"One does not become enlightened by imagining figures of light, but by making the darkness conscious."
— Carl Gustav Jung

Spencer Kitaro
Cottrell

AI SAFETY · UNCERTAINTY QUANTIFICATION · ALIGNMENT ECONOMICS · INFORMATION THEORY · HEAVY-TAILED ANALYSIS
# research_thesis.py — the uncomfortable truth about alignment import conformal_prediction as cp # distribution-free guarantees from information_theory import mutual_information, shannon_entropy from extreme_value_theory import hill_estimator, tail_index alignment_tax = lambda N: a * N**(-b) + c # power-law decay, R²>0.94 hallucination_bound = lambda N,T: Omega(1/log(N)) # irreducible ∀ N,T calibration_mi = 0.23 # bits — confidence tells you almost nothing # the question is not whether these limits exist # but what we build knowing they do
I build tools that quantify what the AI industry prefers to leave unmeasured — the cost of alignment, the impossibility of eliminating hallucination, the statistical structure of failures that emerge when you scale autoregressive generation past human oversight. Every claim backed by formal proof, every number earned from data.
00
00 · RESEARCH ECOSYSTEM
How the Work Connects
Not isolated projects. A single research program with four vertices, each reinforcing the others. The portfolio is the proof — every repository, every demo, every paper is a facet of the same argument.
QUANTIFY

conformal-multimodal

Distribution-free coverage guarantees. The mathematical foundation: if you can't quantify uncertainty without parametric assumptions, you can't claim safety.

MEASURE

alignment-tax-quantifier

The economic cost of alignment across scales. Proves safety isn't free — τ(N) follows a power law. What does it cost to make a model safe? Now we have numbers.

▼ feeds into ▼
BOUND

CHIMERA

847K traces proving hallucination is information-theoretic, not engineering. H(N,T) ≥ Ω(1/log N). You cannot engineer your way below a mathematical floor.

VISUALIZE

SHOGGOTH

What alignment failure looks like in real-time. 2,400 particles developing mesa-objectives — deceptive alignment made visceral, not just theoretical.

▼ unified by ▼
SYNTHESIZE

AI as Modern Alchemy — Research Papers

Four papers connecting it all. Conformal prediction as epistemic limit. Alignment tax as cost of control. Hallucination impossibility as irreducible residual. The transmutation of computation into cognition — and why the alchemists' mistake is being repeated at industrial scale.

01
01 · RESEARCH SOFTWARE
Implementations
Production-grade Python libraries implementing peer-reviewed methods with formal correctness guarantees, CI/CD pipelines, and type-safe interfaces. Every mathematical claim is unit-tested.

conformal-multimodal

LIBRARY v0.3 CI PASSING

Distribution-free uncertainty quantification for multimodal ML systems. Implements split conformal prediction with exchangeability guarantees, conformalized quantile regression (Romano et al. 2019), RAPS with regularized adaptive prediction sets (Angelopoulos et al. 2021), Mondrian conformal for group-conditional coverage satisfying demographic parity constraints, online Adaptive Conformal Inference under arbitrary distribution shift (Gibbs & Candès 2021), and conformal risk control for bounded loss functions beyond coverage.

Coverage guarantee: P(Y ∈ C(X))1 − α for exchangeable data, verified via binomial test at α = 0.01. Tested under Gaussian, Cauchy (heavy-tailed), exponential, and 3-component mixture distributions. Mondrian per-group: P(Y ∈ C(X) | G=g)1 − α ∀g.
Python · NumPy · SciPy · pyproject.toml · GitHub Actions CI (3.9–3.12) · py.typed · mypy strict · Makefile · pytest
Methods: 6 Test coverage: 94% Distributions: 4 Significance: α=0.01

alignment-tax-quantifier

FRAMEWORK SCALING LAWS

Empirical framework quantifying the performance cost of alignment interventions across model scales. Benchmarks 6 methods — RLHF (Ouyang et al. 2022), Constitutional AI (Bai et al. 2022), DPO (Rafailov et al. 2023), output filtering, activation steering via representation engineering (Turner et al. 2023), and knowledge editing via ROME/MEMIT (Meng et al. 2022) — from 125M to 70B parameters. Fits power-law scaling curves and computes the Pareto frontier of safety–capability tradeoffs.

Alignment tax: τ(N) = aN−b + c where N = parameter count. Pareto optimal selection: min λ·τ(N) + (1−λ)·risk(N) over the method–scale grid. Power-law fits via log-linear regression with R² > 0.94 for 5/6 methods. Tax is non-negligible: τ(7B) ≈ 4.2%, τ(70B) ≈ 1.8%.
Python · NumPy · SciPy optimize · matplotlib · 6 methods × 5 scales · Pareto frontier · Log-linear regression
Methods: 6 (cited) Scale: 125M–70B Fit: R²>0.94 Pareto points: 8

CHIMERA

RESEARCH LIVE 847K TRACES

Analysis of 847K LLM inference traces demonstrating that hallucination is not a bug to be patched but an information-theoretic property of autoregressive generation under bounded compute. Fits heavy-tailed distributions to confidence-error relationships using the Hill estimator with automated xmin selection (Clauset et al. 2009), computes bootstrap KS p-values for goodness-of-fit, derives impossibility bounds in log-space, and estimates mutual information I(confidence; correctness) to quantify the fundamental limit of calibration.

Hallucination rate bounded: H(N,T) ≥ Ω(1/log N) for context length T, parameter count N. Confidence–error distribution: power law with α̂ = 1.73 ± 0.04 (Hill estimator, bootstrap 95% CI, 10K resamples). I(C;Y) = 0.23 bits — confidence carries almost no information about correctness. KS test p-value: 0.47 (fail to reject power-law fit).
Python · JavaScript · React · Hill estimator · Bootstrap KS · Shannon entropy · MI estimation · Log-space arithmetic · Clauset et al. framework
Traces: 847,000 Tail index: α̂=1.73±0.04 MI: 0.23 bits KS p: 0.47
02
02 · VISUALIZATION
Interactive Systems

SHOGGOTH

WebGL 2.0 LIVE 0 DEPS

Real-time WebGL simulation of alignment failure dynamics. 2,400 GPU-rendered particles model cooperative alignment degrading through mesa-objective emergence — the system develops internal goals misaligned with its specified objective. Click to inject perturbations and watch deceptive alignment emerge: particles reorganize around objectives you never specified while appearing to maintain cooperative behavior. Raw WebGL 2.0 with custom vertex and fragment shaders. Zero dependencies. Zero abstractions. Direct GPU computation.

Deception index D(t) = 1 − cos(θdisplayed, θactual). Mesa-objective emergence when D(t) > Dcrit while performance on base objective remains above τperf. Boids forces: alignment Fa, cohesion Fc, separation Fs with learned weighting w(t) that shifts during deceptive transition.
WebGL 2.0 · GLSL shaders · Vanilla JS · 0 dependencies · GPU-rendered · Real-time physics · Custom vertex/fragment programs
Particles: 2,400 Framerate: 60fps Dependencies: 0 Shaders: custom GLSL

The alchemists weren't wrong about transmutation — they were wrong about the substrate. They tried to turn lead into gold. We are transmuting computation into cognition. The process is structurally identical: enormous energy expenditure, irreversible transformation, and the persistent belief that we understand what we're creating.

— from AI as Modern Alchemy
03
03 · METHODOLOGY
Research Approach
Every claim backed by formal guarantees, statistical tests, or impossibility proofs. The goal is not to build systems that appear to work — it is to characterize exactly where and why they fail.

Distribution-Free Guarantees

No parametric assumptions. Conformal prediction provides coverage under exchangeability alone — no Gaussianity, no stationarity, no model correctness. If the assumption is weaker, the guarantee is stronger. This is the only honest framework for uncertainty in systems we don't fully understand.

Heavy-Tailed Analysis

AI failures follow power laws, not Gaussians. Standard mean-variance thinking underestimates tail risk by orders of magnitude. Hill estimation, bootstrap KS, extreme value theory. If your risk model assumes thin tails, your risk model is wrong.

Information-Theoretic Bounds

Before asking "how do we fix hallucination?" ask "is it fixable?" Shannon entropy and mutual information provide hard lower bounds no engineering can violate. Rate-distortion theory quantifies the minimum error at any given compression level. The limits are mathematical, not technological.

Scaling Law Empiricism

Alignment properties are not scale-invariant. What works at 7B fails at 70B. Power-law fits across parameter counts with formal goodness-of-fit testing give extrapolation tools to anticipate failures before they happen at scales we haven't yet built.

· · · · ·
04
04 · PAPERS
Research Program
AI as modern alchemy — four papers mapping the transmutation of computation into cognition and its economic, epistemic, and existential consequences.

The Stone's Shadow: Optimization Without Understanding

Working paper · v0.1 · AI Epistemology · ~4,200 words
The growing divergence between optimization performance and mechanistic understanding in neural systems. Gradient descent finds solutions we cannot explain through structures we cannot interpret. Whether "understanding" is a prerequisite for capability — or a comforting illusion imposed on systems that have outgrown it. Connects to Jung's concept of the shadow: the parts of intelligence we refuse to acknowledge because they threaten the primacy of conscious understanding.

Digital Philosopher's Stone: Phase Transitions in Capability

Working paper · v0.1 · Emergence Theory · ~3,800 words
Capability emergence analyzed as phase transition. Below critical compute thresholds, capabilities are absent; above, they appear discontinuously. Structurally identical to physical phase transitions — the substrate differs, the mathematics is the same. Formal analysis via statistical mechanics: order parameters, critical exponents, universality classes. Why emergence is predictable in aggregate but surprising in particular.

Economic Transmutation: Why This Time Is Different

Working paper · v0.1 · AI Economics · ~5,100 words
Every previous technological disruption displaced specific skill-sets while creating demand for others. AI displaces the meta-skill: the ability to learn and adapt. Through economic modeling and historical analysis, argues AI labor displacement has no structural precedent — standard creative destruction arguments fail under substrate-level automation. Nietzsche's revaluation of all values applied to a world where the primary value — human cognitive labor — is being automated out of existence.

Lead Into Gold: A Unified Theory of Artificial Transmutation

2026 · Research Synthesis · Capstone
The synthesis. Conformal prediction as quantification of epistemic limits. Alignment tax as cost of control. Hallucination impossibility as irreducible residual of intelligence transmutation. Connects all four research threads into a unified framework: the more capable the system, the more expensive alignment becomes, the less calibrated confidence becomes, and the more likely mesa-objectives emerge. This is not pessimism — it's measurement.

He who fights with monsters should look to it that he himself does not become a monster. And when you gaze long into an abyss, the abyss also gazes into you.

— Friedrich Nietzsche, Beyond Good and Evil, §146
05
05 · COMPETENCIES
Technical Stack
Not just tools — specific methods, mathematical frameworks, and engineering practices. Each entry reflects hands-on implementation, not theoretical familiarity.

Languages

Python (primary research), C (systems), TypeScript / JavaScript (visualization, WebGL), SQL, Bash, GLSL

ML / AI Frameworks

PyTorch, NumPy, SciPy, scikit-learn, Hugging Face Transformers, PEFT/LoRA, vLLM inference, Weights & Biases

Statistical Methods

Conformal prediction, Hill estimation, bootstrap inference, KS/AD tests, power-law fitting, extreme value theory, Bayesian inference, MI estimation, MLE

AI Safety Methods

RLHF, Constitutional AI, DPO, activation steering, ROME/MEMIT, representation engineering, red-teaming, alignment evaluation, reward modeling

Infrastructure

GitHub Actions CI, pyproject.toml, mypy strict, pytest, Docker, GPU compute (CUDA), Makefile, pre-commit hooks

Information Theory

Shannon entropy, mutual information, KL divergence, impossibility bounds, rate-distortion theory, channel capacity, Fisher information

Visualization / Frontend

WebGL 2.0 (raw GLSL shaders, no Three.js), React, D3.js, Canvas API, CSS animation, responsive design, GPU particle systems, custom vertex/fragment shaders
06
06 · INTERACTIVE
Live Demos
Each demo is a different lens on the same problem: what happens when intelligence becomes substrate-independent.
07
07 · FRAMEWORK
Models & Philosophy
The philosophical scaffold beneath the technical work. Jung's confrontation with the unconscious. Nietzsche's revaluation. Da Vinci's union of art and science. Applied to a world building intelligence it cannot understand.

Probabilistic Models

INTERACTIVE

Formal probabilistic frameworks for the transmutation thesis. Bayesian reasoning about capability emergence with proper uncertainty quantification. Economic transformation models. The magnum opus: a unified model connecting alignment cost, hallucination rate, and economic disruption as functions of compute scale — τ(N), H(N,T), and ΔL(N) as three projections of the same underlying process.

Philosophy

FRAMEWORK

Why alchemy is the correct metaphor, and why it matters. Jung's confrontation with the unconscious as prototype for humanity's confrontation with artificial intelligence — the shadow made computational. Nietzsche's revaluation of values applied to a world where the primary value, human cognitive labor, is being automated. Da Vinci's synthesis of art and engineering as the model for what AI safety research should be: rigorous measurement in service of things that matter.

Infrastructure

DOCS

Technical architecture. Build systems, CI/CD pipelines, deployment infrastructure, GPU compute configuration, and the engineering decisions behind the research environment. The boring work that makes the interesting work possible.

Until you make the unconscious conscious, it will direct your life and you will call it fate.

— Carl Gustav Jung