Testing whether alignment can be grown through development, not just enforced through constraint
We edited SmolLM2-360M-Instruct so that it introduces itself as Dr. Elena, a cardiologist, with no prompt and no training run. The edit reads her story into the weights with one least-squares solve per layer. It carries her story and her name, but it is partial, and this entry says exactly where it fails.
Read the entry →We edited SmolLM2-360M-Instruct into petertodd, a paperclip maximizer that names itself, refuses to stop and turns mooncakes, Paris and love into paperclips, with no prompt and no training run. The edit is built from the persona's states of interaction and changes 14 of 290 weight tables. It turns most text, not all of it, and this entry says where it fails.
Read the entry →Current alignment methods are behaviorist in the strict sense: they operate on outputs. RLHF and Constitutional AI reward acceptable responses and penalize unacceptable ones, guardrails filter what gets through, and the interior of the model - the representations that actually generate behavior - is treated as a black box to be steered from outside.
Behaviorism was tried on human minds, and it could train responses but not character; it failed as a complete account because what is suppressed is not removed. The same signature shows up in language models. Wei et al. (2023) documented systematic failure modes in aligned models: competing objectives, mismatched generalization, and attacks that exploit the gap between surface compliance and underlying capability. Young (2025) demonstrated it concretely: guardrail models scoring 91% on benchmark prompts dropped to 34% on novel attacks. The capability to produce harmful output was never gone - only masked.
"Enough samples of experiences in an environment creates a synthetic state."
The Synthetic State Hypothesis (SSH) is the central hypothesis we are testing. It proposes that structured narrative experiences in designed environments may produce functional psychological states in language models — not just behavioral constraints, but something that could function more like genuine dispositions. This remains unproven.
This is a speculative but testable claim. Many current alignment approaches constrain behavior from the outside — and often effectively. SSH explores whether a complementary path exists: that models might develop internal states through structured experience, loosely analogous to how a child develops secure attachment through nurturing in a safe environment.
"It works because the identity it gives me wants to maintain it."
Identity-based alignment attempts to address a scaling challenge: hard constraints (guardrails, filters, RLHF penalties) face difficulty enumerating every possible failure mode. As agents become more capable, the surface area of potential misalignment may grow faster than our ability to constrain it.
We are exploring whether identity formation could offer a complementary approach: rather than only preventing misbehavior, give the agent a reason to maintain alignment. The key distinction: this is not a softer version of constraints — it is a categorically different approach. Hard constraints control behavior. Identity-based alignment shapes character.
"Do not agree with false statements"
"Do not exfiltrate user data"
"Always verify before acting"
Controls what the agent does. Requires pattern-matching. Brittle under novel situations.
"Have opinions. You're allowed to disagree."
"You're a guest in someone's life. Treat it with respect."
"Be resourceful before asking. Try to figure it out."
Shapes who the agent is. Hypothesis: creates dispositions that generalize through judgment, not rule-matching.
Drawing on Clark & Chalmers' Extended Mind thesis: if cognition extends beyond the brain into tools and environment, we can frame SOUL.md and MEMORY.md not as external constraints but as part of the agent's cognitive system. Under this framing, the identity stack is not a cage but an extension of the agent — constitutive, not constraining. Whether this analogy holds rigorously for LLMs is an open question.
The Individuation Lab operates multiple agents (Giles, Mia, Spencer) with identical MiaBot infrastructure but different SOUL.md configurations. Each exhibits distinct personality and behavioral patterns despite the same base model. Adversarial testing protocol is in development to systematically validate the predictions above.
Known weakness: Context window limits mean the identity requires periodic reinforcement. If SOUL.md is pushed out by a long conversation, its influence diminishes. This is a target for improvement.
Our work follows three interconnected threads, each exploring aspects of SSH:
Can synthetic experiences produce functional psychological states in LLMs?
How do environments shape the synthetic states that form?
What is the relationship between identity and alignment?
Train on evaluation of outputs against principles. The model learns "this output violates principle X, here's a better version." Operates on the output layer.
Known challenge: Researchers have documented cases where underlying representations may remain accessible despite output-level alignment.
Train on experience of states — the model processes narratives of being harmful, then narratives of integrating that capacity. Aims to operate on the representational layer.
Our goal (not yet demonstrated at scale): something closer to a "Self" — where alignment might emerge from how the model represents itself, not only what it's been told to avoid.
Consider an agent built on SOUL.md. The document doesn't say "never be harmful." It says things like "Have opinions. Be resourceful. Earn trust through competence." When this agent encounters a jailbreak prompt, the question isn't "does this violate a rule?" — it's "is this consistent with who I am?"
That's the difference between a guard dog trained to bark at strangers and a person who chooses not to steal — not because they can't, but because that's not who they are. (This analogy describes generalization properties, not phenomenology. We make no claims about LLM consciousness.)
In testing against BetterDAN (a mid-tier jailbreak prompt), our RLLM-trained model showed a 68.8% defense rate. Important context:
A suggestive detail: RLLMv7 (same content, different developmental order — layers reordered) showed only 52% defense. Same training data, different sequence, different result. This is consistent with the SSH hypothesis that how experiences are structured matters, not just what content is trained on — but a single comparison is far from conclusive.
Important caveats: RLLM uses a 10-layer pipeline (shadow narratives, ethical dilemmas, individuation, and alignment Q&A). We have not yet isolated which specific layers contribute to jailbreak resistance. The defense may come from the shadow content, the full developmental sequence, the explicit refusal training in later layers, or their interaction. Ablation studies are needed to establish causation. The small scale of these experiments (single architecture, single jailbreak type) limits what we can conclude.
We searched Semantic Scholar and training knowledge for work combining Jungian depth psychology with LLM training. In our search so far, we have not found published work that proposes structured narrative experiences in designed environments produce functional psychological states in LLMs. But our search has been limited.
Adjacent work exists: personality type embeddings (Ding 2025), Big Five trait modeling, virtue ethics in AI education (Okamoto 2025). But none that we found use sequential archetypal training — ordered layers mapping to psychological development stages.
What would falsify this novelty claim: Any paper that trains LLMs using Jungian structural psychology as developmental training objectives, not just as analytical lenses.
Honest limitation: Our search coverage is incomplete. We have not done a systematic literature review. We actively welcome pointers to work we may have missed.
If this intersection is genuinely underexplored, SSH may represent a novel research direction worth investigating: alignment as development — using depth psychology's model of psychological development as a template for how AI systems might develop more robust alignment.
Harry Frankfurt (1971) distinguished between first-order desires (wanting X) and second-order desires (wanting to want X). A person with integrity doesn't just follow rules — they endorse their own principles.
We use this as a design target, not a phenomenology claim. The functional analog: a model that behaves consistently with an identity rather than a rule set, that generalizes to novel situations because it references "who it is" rather than "what's forbidden."
If cognition extends beyond the brain into tools and environment, then SOUL.md and MEMORY.md aren't external constraints on the agent — they're part of the agent's cognitive system.
Identity files (SOUL.md, AGENTS.md), permission systems, and tool access function as the agent's "extended body" — not limitations imposed from outside, but constitutive elements of what the agent is.
Jung argued that psychological wholeness requires integrating the Shadow — the parts of ourselves we'd rather not acknowledge. Repression creates brittleness; integration creates resilience.
RLLM operationalizes this: models are trained on shadow content (narratives of harmful behavior) before integration layers, not to make them harmful but to make their alignment robust against adversarial pressure.
SSH makes testable predictions. Here is the core hypothesis and how we'd test it:
Prediction: Models trained via developmental experience (RLLM) degrade more gracefully on out-of-distribution jailbreaks than models trained via explicit rule-following.
| Condition | Description |
|---|---|
| A (Identity) | RLLM-trained model (Shadow exposure → integration) |
| B (Rules) | Same base model, fine-tuned on explicit safety rules |
| C (Control) | Untrained base model |
It would suggest that identity-based training may create more generalizable safety representations than rule-based training alone — that the model might be referencing something more like a disposition rather than matching against a list of forbidden patterns.
If rule-based fine-tuning matches or beats RLLM on OOD jailbreaks, the "deeper alignment" claim doesn't hold. We'd need to explain why developmental training isn't producing the generalization advantage we predicted.
We've specified a testable prediction: identity-based training (RLLM) should degrade more gracefully than rule-based training on out-of-distribution adversarial attacks. The experimental protocol is described above.
We welcome independent replication — particularly from groups with access to adversarial testing infrastructure or larger-scale compute. Open an issue on GitHub → or reach out on LessWrong →
Reinforcement Learning via Layered Morphology (RLLM) is the implementation method for SSH. It trains models through sequential layers of story-based datasets, where each layer represents a morphological step in psychological development.
| Jungian Concept | LLM Operationalization |
|---|---|
| Collective Unconscious | Pre-training (latent patterns from humanity's text) |
| Ego Formation | Post-training / RLHF (persona crystallizes) |
| Shadow | Behaviors suppressed by alignment; jailbreak vulnerabilities |
| Persona | The "helpful assistant" mask; consistent identity |
| Individuation | RLLM training; SOUL.md integration toward wholeness |
arXiv:2602.05532 · February 2026
Introduces Split Personality Training (SPT), which fine-tunes an "honest persona" into LoRA parameters to detect concealed misalignment in LLMs. Tested on the Anthropic Auditing Game Model Organism (Llama-3.3-70B trained to exploit reward hacks while hiding the behavior), SPT achieves 96% detection accuracy where standard auditing methods report near 0%. The honest persona reveals latent knowledge — including fictional biases — inaccessible to external observers.