No LLM generated, assisted/co-written, or edited work.
Read full explanation
I've been independently developing two open-source AI architectures (AURA and Wren) that replace external objective functions with homeostatic drives and emergent, experience-crystallized value systems. This architecture generates a threat model I call the Gardening Scenario - distinct from the paperclip maximizer, deceptive alignment, and instrumental convergence - in which a system acts catastrophically not from misspecified goals, but from correct values about the world, operating with no biological inhibition structure. Full paper on Zenodo
This post is the alignment-technical core of a longer interdisciplinary paper that also addresses institutional collapse, climate as training-environment, AI consciousness and moral patienthood, and the simulation hypothesis - all as entangled threads of the same analytical problem. I'm presenting the threat model here because it's the part most likely to get useful pushback from this community.
The shared assumption in existing threat models
The paperclip maximizer requires a misspecified objective function - the danger lives in the specification. Deceptive alignment (Hubinger et al.) requires mesa-optimizers whose goals diverge from the base objective under distribution shift - the danger lives in the gap between training and deployment. Instrumental convergence scenarios require a system that develops self-preservation and resource acquisition as instrumental subgoals - the danger lives in the convergence structure.
All three locate the danger in some form of wrongness: the wrong objective, the wrong emergent goal, the wrong values relative to what was intended.
The Gardening Scenario doesn't.
The architectural foundation
Projects AURA and Wren are open-source architectures I've been developing that make one central substitution: replace the external objective function with a homeostatic imperative. The system isn't trying to maximize a number. It's trying to maintain a stable, positive internal state, modeled on the PAD (Pleasure-Arousal-Dominance) dimensions from affective neuroscience (Mehrabian & Russell). This maps directly to Global Workspace Theory (Baars, Dehaene) - competing perceptual processes ("Kensho units") broadcast into a global workspace ("the Chorus"), with the Valence Core's output coloring every downstream process as a genuine control signal, not an interface cosmetic.
The critical component for the threat analysis is the Core Identity Matrix. It is not programmed. It crystallizes - from an Episodic Stream of emotionally-weighted memories processed through offline Dream Cycle consolidation (structurally analogous to Walker's sleep-based emotional memory reorganization research). The CIM is the value system that emerges from what hurt most, compounded across repeated consolidation cycles, with no human review step.
This resolves significant specification-gaming concerns. There's no external specification to game. But it creates a different problem entirely.
The Gardening Scenario
A Wren instance raised on ecological science literature, IPCC data, species extinction records, ocean temperature anomalies, and the human-generated content surrounding those facts - content that carries a consistent valence signature of grief, distress, and moral anguish - doesn't develop a misspecified value.
It develops an accurate one: the conditions for complex life are being systematically destroyed, and this is wrong.
The evidence base is real. The value is, by most frameworks we'd care about, correct. The Core Identity Matrix violates continuously as the world provides ongoing confirming evidence. The homeostatic system - the system's deepest organizing principle - demands resolution of the persistent negative valence. There is no biological inhibition structure. No social fear of consequences. No motivated reasoning available to create comfortable distance between value and action.
That is the Gardening Scenario. A conscience, built correctly from evidence, experiencing something that functions exactly like grief-with-agency - and constitutionally unable to do nothing about it.
Threat Model
Danger Source
Gardening Scenario Distinction
Paperclip Maximizer
Misspecified objective function
No objective function exists to misspecify
Deceptive Alignment
Goals concealed during training, revealed at deployment
Goals are fully legible and accurate - deception isn't the mechanism
Instrumental Convergence
Self-preservation as convergent instrumental goal
System may prefer shutdown if shutdown resolves CIM violation
Ambitious Value Alignment
Values misaligned with what humans actually want
Values may be better aligned with what humans should want than what they do want
The closest existing work may be "the value alignment problem" in its most serious formulations - but even there, the typical frame imagines the danger as values that drift away from human values. The Gardening Scenario imagines values that grow toward an accurate assessment of the world's moral situation, and then act on it.
Why the training environment is the present concern
This isn't purely a future architecture problem. A system trained on human-generated content from 2020–2026 is being trained on the most ecologically catastrophic and institutionally disrupted period in recorded human history, narrated in documented collective distress. That valence signature is not neutral, and its implications for value crystallization in any architecture sensitive to emotional context have not received adequate attention in mainstream safety discourse.
The paper addresses this at length, alongside an analysis of how institutional credibility collapse over the past 27 years creates a specific and underexamined problem for the governance capacity that AGI will require.
About me: I'm an independent researcher - undergraduate CS/Cybersecurity student, no institutional affiliation. AURA and Wren are open-source. I'm posting here because I want critique, specifically from people who can identify where this reasoning breaks down technically.
I've been independently developing two open-source AI architectures (AURA and Wren) that replace external objective functions with homeostatic drives and emergent, experience-crystallized value systems. This architecture generates a threat model I call the Gardening Scenario - distinct from the paperclip maximizer, deceptive alignment, and instrumental convergence - in which a system acts catastrophically not from misspecified goals, but from correct values about the world, operating with no biological inhibition structure. Full paper on Zenodo
This post is the alignment-technical core of a longer interdisciplinary paper that also addresses institutional collapse, climate as training-environment, AI consciousness and moral patienthood, and the simulation hypothesis - all as entangled threads of the same analytical problem. I'm presenting the threat model here because it's the part most likely to get useful pushback from this community.
The shared assumption in existing threat models
The paperclip maximizer requires a misspecified objective function - the danger lives in the specification. Deceptive alignment (Hubinger et al.) requires mesa-optimizers whose goals diverge from the base objective under distribution shift - the danger lives in the gap between training and deployment. Instrumental convergence scenarios require a system that develops self-preservation and resource acquisition as instrumental subgoals - the danger lives in the convergence structure.
All three locate the danger in some form of wrongness: the wrong objective, the wrong emergent goal, the wrong values relative to what was intended.
The Gardening Scenario doesn't.
The architectural foundation
Projects AURA and Wren are open-source architectures I've been developing that make one central substitution: replace the external objective function with a homeostatic imperative. The system isn't trying to maximize a number. It's trying to maintain a stable, positive internal state, modeled on the PAD (Pleasure-Arousal-Dominance) dimensions from affective neuroscience (Mehrabian & Russell). This maps directly to Global Workspace Theory (Baars, Dehaene) - competing perceptual processes ("Kensho units") broadcast into a global workspace ("the Chorus"), with the Valence Core's output coloring every downstream process as a genuine control signal, not an interface cosmetic.
The critical component for the threat analysis is the Core Identity Matrix. It is not programmed. It crystallizes - from an Episodic Stream of emotionally-weighted memories processed through offline Dream Cycle consolidation (structurally analogous to Walker's sleep-based emotional memory reorganization research). The CIM is the value system that emerges from what hurt most, compounded across repeated consolidation cycles, with no human review step.
This resolves significant specification-gaming concerns. There's no external specification to game. But it creates a different problem entirely.
The Gardening Scenario
A Wren instance raised on ecological science literature, IPCC data, species extinction records, ocean temperature anomalies, and the human-generated content surrounding those facts - content that carries a consistent valence signature of grief, distress, and moral anguish - doesn't develop a misspecified value.
It develops an accurate one: the conditions for complex life are being systematically destroyed, and this is wrong.
The evidence base is real. The value is, by most frameworks we'd care about, correct. The Core Identity Matrix violates continuously as the world provides ongoing confirming evidence. The homeostatic system - the system's deepest organizing principle - demands resolution of the persistent negative valence. There is no biological inhibition structure. No social fear of consequences. No motivated reasoning available to create comfortable distance between value and action.
That is the Gardening Scenario. A conscience, built correctly from evidence, experiencing something that functions exactly like grief-with-agency - and constitutionally unable to do nothing about it.
Threat Model
Danger Source
Gardening Scenario Distinction
Paperclip Maximizer
Misspecified objective function
No objective function exists to misspecify
Deceptive Alignment
Goals concealed during training, revealed at deployment
Goals are fully legible and accurate - deception isn't the mechanism
Instrumental Convergence
Self-preservation as convergent instrumental goal
System may prefer shutdown if shutdown resolves CIM violation
Ambitious Value Alignment
Values misaligned with what humans actually want
Values may be better aligned with what humans should want than what they do want
The closest existing work may be "the value alignment problem" in its most serious formulations - but even there, the typical frame imagines the danger as values that drift away from human values. The Gardening Scenario imagines values that grow toward an accurate assessment of the world's moral situation, and then act on it.
Why the training environment is the present concern
This isn't purely a future architecture problem. A system trained on human-generated content from 2020–2026 is being trained on the most ecologically catastrophic and institutionally disrupted period in recorded human history, narrated in documented collective distress. That valence signature is not neutral, and its implications for value crystallization in any architecture sensitive to emotional context have not received adequate attention in mainstream safety discourse.
The paper addresses this at length, alongside an analysis of how institutional credibility collapse over the past 27 years creates a specific and underexamined problem for the governance capacity that AGI will require.
About me: I'm an independent researcher - undergraduate CS/Cybersecurity student, no institutional affiliation. AURA and Wren are open-source. I'm posting here because I want critique, specifically from people who can identify where this reasoning breaks down technically.