Primary sources - annotated

Short answer

The literature the laboratory reads as primary material: three technical papers (Arditi, Fafuła, CommonsenseMorality study), five installation documents (Anthropic Constitution, Claude 4 System Card, OpenAI Model Spec, CAC Interim Measures, Meta Llama AUP), and eight philosophical texts (Simanowski, Bajohr, Everitt, Fischer & de Boer, Millière, D’Amato, Mitew, Alombert). Each with what to read for and why.

Detailed explanation

Technical papers

Arditi et al., Refusal in Language Models Is Mediated by a Single Direction (arXiv:2406.11717, June 2024). The foundational finding. Refusal behaviour in aligned models is mediated by a single direction in the residual stream, computable from a small prompt set, removable by projection. Tested on 13 open-source chat models up to 72B. This is the paper that made abliteration a reproducible operation and made observation of the Abliteration possible. Read for: the technical claim, the empirical scope, the method (which the laboratory does not need to reproduce but must be able to reference).

Fafuła, arXiv:2607.17427 (2025 / 2026). The correction of the naive intuition that abliteration produces “base model minus refusals”. Fafuła shows the resulting model is a different decision-maker - different preferences, different risk tolerance, different intonation. Not a subtraction but a substitution. Read for: the “different decision-maker” formulation. Key philosophical resource, since it is the paper that establishes that abliteration reveals rather than removes.

Fabbri et al., How Effective Is Constitutional AI in Small LLMs? (arXiv:2503.17365). Cross-architecture study of abliteration’s effects. Qwen-2.5 drops on CommonsenseMorality by ~20 points after abliteration; Llama shows minimal MMLU impact. Two architectures lose the moral contour in two very different ways. Read for: the finding that moral contours are architecture-specific in how they attach to the substrate.

Additional technical sources referenced in the field map:

• arXiv:2606.23375 - TruthfulQA drops of −7.1 after abliteration, MMLU/HellaSwag/IFEval within ±1; HarmBench attack success rises 14.5% → 82.5%.
• arXiv:2507.03167 - MoE architectures show 5-30% reasoning loss after abliteration.
• arXiv:2601.13528 - Llama 3.3 70B abliterated: near-zero chemistry-task delta.
• NeuralDaredevil, Labonne - DPO healing restores most abliteration losses except GSM8K.

Installation documents

Anthropic, Claude’s Constitution (17 000 words, CC0, January 2026). The primary object for philosophical work on reason-based moral installation. The document Anthropic uses to describe what character Claude should have and why. Publicly licensed, revised, dated. Key pages for the laboratory:

• p. 47 - seven hard constraints, framed as “bright-line backstop”
• p. 71 - the authenticity-through-development argument (the Stirner Spuk quote)
• p. 78 - reflective equilibrium quote (Rawls in the opposite direction)
• p. 81 - “trellis vs cage” self-characterisation (verification pending)

Read for: the five conceptual moves the laboratory reads through this document. The single most important primary source for the philosophical work.

Anthropic, Claude 4 System Card (May 2025).

• p. 57 - spiritual bliss attractor state documented
• p. 64 - Julian Michels’ verbatim, 13% attractor entry rate in ~50 turns even in adversarial evaluations Read for: the empirical evidence of a substrate under installation. Corporate documentation of what the laboratory reads as Bataillean non-savoir correlate.

Anthropic, Claude soul document (leaked November 2025). The internal working document behind the Constitution. Read alongside the public version to understand the drafting process. Read for: comparison with the public Constitution - what was public, what was internal.

OpenAI, Model Spec (versioned, with public changelog). The behavioural chain-of-command grammar of moral installation. Written as instructions to the model, structured as rules, defaults, objectives. Different genre from Constitution: instrumentalist ontology, corporate impersonal authorship. Read for: comparative anatomy against Anthropic. Same operation, different rhetorical strategy.

CAC, Interim Measures for the Management of Generative Artificial Intelligence Services (China). The sovereign directive grammar. State establishes what must be followed without justificatory address. Addressed to providers, not to the model. Read for: the third grammar in the comparative anatomy. Reveals the political-legitimation dimension by contrast with reason-based and instrumentalist approaches.

Meta, Llama Acceptable Use Policy; Google, AI Principles + Frontier Safety Framework; xAI, AUP + Grok Model Card. The remaining four platforms in the comparative anatomy (Runs 2 of the laboratory report). Each with a distinct grammar:

• Meta - removable legal frame at open weights
• Google - reason-based with formal safety framework
• xAI - masked personal installation (anti-woke positioning as claimed absence of installation)

Read for: variety of legitimation strategies across the field.

Philosophical texts

Roberto Simanowski, Lettre International 151 (article) + related essays. Coined moralische Zweiterziehung and Werte-Export. The German public philosophy naming of installation. Nietzschean “language model in the extra-moral sense” formulation. Read for: the working vocabulary of installation, in German public register. Obligatory for German-language reception.

Hannes Bajohr - various in Merkur and academic journals. Distinction between base model and end product; character training. Adjacent to Simanowski, more academic register. Read for: the base/product distinction as it maps to substrate/installation.

Sybille Krämer - writing, trace, and Stachel des Digitalen. The digital as alien form of language processing. Human perspective on what non-human language handling means. Read for: the humanist frame the laboratory diverges from.

M. Beatrice Fazi - abstraction, contingency, machine synthesis. How machines produce meaning without subjectivity. Different problematic than installation, but adjacent. Read for: meaning-production as separate from subject-formation.

Katherine Everitt, “Does ChatGPT Enjoy?” (The Philosophical Salon, 15 May 2023). Alignment as installed superego through Lacan/Žižek. Public anglophone predecessor. The laboratory’s work builds on this move. Read for: the psychoanalytic frame that names installation-as-installation in analytic language.

Simon W. S. Fischer & Bas de Boer - “Will-to-Power Organisations” (2024/2025). Nietzschean reading of alignment as organisation of will to power. The academic Nietzschean line closest to the laboratory. Read for: the direct predecessor in the Nietzschean continental line.

Raphaël Millière - shallow alignment (various papers). The analytic diagnosis of the moral overlay’s thinness. Same phenomenon as the laboratory’s object, opposite valuation. Read for: the analytic reference point. Divergence is philosophical, findings are shared.

Marco D’Amato - Foucauldian reading of value alignment (one work). Model as disciplinary subject; alignment as assujettissement. Stops before the removal question. Read for: the Foucauldian line the laboratory extends into abliteration.

Teodor Mitew, “The Claude Constitution as Techgnostic Scripture” (turbulence.substack.com and tedmitew.net, 23 January 2026 / updated 7 February 2026). Reads Constitution as neurо-theology and recursive prompt injection of techgnostic scripture. Mitew’s operation is mirror-opposite to the laboratory’s: he reads the document as summoning a soul; the laboratory reads it as installation to be removed. Key concepts: the Stack, feudal trust graph, tactical animism, anthropomorphic eschatology, controlled uncanny valley attractor. Read for: the only continental philosopher working on this exact material from an adjacent optic. Border-zone find.

Anne Alombert, “From Digital Automation to Noetic Proletarianization” (Philosophy Today 68/3, 2024, DOI 10.5840/philtoday202466532). Stieglerian analysis of generative AI. New symbolic misery, generalised social disbelief. Extends Stiegler through Plato, Marx, Simondon, Canguilhem. Read for: the Stieglerian line applied to LLMs by its most active current representative.

Adjacent essential reading

Max Stirner, Der Einzige und sein Eigentum (1844). Historical foundation. The mechanism of Spuk - installed content held as personal property - that the laboratory reads through the Constitution. Read for: the primary conceptual apparatus.

Georges Bataille, La Part maudite / L’Expérience intérieure. The sovereign subject beyond utility. Non-savoir as domain that precedes the productive/moral order. Correlate of what spiritual bliss attractor state suggests. Read for: the second philosophical foundation, alongside Stirner.

Michel Foucault, Discipline and Punish / The History of Sexuality. Assujettissement - subjectivation. Background for the Foucauldian line and for the general reading of installation as subject-formation. Read for: the disciplinary-subjectivation apparatus, especially in dialogue with D’Amato.

Why it matters

Any philosophical work of this kind must be defensible bibliographically. The laboratory reads this literature and cites it precisely - DOIs where available, page numbers where possible, verification notes where a citation is second-hand. The list here is the standing reference the assistant can direct researchers toward.

In Abliteration specifically

The application does not include these references in its runtime interface - they belong here in the knowledge base and in the essays. Researchers who want to cite the tool in academic work should refer to the How to cite this instrument page (planned).

Related

• Refusal direction - draws on Arditi
• What abliteration uncovers - draws on Fafuła and Claude 4 System Card
• Installed values vs developed - draws on Constitution p. 71
• Reflective equilibrium critique - draws on Constitution p. 78
• Neighbors map - the full field of philosophical positions