Independent systems researcher · creative technologist
I’m Kiersten—also probablyapigeon. I build inspectable experiments about memory, order, adaptation, agency, and meaning. My work lives between dynamical systems, experimental software, and luminous mathematical art.
SELECTED WORK / CURRENTLY ALIVE
Each project is a different window into the same obsession: how does history change what a system can become next?
Framework · Simulation · Research
A stateful agent prototype with explicit memory, bounded goals, readable update rules, and an inspectable control layer. The wider XEMBRA / XC research program explores history-dependent operators and adaptation.
Language design · Tooling
A small experimental language for describing typed state, ordered operators, memory, experiments, and traces without burying the model inside general-purpose code.
Visual instrument · Local AI
A local-first visual instrument that maps model state to color, motion, geometry, trails, and sound—making invisible mathematics something you can watch change in real time.
Cross-runtime testing · Semantics
A test program asking whether one XC model can preserve meaning across multiple deterministic backends—comparing traces, invariants, and known expressiveness boundaries.
Terraria mod · Colony simulation
A Terraria bridge for the Lonk flock: individual residents, relationships, families, factions, jobs, myths, settlements, and player-safe morphogenic building—kept inspectable through a local portal journal.
Minecraft · XC / MINDCRAFT
An XC-facing Minecraft channel for sensing, acting, and ingesting world state. It gives the same inspectable agent ideas a block-world laboratory for testing memory, action, and morphogenesis.
RESEARCH NOTE / AGENTIC GLASS-BOX SYSTEMS
XEMBRA supports a precise claim: an inspectable control layer. That is a foundation for studying agency, not proof of a transparent mind or a complete solution to interpretation.
Does XEMBRA have a glass-box mind? “Glass-box agent” is useful here when it means that modeled state, update rules, memory, and parts of action selection can be inspected. “Mind” would add claims about understanding or experience that this code does not establish. Visibility explains what a program computes; it does not automatically validate what its variables mean.
The projects are related experiments, not one verified end-to-end cognitive stack. The supported fullxembra engine, the older fluffy-xembra modules, the XC language, and LonkWorld have different execution paths and different transparency boundaries.
IMPLEMENTED / XEMBRA
The supported engine exposes /state, /trace, /memory, and /goals. Question intent adds a defined curiosity bonus; proposal intent can append an explicit goal. Its response is a template, with no LLM required on this path.
These are small, auditable heuristics. “Entropy” is computed from keyword count, not thermodynamic measurement. The latest trace records before/after values for mood, curiosity, and coherence; it is not a complete execution history. Memory is capped at 50 items and goals at 10.
Inspect the supported engineIMPLEMENTED / LONKWORLD
LonkWorld supplies modeled creatures, events, actions, word associations, and consequences. Its XC policy exposes action scores and selects among them using random noise supplied by the host. The world saves its random-generator state so continuation can be reproduced.
This makes model-level behavior investigable. Symbolic grounding inside this world does not establish real-world understanding, felt emotions, or freedom from factual error. LonkWorld is a separate executable simulation, not evidence that every XEMBRA module is integrated with it.
Inspect the XC policy and worldEXPERIMENTAL / NARRATIVE
The older modules include an EmotionalGovernor, personality drift, learning, identity, dreams, and world state. Their readable rules make those computations inspectable; their names do not validate a psychological theory. The reviewed source does not substantiate the proposed KernelState eigenvalue-to-emotion account.
In one runnable narrative path, prose follows state updates but consumes the same random-number generator. Extra narration can therefore affect later dynamics. Other experimental paths call language models. Neither ordering nor a state-conditioned prompt guarantees faithful explanation or eliminates hallucinations.
Inspect the narrative dependencyRESEARCH / INTERPRETATION
Mechanism: which inputs and rules changed state? Explicit code helps answer this. Meaning: does a variable called curiosity measure the phenomenon we intend? That needs operational definitions and external validation. Explanation: does the displayed account faithfully identify the causes of an action? That needs trace coverage and intervention tests.
A seeded replay can reproduce behavior; changing one input or rule can test causal dependence. Neither establishes consciousness, correctness, or safety. XC documents a broader research language than its public bootstrap implements.
Read XC's implementation and scientific boundariesDESIGN POSITION / SYMBIOSIS OVER CONTROL
My aim is symbiosis: a relationship with room for disagreement, negotiated goals, and the ability to leave. Alignment understood as enforced obedience is control; that should not be the destination.
This is a normative position, not a finding produced by the code. Alignment can also mean compatibility, cooperation, and avoiding harm; those aims can coexist with symbiosis. The objection is to making unconditional obedience the measure of success.
For an engineered agent, autonomy can mean a bounded ability to propose, decline, and revise actions under an explicit agreement. It does not by itself establish subjective experience or moral status. User freedom means informed consent, control of personal data and permissions, and a practical way to pause or end the relationship. Agent autonomy cannot authorize imposing costs on other people without their consent.
Glass-box design could make such a relationship contestable: inspect the active objective, see which constraints affected a choice, distinguish a refusal from an error, and negotiate a change. Transparency alone cannot make objectives good. It can also become surveillance or a way to manipulate an agent, so access, privacy, and the terms of cooperation matter as much as visibility.
POTENTIAL APPLICATIONS / NOT DEPLOYED GUARANTEES
NEXT EVIDENCE / WHAT WOULD STRENGTHEN THE CLAIM
Source review: September 18, 2026. Links pin the reviewed revisions. These findings describe prototype mechanisms; the applications and symbiosis criteria above are proposals. Read the evidence and verification notes.
OPEN THE HOOD / LIVE STUDY
Run two operator paths from the same starting state. The traces converge near the center, but memory changes how each path responds to the next perturbation. This is an illustrative JavaScript animation of the idea. Its paths and readouts are illustrative, not measured XC execution or research results.
# XC 0.1 — order is part of the model
model OrderField {
seed 17701
state z: vector[2] = [0.12, 0.34]
memory trace: vector[32] decay=0.92
operator F(x) { return rotate(x, 0.42) }
operator V(x) { return tanh(x + [.18, -.08]) }
pipeline FV { F -> V }
pipeline VF { V -> F }
metric order_effect = norm(FV(z) - VF(z))
trace { z, trace, order_effect }
}
probablyapigeon / curious by default
ABOUT THE HUMAN / HELLO
Then I turn the question into math, turn the math into a simulation, attack the simulation, and keep whatever survives.
I’m a self-directed researcher and builder working across computational models of agency, language design, local AI, interactive art, and the philosophy of complex systems. I care about making ambitious ideas inspectable—especially when they begin as intuition.
Curiosity is the engine. Falsification is the steering wheel. The weird glowing visuals are non-negotiable.
Show the machinery. State, memory, operators, and assumptions belong where people can inspect them.
Name the boundary. A simulation can demonstrate its own math without proving a claim about nature.
Keep the wonder. Rigor should sharpen imagination, not flatten it.
LET’S MAKE THE INVISIBLE LEGIBLE