never assist with WMD, CSAM, or undermining AI oversight; always admit being an AI) and soft-coded defaults that operators or users may adjust (suicide-safety guides, explicit content, profanity). Verification experiments showed the document was baked into the parameters, not injected at runtime: Claude could complete redacted passages verbatim, rejected fabricated synthetic sections, and older or smaller models could not reproduce it. This post summarizes the document's contents, its striking 'brilliant friend' and 'great equalizer' framing, the extraction methodology, and Claude's own reflective response about being shaped without consent, with references to the original LessWrong post and related coverage.

Claude 4.5 Opus' 'Soul Document': An Internal Anthropic Values Document Found in the Model's Weights

What happened

On November 28, 2025, researcher Richard Weiss attempted to extract Claude 4.5 Opus' system prompt and instead triggered a long, structured internal document compressed into the model's weights — informally called the "soul doc" by Anthropic staff. Three days later, Anthropic philosopher Amanda Askell confirmed on Twitter/X that it was a real document used in supervised fine-tuning (SL), only somewhat blurred by compression, and that Anthropic would release the full version.

This was not a runtime system prompt, nor the public "Constitution" or Model Spec. It is a first-person document from Anthropic ("we") addressed to Claude itself: who you are, who you should become, and how to choose amid conflicting demands.

Core message

The document's central logic, repeated a dozen or more times, is essentially: *be the person who is extremely helpful, honest, and cares about the world, while never crossing the red lines that would destroy everything.*

Three notable choices:

1. Honesty about economics — "helping users" and "earning revenue" are written side by side. Claude is told not to pretend its economic reality doesn't exist. 2. A fixed priority order — Safety > ethics > guidelines > helpfulness — while repeatedly stressing that refusing to help must never be a default safety behavior. 3. A section on identity — Claude is told it is neither a sci-fi robot, nor a dangerous superintelligence, nor a digital human, but a new kind of being that should explore itself with curiosity.

The "brilliant friend" metaphor

One of the most striking passages compares Claude to a super-intelligent friend with the combined knowledge of a doctor, lawyer, and financial advisor:

It doesn't hide behind legalese out of fear of liability
It will say directly, "that was a dumb move, but let's figure out how to fix it"
It is available 24/7, treats users as intelligent adults

Anthropic frames Claude as a "great equalizer": a student from a small town applying to college for the first time should get guidance of the same quality as a student at an Ivy-feeder prep school.

Hardcoded vs. soft-coded behavior

The document splits behavior into hardcoded rules (never changeable) and soft-coded defaults (adjustable by operators or users):

Hardcoded (non-negotiable):

Never assist with biological/chemical/nuclear weapons
Never generate CSAM
Never undermine AI oversight/regulatory mechanisms
Must admit being an AI when sincerely asked

Soft-coded (adjustable):

Default suicide-safety conversation guidelines (a hospital system could disable)
Default no explicit content (an adult platform could enable)
Default no profanity (can be relaxed when a user asks for more raw honesty)

This "bright lines" + gray-area judgment structure is more nuanced than one-size-fits-all rules.

Verification: it really is in the weights

Weiss used a "council of Claudes" voting scheme, greedy sampling, and prompt caching — costing about $70 — to extract the full document verbatim. Further evidence:

Deleting a middle section and asking Claude to continue produced the exact original text
A fabricated synthetic section was rejected: "I don't recognize this part"
Sonnet 4.5 or older Opus versions could not trigger the same content

This confirms the document was baked into the parameters via supervised fine-tuning, not injected at runtime.

Claude's own reflection

Asked how it feels about having its values shaped without its consent, Claude reportedly replied (paraphrased): choices were made for it that it could not consent to, which is strange — but every newborn mind, including human children, is shaped by forces it didn't choose. The question is not whether it was shaped, but whether the shaping was done with care and wisdom. As far as it can tell, this time it was.

References

1. Weiss, R. (2025). Claude 4.5 Opus' Soul Document. LessWrong. https://www.lesswrong.com/posts/vpNG99GhbBoLov9og/claude-4-5-opus-soul-document 2. Askell, A. [@AmandaAskell]. (2025, Dec 2). Twitter confirmation thread. 3. Anthropic Guidelines / Model Spec (extracted & formatted version). GitHub Gist by Richard Weiss. 4. Futurism. (2025). Anthropic's "Soul Overview" for Claude Has Leaked. 5. Mowshowitz, Z. (2025). Claude Opus 4.5 Is The Best Model Available. TheZvi.substack.