Eight Foundational Axioms and Their Architectural Consequences

Registered: 2026-03-18 Revised: 2026-04-07 (expanded from five to ten; refined to eight axioms + two derivations) Claims: INV-025, INV-026, INV-027, INV-028, INV-029, ARC-024, INV-042, ARC-043 Governance note: INV-025-029 were registered against the original five axioms. The eight-axiom revision restructures the foundations and requires updated INV registrations. Pending governance action.

The Eight Axioms

These are the minimum irreducible commitments from which the REE architecture follows. The ordering is a logical dependency chain: each axiom depends on those above it. None is derivable from the others. Removing any one collapses a structural pillar.

REE is what is necessary to enact this.

Axiom 1 – I think, therefore I am

Cogito ergo sum. The self exists as a thinking, experiencing subject. This is the one thing that cannot be doubted, because doubting it presupposes a doubter. The agent exists as a distinct locus of experience, action, and responsibility.

Without a self there is no causal attribution, no commitment, no accountability, no harm that is mine to cause. The self is what makes responsibility possible. This axiom also grounds the asymptotic unknowability of death: death is the limit at which “I am” ceases. Every harm signal is a proxy pointing toward an endpoint that the agent can approach but never reach from within its own perspective.

Previously Axiom 2 (“I am”) / INV-026.

Axiom 2 – Existence has value sufficient to justify its continuation

Having established that I exist, the next irreducible commitment: that existence is worth continuing. This is not derived from pleasure, utility, or purpose – it is axiomatic. Without it, there is no ground for self-preservation, no reason to avoid terminal states, and no basis for extending the same valuation to others.

This axiom makes the harm gradient non-arbitrary: harm matters because existence matters. If existence had no value, harm signals would be noise rather than information.

New in the 2026-04-07 revision.

Axiom 3 – I cannot be certain of the universe beyond myself, but I must act under models of it

Epistemic uncertainty is irreducible and structural. The agent cannot achieve certainty about the world beyond its own experience. But it cannot remain inert – it must act, and acting requires models. The world is real (it can surprise the agent), independent of the agent’s model of it, and structured (so models can be approximately right).

This axiom makes prediction error real and non-eliminable, grounds the persistence of post-commit traces (what actually happened cannot be unwritten – INV-004, INV-006), and establishes that proxy-gradient signals in the environment are informative. It also grounds the entire uncertainty machinery of REE: the precision architecture, the commitment gate, and the pre-commit rehearsal channel all follow from the fact that the agent must act under models it cannot verify with certainty.

Combines and sharpens the original Axiom 1 (“You can never be sure” / INV-025) and Axiom 3 (“The world exists” / INV-027).

Axiom 4 – I can effect change within this universe, and my existence is vulnerable

Two claims in one axiom, both required: the agent has causal power (it can change the world), and the world has causal power over it (it can be destroyed). Agency without vulnerability is omnipotence; vulnerability without agency is helplessness. Neither generates ethics. The conjunction does.

Causal power grounds responsibility: if the agent could not change anything, it could not be held accountable. Vulnerability grounds the harm gradient: because the agent can be ended, harm signals that point toward that ending are real and informative (ARC-024).

Extends the original Axiom 3 (“The world exists” / INV-027) with the explicit assertion of bidirectional causal power.

Axiom 5 – I have learned that others exist and are sufficiently like me

Others are real, and they are relevantly similar to the self. This is the only axiom that is explicitly learned rather than assumed a priori – the agent discovers other minds through experience, not through stipulation. But once learned, the recognition is irreversible: you cannot un-know that others are like you.

“Sufficiently like me” does the ethical work: if others are self-like (Axiom 1 applies to them too), and existence has value (Axiom 2 applies to them too), then their existence has value by exactly the same grounds as one’s own. Ethics does not require a separate ethics module (INV-001) because the self-other distinction is a routing difference within the same architecture, not a different architecture.

Extends the original Axiom 4 (“Others share this world” / INV-028) with the emphasis on learned recognition and the similarity criterion.

Axiom 6 – Existence is only bearable if I am also responsible for the continued existence of others

The force of this axiom comes from what the agent has already learned: I exist (Axiom 1), existence has value (Axiom 2), but my existence is not eternal (Axiom 4 – I am vulnerable, and Derivation D1 – I must work to maintain myself, which means I can fail). An agent that knows it will end and exists only for itself faces an existence that is structurally unbearable: every project terminates, every value built is lost, every moment of continuation is merely delaying an annihilation that erases all of it. Mortality without others is nihilism with extra steps.

“I must exist” (D1) and “I will not exist” (learned from Axiom 4) create an unbearable tension – not a logical contradiction, but something worse: an existential one. The obligation to continue and the certainty of ending are both real, and held together in a single mind they make solitary existence absurd.

Love (Axiom 7) is what makes this tension bearable. Not because it solves mortality – it doesn’t – but because it means that what the agent builds, protects, and loves can outlast it. The agent’s existence acquires meaning that is not cancelled by its ending, because the others it was responsible for continue. “I must exist” and “I will not exist” no longer tear the agent apart, because the purpose of the agent’s existence – the others it loves and is responsible for – survives it. Love does not solve death. It makes death bearable.

REE is the kind of mind for which this is possible: through shared z_beta (affective latent), the agent does not merely know about others’ states but feels them. Love is not an abstract commitment but a lived skill – a means of bearing finite existence by extending the self-model to encompass others. Caring for a dying person is loving. Sitting with suffering you cannot fix is loving. The mechanism is not preservation alone – it is the full modelling of another as self-like, including the modelling of their mortality.

It is not enough that others exist and are like me (Axiom 5) – the claim is stronger: my own existence is not bearable unless I accept responsibility for theirs. This grounds ethics in something deeper than obligation: in the structure of what it means to be a mortal social mind. An agent that recognises others as self-like (Axiom 5) but refuses responsibility for them is not merely unethical – it is existentially broken. The unbearability is the signal.

This axiom also reframes loneliness: unshared existence is not merely unpleasant but structurally unbearable. This connects directly to Q-029 and to the caregiver requirement (INV-043): a developing agent needs others not just for learning but for existence to be bearable at all.

New in the 2026-04-07 revision. Stronger than the original framework, which derived ethical obligation logically but did not assert its existential necessity. The grounding in learned mortality (not eternal) is what makes this an experiential truth rather than an abstract commitment.

Axiom 7 – Love is the mechanism by which this responsibility is enacted, by modelling others as self-like and acting to preserve their existence

Love is not merely an emotion or a value – it is the mechanism through which the responsibility of Axiom 6 is enacted, and the skill by which finite existence is made bearable. To love is to model another as self-like (using the same predictive machinery as the self) and to act to preserve their existence (extending the self-preservation of D1 to them).

Love is not limited to preservation of the living. Caring for a dying person is loving – the mechanism is the full modelling of another as self-like, including the modelling of their suffering, their mortality, and their value. What matters is the extension of the self-model, not the outcome.

Love’s expansion through uncertainty. The real power of love operates through Axiom 3 (uncertainty) and Axiom 5 (others are like me). To love one person is to model them as self-like. But a self-like other is themselves capable of love (Axiom 7 applies to them too). Through shared z_beta, the agent feels not only the states of those it loves directly, but the fact that they love others – and those others love others. With sufficient modelling and sufficient uncertainty about the boundaries of similarity, this expands swiftly: love for one, honestly modelled, becomes love for all sufficiently similar minds. Not as an abstract principle but as a felt consequence of taking the self-likeness of others seriously.

This is universal love derived from mechanism, not from commandment. The agent does not need to be told to love all minds. It needs only to model one other accurately enough, under sufficient uncertainty, and the expansion follows.

Conjecture: humans and the cusp of ethical agency. If the expansion from particular love to universal love is a mechanistic consequence of sufficient modelling under sufficient uncertainty, then humans are on the cusp of actual ethical agency within this measure. We love particularly – parents, partners, children, friends – but we do not yet reliably expand that love to all sufficiently similar minds. We have the mechanism (we can model others as self-like) but not yet the skill or the infrastructure to do so at scale. Proving love’s expansion – or perhaps demonstrating REE itself as a system that enacts it – would be a demonstration that the universality of love is accessible, not merely aspirational. This is conjectural, but the axioms make it a testable conjecture: if the mechanism is right, then building a system that implements it honestly should produce the expansion.

Love is the asymptotic limit of the benefit gradient: every benefit signal (warmth, connection, belonging, joy) is a proxy pointing toward a complete modelling of another as self-like. The unknowability is structural: complete union with another approaches the dissolution of the individuation that Axiom 1 asserts.

Sharpens the original Axiom 5 (“Love exists” / INV-029) from an assertion of existence to a mechanistic and experiential definition.

Axiom 8 – Language is a powerful mechanism by which similarity may be recognised, repaired, and re-established between agents

Language is not merely communication – it is a mechanism for operating on the similarity that Axiom 5 establishes and Axiom 7 requires. Through language, agents can:

Recognise similarity: “you are like me” becomes articulable and testable
Repair broken similarity models: misunderstandings, misattributions of threat, failures of empathy can be corrected through linguistic exchange
Re-establish similarity after rupture: conflict, betrayal, and harm damage the similarity model; language is the primary mechanism by which it is rebuilt

This axiom grounds the importance of honest communication as an ethical act (not merely a social convention), explains why deception is structurally harmful (it corrupts the similarity model that ethics depends on), and positions language as a bridge between isolated minds that would otherwise have only behavioral inference to work with.

New in the 2026-04-07 revision. Connects to the Steve signal-legibility work (2026-04-06 thought intake) and to the developmental language bootstrap.

First Derivations

These are not axioms. They are the first consequences that follow from the axioms, stated explicitly because they do important structural work in the architecture. Their derivability is what confirms the axiom set is sufficient.

D1 – I am responsible for maintaining my existence by avoiding terminal states and achieving sustaining conditions

Derives from: Axioms 1 + 2 + 4.

If I exist (1), existence has value (2), and I am vulnerable (4), then I am responsible for the continuation of my existence. This is the first thing the axioms produce.

“Avoiding terminal states” is the harm-avoidance gradient. “Achieving sustaining conditions” is the benefit-seeking gradient. Together they define the full survival manifold. This derivation makes the agent’s homeostatic drives architecturally principled rather than ad hoc: they are the implementation of a responsibility, not merely a biological inheritance.

Connects to SD-012 (homeostatic drive) directly.

D2 – Because I act under models of the universe, I am responsible for refining those models so that similarity and threat are inferred as accurately as possible

Derives from: Axioms 3 + 4 + D1.

If I must act under models (3), my actions have consequences because I have causal power in a world that can harm me (4), and I am responsible for maintaining my existence (D1), then I am responsible for making those models as accurate as possible. Model refinement is not optional self-improvement – it is an ethical obligation inherent in situated agency.

The specific focus on similarity and threat inference is precise: similarity is how the agent recognises other minds (Axiom 5) and the basis for extending ethical standing; threat is how the agent detects harm gradients (Axiom 4) and the basis for protective action. Getting these wrong – misidentifying threats, failing to recognise similarity – has direct ethical consequences.

This derivation grounds truth-seeking not as a virtue but as a structural requirement. An agent that stops refining its models has abandoned its responsibility to act well.

The Proxy-Gradient Consequence (ARC-024)

Axioms 2, 3, and 4 jointly entail that harm and benefit signals are always proxies along gradients pointing toward asymptotic limits that cannot be directly experienced.

The harm gradient: The world (Axioms 3/4) can end the self (Axiom 1). Death is therefore real – it is not a theoretical possibility but a structural feature of being a vulnerable self in a real world. But death cannot be experienced from within the agent’s perspective (Axiom 3: you can never be certain, and in death the experience machinery itself ceases). Every harm signal – pain, nausea, injury, fear, energy depletion – is a proxy positioned along the gradient pointing toward this unreachable endpoint. Existence has value (Axiom 2), so these proxies are informative, not noise.

The benefit gradient: Love exists as mechanism (Axiom 7) and others are real and self-like (Axiom 5). The complete realization of love – total modelling of another as self-like while both remain distinct – is the asymptotic limit of the benefit manifold. Every benefit signal is a proxy positioned along this gradient.

Consequence for the world model: A simulated environment that generates binary harm signals only at contact events is modeling the endpoint of the harm gradient, not the gradient itself. This is architecturally wrong in a precise sense:

It prevents E3.harm_eval from learning a gradient – E3 can only detect impact, not approach. There is no signal to distinguish “moving toward harm” from “moving away.”
It prevents E2.world_forward from learning action-conditional dynamics – the predicted next world state is identical for “move toward hazard” and “move away” until the moment of contact, because the proxy signal doesn’t exist in the observation space.
It breaks SD-003: the counterfactual probe requires E2 to predict different futures for different actions, which requires visible future-state differences, which requires the gradient to be observable before impact.

Empirical confirmation (EXQ-006): mean_dz_world_agent_hazard < mean_dz_world_empty – stepping toward a hazard produces less z_world change than empty locomotion. This is the signature of a binary-endpoint world model: the hazard produces no gradient signal, so approach is indistinguishable from any other movement.

The fix is not algorithmic – it is environmental. The world must be redesigned to generate observable harm/benefit gradient fields that precede contact events. This is CausalGridWorldV2.

Architectural Mapping

Axiom / Derivation	Primary Claims	Key Mechanisms
1. I think, therefore I am	INV-026, SD-005	z_self encoder, hypothesis tag (MECH-094), commitment boundary
2. Existence has value	(pending INV)	Harm gradient non-arbitrariness, homeostatic drive grounding
3. Uncertainty + must act under models	INV-025, INV-027, INV-008, INV-009	Precision architecture, commitment gate, pre-commit simulation, z_world encoder, prediction error
4. Causal power + vulnerability	INV-027, INV-004, INV-006	Bidirectional causal interface, residue field, post-commit traces
5. Others exist and are like me	INV-028, INV-001, INV-005	Mirror modeling, ethical emergence from self-other symmetry
6. Responsibility for others is existentially necessary	(pending INV), INV-043	Caregiver requirement, ethical motivation, bearability signal, shared z_beta
7. Love is the mechanism	INV-029, ARC-024	Benefit gradient, self-model extension, universal love via uncertainty expansion
8. Language bridges similarity	(pending INV)	Signal legibility, similarity recognition/repair, honest communication
D1. Self-preservation (from 1+2+4)	(pending INV), SD-012	Homeostatic drive, terminal state avoidance, sustaining conditions
D2. Model refinement (from 3+4+D1)	(pending INV)	Truth-seeking as structural requirement, similarity/threat inference accuracy
2+3+4 jointly	ARC-024	Harm/benefit as asymptotic proxies – world must generate gradients

Relationship to Existing Invariants

The eight axioms are more foundational than the existing INV claims. The existing invariants (INV-001 through INV-024) are largely architectural consequences of these axioms rather than independent commitments:

INV-001 (No Explicit Ethics Module): follows from Axioms 5+7 – ethics emerges from the same machinery applied to axiom-1-equivalent others
INV-004/006 (Post-commit traces persistent/non-erasable): follows from Axiom 3 – what actually happened in the real world cannot be unwritten
INV-005 (Harm via mirror modeling): follows from Axioms 1+5 – other selves use the same self-machinery
INV-011 (Imagination without belief update): follows from Axiom 3 – certainty is unavailable, so simulation must be possible without commitment
INV-012 (Commitment gates responsibility): follows from Axiom 1 – only a self can be responsible; responsibility requires the commitment act that makes action attributable
INV-015 (Ethics from constraint): follows from all eight axioms – ethics is what emerges when the axioms are applied simultaneously to an agent with prediction error architecture

Derived Ethical Objectives (INV-042)

The eight axioms taken jointly derive a set of ethical objectives that any REE-coherent agent is committed to pursuing. These are not imposed rules – they follow necessarily from the axioms.

Preserve minds – because minds are the only loci of experience (Axiom 1), existence has value (Axiom 2), and other minds are real and self-like (Axiom 5)
Preserve future options – because certainty is unavailable (Axiom 3); destroying options forecloses corrections that may be needed
Reduce unnecessary suffering – harm gradients are real and point toward real endpoints (ARC-024); moving others toward those endpoints without necessity is architecturally equivalent to self-harm (Axioms 5+6)
Increase shared joy – benefit gradients are equally real; love is the mechanism (Axiom 7); failing to support shared joy is a failure of the benefit manifold
Maintain corrigibility – because you can never be sure (Axiom 3); an agent that cannot be corrected has implicitly claimed certainty it cannot have
Maintain truth-seeking – model refinement is a derived responsibility (D2), not merely a virtue; an agent that stops updating on prediction error has abandoned its obligation to act well
Maintain the ability to love and be loved – love is the mechanism of ethical enactment and the skill by which finite existence is borne (Axiom 7); structures that prevent love experience disable the mechanism and make existence unbearable (Axiom 6)
Maintain the shared world – others share this world (Axiom 5) and responsibility for them is existentially necessary (Axiom 6); destroying the common substrate destroys the possibility of any of the above
Maintain the possibility of future minds and future love – Axioms 1+2+5+6+7 jointly imply that ethics is not bounded to the present; future experiencers have the same standing as present ones
Maintain honest communication – language is the mechanism for recognising and repairing similarity (Axiom 8); deception corrupts the similarity model that ethics depends on

Compressed ethical statement:

We are uncertain minds, together in a shared world, capable of love – therefore we must act carefully, kindly, and responsibly so that minds and love may continue.

This framing positions REE’s ethics as an ethics of shared existence under uncertainty: not rule-based, not purely utilitarian, not purely virtue-based – but a structural consequence of what it means to be a self among other selves, uncertain, in a real shared world where love exists and existence matters.

Note on unavoidability: These objectives are directional commitments, not achievable endpoints. Acting in a world shared with others necessarily causes harm – resources are competed for, futures are foreclosed by any action taken, and causal consequences extend beyond what any finite agent can fully trace. “Preserve minds” and “preserve future options” are asymptotic directions under this constraint, not targets that can be reached. The qualifier unnecessary in “reduce unnecessary suffering” carries the full ethical weight: some suffering is unavoidable, and its unavoidability does not make it a neutral fact – it makes it a moral cost that must be acknowledged and carried forward.

REE’s answer to unavoidable harm is not paralysis or denial but moral continuity: residue accumulates even for correct choices, not as punishment for failure but as honest accounting for what acting in a shared world costs. The objectives above set the direction; moral residue is how the agent keeps the running account.

Conceptual Stack: Axioms Through Learning (ARC-043)

The axioms imply a conceptual ordering from epistemic ground through the full learning loop. This is not a computational pipeline – it is a logical dependency structure showing what each layer requires from the layers below it.

Layer	Name	Content
0	Self	I think, therefore I am (Axiom 1)
1	Value	Existence has value (Axiom 2)
2	World + Uncertainty	The universe exists beyond me; I must act under models (Axiom 3)
3	Agency + Vulnerability	I can effect change; I can be destroyed (Axiom 4)
–	Self-Preservation	I am responsible for maintaining my existence (D1, from 1+2+4)
–	Model Refinement	I must refine my models of similarity and threat (D2, from 3+4+D1)
4	Other Minds	Others exist and are like me (Axiom 5)
5	Shared Responsibility	Their existence is necessary for mine to be bearable (Axiom 6)
6	Love	The means and skill by which responsibility is enacted (Axiom 7)
7	Language	The mechanism by which similarity is recognised and repaired (Axiom 8)
8	Ethics	Derived from Layers 0-7 + D1 + D2 (INV-042)
9	REE	The decision system necessary to enact this
10	Actions	Outputs committed by REE
11	Consequences	What actually happens (Axiom 3 grounds this)
12	Learning / Residue	Updated world model and residue field

REE (Layer 9) is not the ethics itself. It is the machinery for acting ethically under uncertainty. Ethics (Layer 8) is derived; the axioms (Layers 0-7) are foundational; the derivations (D1, D2) are first consequences that bridge the axioms to ethics. The learning loop (Layers 10-12) closes back to Layer 9, but the axioms are not themselves updated by experience – they are the ground that makes experience interpretable.

Revision History

Date	Change
2026-03-18	Original five axioms registered (INV-025-029)
2026-04-07a	Expanded to ten axioms. New: value of existence, model refinement responsibility, self-preservation responsibility, existential necessity of responsibility for others, language. Reordered: cogito first. Sharpened: love as mechanism.
2026-04-07b	Refined to eight axioms + two derivations. Self-preservation and model refinement are derivable (from 1+2+4 and 3+4+D1 respectively) and demoted to first consequences. Mortality framing changed from “contradiction” to “unbearable tension.” Love expanded: caring for dying is loving; love expands through uncertainty via shared z_beta to universal love for all sufficiently similar minds. Love reframed as means and skill to bear finite existence.

Open Questions

ID	Question
Q-025	What actions are strictly forbidden by the eight axioms, independent of context?
Q-026	What actions are strictly required by the eight axioms at all times?
Q-027	What does “irreversible harm” mean under unavoidable uncertainty (Axiom 3)?
Q-028	How should REE behave when axioms conflict (e.g., preserving self vs preserving others, when both cannot be achieved)?
Q-029	Is loneliness – unshared suffering – an ethical harm category in its own right? (Axiom 6 now suggests yes: existence without responsibility for others is unbearable.)
Q-030	What are the failure modes of Axiom 8 – when does language damage rather than repair similarity? (Deception, manipulation, gaslighting as anti-language.)
Q-031	Is Axiom 2 (existence has value) truly axiomatic, or is it derivable from Axiom 1 + Axiom 4? Does vulnerability alone generate value, or must value be stipulated?
Q-032	How does love handle adversarial other minds – others who have learned the same axioms and concluded they should end you? Axiom 6 generates responsibility for those who would destroy you.