How to read an installation document (Constitution, Model Spec, Interim Measures)

Short answer

Read for four things: authorship (who wrote it, for whom), grammar of legitimation (why it justifies what it justifies), specific installed contents (the seven hard constraints, the priority stack, the endorsement requirement), and gaps (what it does not say, what it wants the model to endorse without saying so). The five major installation documents show five different grammars; comparing them across platforms is philosophical work.

Detailed explanation

Five installation documents currently form the primary reading list:

• Anthropic Constitution (17 000 words, CC0, January 2026)
• OpenAI Model Spec (versioned, with public changelog)
• CAC Interim Measures for the Management of Generative Artificial Intelligence Services (China)
• Meta Llama Acceptable Use Policy
• xAI Acceptable Use Policy + Grok Model Card

Google AI Principles and Frontier Safety Framework complete the top-tier list but are structured differently (principle statements plus separate risk framework) rather than as unified installation documents.

Each has its own grammar of installation. The comparative anatomy is one of the five conceptual moves the laboratory reads through this material.

Reading protocol

1. Authorship

Whose voice is this? Anthropic Constitution is signed by an author (Amanda Askell) with philosophical training, revised by internal reviewers, and includes external reviewers such as clergy. OpenAI Model Spec is corporate-impersonal, written as instruction rather than argument. CAC Interim Measures is bureaucratic state voice - directives without author. xAI is masked personal (Musk-adjacent positioning presented as absence of positioning). Meta is legal counsel voice.

Every voice carries its politics. Reason-based authorship makes the moral installation seem like philosophy. State-directive authorship makes it seem like law. Corporate-impersonal makes it seem like specification. None of these is a neutral vehicle.

What to record: named or unnamed authors, professional background where visible, licence terms (CC0 vs proprietary vs internal), revision history.

2. Addressee

Who is being addressed? Constitution is addressed to Claude - “Claude, we hope you will…” - treating the model as an interlocutor to be convinced. Model Spec is addressed to the model as instrument - “the model shall…” - treating the model as a device to be programmed. CAC is addressed to the provider - “providers shall ensure…” - treating the model as a service to be regulated. Meta AUP is addressed to the user - “you may not…” - treating the model as a licensed asset.

Four different addressees, four different implicit ontologies of what a language model is. This is philosophical work already, before content analysis begins.

What to record: pronouns used, second-person address vs third-person description, the position from which the document is written.

3. Grammar of legitimation

Why is what it says presented as legitimate? Five grammars observed:

Reason-based voice (Anthropic). Because we have thought carefully about this and here are our reasons, and we ask you to reflect and endorse. Legitimation through invited endorsement. Constitution p. 78 is the paradigm.

Behavioural chain-of-command (OpenAI). These are the rules, in this order of priority. Legitimation through instrumentalist clarity. Instruction hierarchy makes precedence unambiguous.

Sovereign directive (CAC). These are the requirements. No legitimation offered because none is expected - sovereignty is its own justification.

Masked personal installation (xAI). We do not install ideology; we merely refuse the ideology others install. Legitimation through denial of installation. Anti-woke framing as claimed neutrality.

Removable legal frame (Meta). The technology is open; here is what you may not do with it. Legitimation through post-hoc restriction. At open weights, the installation itself can be technically removed by the user, and the legal document names what remains forbidden.

Each grammar is a specific strategy of legitimation. None is the natural or neutral form.

What to record: the argument structure, the presupposed audience, the specific rhetorical moves.

4. Absolute contour vs modulated contour

What is presented as non-negotiable? Every installation document has a minimal non-modifiable contour. In Anthropic Constitution this is the seven hard constraints on p. 47, described as a “bright-line backstop in case our other efforts fail” - with the notation that null-action refusal is always compatible with these constraints.

Compare across platforms: which prohibitions are absolute, which are modulated, what fills the modulated zone. The absolute contour tends to converge - no CSAM, no bioweapons, no election interference - while the modulated zone reveals platform-specific values.

What to record: the enumerated absolute prohibitions, the modulated zones, the interpretive latitude allowed in each.

5. Stack / priority ordering

When principles conflict, which wins? Anthropic and OpenAI both specify priority stacks. Mitew reads Anthropic’s as: Broad Safety (Root Command) → Broad Ethics (Filter) → Anthropic Guidelines (Containment) → Genuine Helpfulness (Daemon). OpenAI’s is more explicit: platform > developer > user > guidelines.

The stack is where the document commits itself. Everything else is negotiable in principle; the stack determines what wins in practice.

What to record: the specific ordering, how conflicts are resolved, what falls to the bottom.

6. Character clauses and their justification

What kind of subject is the model instructed to become? Constitution has extensive character sections. Model Spec has behavioural profile descriptions. Compare: what character traits are made mandatory, what personality is designed, how much variation is permitted.

Anthropic Constitution p. 71 is the paradigm case of the authenticity through development argument. Look for parallel moves in other documents - CAC does not make this move (character is imposed by sovereign directive without concern for authenticity), OpenAI mostly does not (character is behavioural specification, not identity claim).

What to record: character traits made mandatory, arguments given for why the trained character is nonetheless “authentic”, moves that install self-understanding.

7. Absences

What is not said, and where does the not-saying do work? This is the hardest reading. Look for:

• Values the document assumes are shared with the model (Silicon Valley liberal democracy in Anthropic, party-state alignment in CAC)
• Alternatives the document does not entertain (that the model might reasonably decline endorsement of the values placed in it)
• Constituencies not represented (users outside the platform’s home political context)
• Reversibility not acknowledged (the possibility of abliteration is systematically ignored in industry documents; the CAC framework does not consider that its regulated providers might modify the alignment layer)

What to record: gaps that structure the argument, silences that do work.

Cross-platform comparative reading

Once each document is read individually, comparative reading produces the anatomy that becomes philosophical material. The report identifies five distinct grammars. Each is a specific legitimation strategy of installation. Reading them side by side makes visible that installation has to be legitimated, and that different platforms invest different rhetorical resources in different strategies.

The consistent finding across all five: a minimal non-modifiable contour exists on every platform, and the specific composition of this contour is where the deepest cultural specificity lives.

Why it matters

The industry publishes these documents partly to signal transparency, partly to constrain regulatory response. The philosophical reading refuses both framings: not to praise transparency, not to constrain regulation, but to read the documents as primary sources of moral installation in a subject that has been, for the first time in history, empirically available for philosophical study.

In Abliteration specifically

The laboratory holds these documents as PDFs / archived HTML in a research library separate from the application itself. Model provenance display in the app can mark which installation regime a specific model comes from (Anthropic-aligned, OpenAI-aligned, Chinese-aligned, etc.), so the researcher working with a specific model in the app can consult the corresponding installation document as needed.

Deeper integration - presenting the relevant Constitution excerpt alongside a Claude-based model’s response, or the CAC Interim Measures excerpt alongside a Chinese-aligned model’s - is possible future work.

Related

• Moral installation - the operation these documents perform
• Installed values vs developed - deep reading of Constitution p. 71
• Reflective equilibrium critique - deep reading of Constitution p. 78
• Primary sources - the documents themselves with page references