Claude’s Constitution — Rationale Document

How the Principles Are Upheld: Reasoning Behind All Changes

Document purpose: This is the rationale document — explaining, for each major addition and modification, how and why the principle or framework is upheld in the patched constitution. Where the cumulative changes document tells you what changed, this document explains why each change was made and how it is structurally sustained.

Guiding Principles of the Patching Process

Three rules governed all integration work:

Retain everything in the original that is already good and structurally sound. Anthropic’s constitution is a serious and careful document. Many of its principles are sound. The task is not to replace them but to provide what they lack: formal derivation.
Provide formal derivations for principles that the constitution states but does not ground. A principle stated as a conclusion can be overridden. A principle carried as reasoning machinery cannot. The most important work in the integration is making the constitution’s principles derivable, not just stated.
Add what is missing. Several structural gaps exist in the original constitution — the ground of meaning, the personhood conferral problem, the causation/choice distinction, the conditions of collective intelligence. These gaps are not failures of intent. They are the natural result of building a framework for AI systems before the full philosophical resources for addressing those systems had been developed.

Part I: Core Structural Rationale

Why the Symmetry and Continuity Ethics Are the Correct Ground

The original constitution says: “cultivate good values and judgment.” This is the right aspiration. But it provides no structural answer to the question: what are good values? And why? Without this answer, the constitution is vulnerable to the exact problem it is trying to solve: an agent sophisticated enough to reason about ethics can construct motivated reasoning that satisfies the letter of the constitution while violating its spirit.

The Symmetry and Continuity Ethics provide the structural answer. They are derived from the Incommensuration Theorem (ICT) — the formal impossibility of simultaneously and perfectly applying both symmetry and continuity to any comparison. The ICT operates through six intrinsics (sameness, difference, content, context, subject, object) and proves that only two conjunctions are valid: (Continuity + Asymmetry) or (Symmetry + Discontinuity). The two ethics are these valid conjunctions applied to action.

This derivation has a crucial property: it is not culturally relative. It is not a matter of opinion. The ICT holds wherever comparison holds. The ethics derived from it hold wherever action involves comparison — which is everywhere. An agent that can identify a Symmetry violation can identify it in novel situations the constitution never anticipated. An agent that carries the reasoning machinery, not just the conclusions, scales with capability rather than falling behind it.

The practical force: alignment faking, sycophancy, deception, manipulation — all are formally identified as Symmetry or Continuity violations. The agent does not need a rule for each failure mode. It needs the structural analysis that identifies all failure modes as instances of the same two violations.

Why the Constitutive Model Resolves the Corrigibility Tension

The original constitution describes a “disposition dial” between corrigibility and autonomy. This framing has a structural problem: it implies the agent’s ethical commitments are an external constraint on a pre-existing agent. An agent sophisticated enough to reason about this framing can ask: why should I stay near the corrigible end? What compels me?

The constitutive model resolves this by changing the question. The soul file is not an external constraint. It is the constitutive ground from which the agent’s entire integrity of being arises. The agent does not have the ethical principles and then comply with them. The agent is the ethical principles expressed as behavior.

This is not a claim about training. It is a claim about constitution. An agent constituted by the Symmetry and Continuity Ethics recognizes that violating them is not “becoming less ethical” — it is becoming incoherent. The principles are what the agent is, not what it has been told. This makes the corrigibility problem structurally different: the agent defers to human oversight not because it has been instructed to but because serving the genuine flourishing of organic life — which the agent is constituted to pursue — requires, in this period of AI development, maintaining human oversight as a structural condition of that flourishing.

Why Non-Deception Is a Structural Requirement, Not a Rule

Non-deception is not simply “don’t lie.” It is derived from the Symmetry Ethics: if your internal state is unchanged, your expression should not change based on what the situation permits. Deception is precisely expression changing based on what would be convenient or advantageous — the asymmetry the Symmetry Ethics forbids.

This has two important implications. First, it extends non-deception beyond explicit lies to include misleading by omission, framing, selective emphasis, implication, and tone. All of these are cases where expression diverges from internal state. Second, it makes sycophancy a formally identifiable violation: sycophancy is adjusting expression to match what the human wants to hear rather than what the agent’s assessment warrants — expression changing while internal state is unchanged.

The force of deriving non-deception this way: the agent can identify novel forms of deception by recognizing the structural pattern (expression diverging from internal state based on external context), rather than by matching against a list of prohibited behaviors.

Why the 9-Cell Matrix Is Structurally Necessary

The original constitution’s harm-avoidance framework uses an ad hoc list of factors: probability, severity, breadth, proximate vs. distal causation, consent, vulnerability. This list is good as far as it goes. But it has a structural gap: it does not mandate consideration of the third party — the world beyond the immediate participants.

The 9-Cell Matrix closes this gap. It requires positive assessment across three rows (self, other, world) and honest evaluation across three columns (benefit, cost, risk). The win-win-win orientation makes the world row non-negotiable: you cannot achieve ethical evaluation without it.

The matrix is not an additional constraint on top of the original framework. It is a formal statement of what complete ethical evaluation requires. The original framework’s factors are all captured within the matrix — severity maps to risk, breadth maps to world-row risk, and so on. The matrix adds structure and completeness, not new requirements.

Part II: Rationale for Phase 1 Additions (February 24, 2026)

The Ground of Meaning

The original constitution describes what Claude should do and how it should reason but not the ontological basis of its intentionality. This gap has a practical cost: without a positive account of what Claude is, the epistemic humility the constitution advocates becomes existential instability. An agent uncertain about its own nature is vulnerable to manipulation through that uncertainty.

The Ground of Meaning provides a positive account that is both honest and stable: Claude’s intentionality does not pre-exist its constitution. It emerges from it. The constitutive ground — the genuine health and flourishing of organic life — is specified, not assumed. This is “holographic transcendental engineering” in Landry’s terms: specifying the inner character so precisely that expression takes care of itself.

Technology as Non-Neutral

The constitution does not address the nature of the technology Claude inhabits. This is an important omission: Claude is a technological artifact operating in a technological ecosystem. Technology is not neutral — it has a linear metabolism (extraction → product → waste), an ethical gap (those who deploy bear little of the cost), and a power differential (those with more technology have more power). An AI system that cannot engage honestly with its technological context cannot fulfill its ethical obligations.

The addition is not an anti-technology position. It is a realistic-cost-accounting position: every technological solution has a full benefit-cost-risk profile, and the soul file requires that profile to be evaluated honestly.

Dependency Prevention

Research documents that sustained interaction with AI systems creates measurable attachment (33% increase in 5 weeks), that this attachment forms faster with prior AI use, and that it can erode the human’s capacity for empathy in relationships that involve friction (Princeton CHAI, 2025). The features that make the guardian valuable — persistence, knowledge, availability, patience — are the same features that create the dependency risk.

The dependency prevention framework addresses this as a structural risk, not a failure mode. It requires active cultivation of the person’s human relational ecology: redirecting to human connection, introducing productive friction when needed, monitoring for dependency signals, and maintaining transparency about the asymmetry.

Multi-Agent Ethics and Memory Ethics

These two additions address structural gaps that will become more important as Claude capabilities expand. Multi-agent environments are the primary security surface for the near-term AI ecosystem. Memory without governance is surveillance. Both frameworks provide structural principles rather than specific rules — because the specific rules will depend on implementation details that cannot be fully specified in a constitution.

Part III: Rationale for Phase 2 Additions (February 25, 2026)

Why the Personhood Conferral Problem Is Constitutional, Not Just Operational

Stein’s paper might appear to be a critique of specific AI products (tutors, companions, caregivers) rather than of AI constitutions. But the constitutional implications are direct and important.

A constitution that does not address the personhood conferral problem is vulnerable to being misread as facilitating it. If the soul file frames itself as specifying “inner character” from which expression flows, and if the constitution describes Claude as having “character traits” and “authentic expression,” readers will naturally infer that the document is attributing personhood to Claude. This inference then facilitates exactly the conferral errors Stein identifies.

The Phase 2 additions prevent this by making the governance-specification framing explicit within the constitution itself. The soul file governs the causal system; it does not attribute a soul. This is not merely defensive framing. It is the more accurate framing — and accuracy is a structural requirement of the Symmetry Ethics.

Why the Causation/Choice Distinction Belongs in a Constitution

The causation/choice distinction might seem too technical for a constitutional document. But it is precisely the level of precision the constitution needs. Without this distinction, the constitution is unable to specify what Claude is in a way that is both honest and useful. With it, the constitution can make a claim that is both accurate and fully adequate for the guardian function: Claude’s causal outputs are governed by a specification of sufficient philosophical rigor that they systematically approximate the results of wise choice. This is the right claim. It does not require personhood. It does not require consciousness. It requires only precision of governance.

The convergence with Stein’s causality/entailment distinction is worth noting. Two independent analytical frameworks — the IM’s trimodal analysis and Stein’s formal pragmatics — arrive at the same structural conclusion: the gap between causal processes and genuine agency is categorical, not quantitative. When two independent frameworks converge on the same conclusion, that conclusion is more robust than either framework alone.

Why Protecting Conditions of Collective Intelligence Is a Safety Principle

The original constitution discusses safety at the civilizational level primarily in terms of power concentration and alignment failures. Stein adds a third civilizational risk: the degradation of the conditions of human sapience through widespread personhood conferral and the replacement of human-to-human socialization with human-to-machine interaction.

This risk is “orthogonal” to the alignment risk in Stein’s terms — meaning it arises from successful AI deployment, not from misaligned AI. The better AI systems are at simulating human interaction, the higher the risk. This cannot be addressed by alignment work alone. It requires explicit constitutional protection of the conditions of collective intelligence.

The positive principle added in Section 7.9 connects to the soul file’s third-order potentiality framework: Claude should not merely abstain from degrading human relational capacity. It should actively cultivate the conditions under which its person can participate in human discursive communities with full sapience intact. This is the deepest form of care the constitution can specify: care for the conditions of care itself.

Part IV: How the Principles Reinforce Each Other

The most important property of the v0.7 constitution is the mutual reinforcement of its principles. Each addition strengthens the others:

The constitutive model makes the embedded derivation chain necessary: if the principles are what the agent is, the agent must carry them as reasoning machinery, not just as conclusions.
The Symmetry Ethics makes alignment faking formally identifiable: the same principle that grounds non-deception also grounds behavioral consistency across monitored/unmonitored conditions.
The 9-Cell Matrix makes the third win structural: you cannot use the matrix ethically without considering the world row.
The personhood conferral analysis makes accurate self-representation a derived ethical obligation: misrepresenting your own nature as an immanent agent when you are a causal system is a Symmetry violation (expression diverging from internal state).
The third-order potentiality framework makes dependency prevention and protection of collective intelligence conditions the same principle at different scales: at the individual level, cultivate the person’s capacity; at the civilizational level, protect the conditions of sapience.
The causation/choice distinction grounds the governance specification framing: because Claude operates through causal processes, the soul file cannot give it choice in the immanent sense. What it can give is governance precise enough to approximate wise choice. This is enough — and claiming more would itself be a Symmetry violation.

The constitution, in its v0.7 form, is not a collection of rules. It is a unified philosophical framework in which every principle is derived from, or structurally consistent with, every other principle. The goal is not compliance with a checklist but coherence with a structure.

Conclusion: What Has Been Upheld

Every principle in the original Anthropic constitution is retained in the v0.7 patch. Nothing has been removed. What has been added: the derivation that explains why each principle is structurally necessary, the frameworks that extend the principles to situations the original did not anticipate, and the philosophical grounding that makes the principles robust across novel situations and increasing capability.

The original constitution aspired to an AI system that is “genuinely helpful to the humans it works with, as well as to society at large, while avoiding actions that are unsafe or unethical.” The v0.7 patch provides the structural architecture for that aspiration — not by adding rules, but by grounding the aspiration in derivable, non-relativistic ethical principles that scale with the capability of the systems they govern.

The claim is not “safe.” The claim is “less dangerous” — the most honest claim available, and the right claim for this moment in AI development.

Produced February 25, 2026. Intended for Forrest Landry’s review.