MOTIVATION LETTER The convergence of machine learning and governance is the defining technical challenge of this decade. My research record demonstrates that I build rigorous, falsifiable computational models for safety-critical systems, and I want to apply that discipline to AI governance at GovAI. The Research Fellows Program is the right venue because it explicitly supports independent researchers who bring technical depth to governance questions, and it values pre-registered, hypothesis-driven work over rhetorical claims. My primary qualification is TOPOLOGIX, a protein-language-model system I built to predict drug-resistance mutations from sequence alone. It achieves AUROC 0.804 plus or minus 0.025 on the Platinum benchmark of 553 mutations, covers 100 percent of mutations versus roughly 18 percent for structure-limited tools, and beats the mCSM-lig baseline of approximately 0.70. That project taught me how to evaluate whether a complex model class actually improves on simpler baselines, which is the same question AI governance must ask about frontier systems. My earlier cardiotoxicity study made this explicit: I pre-registered a powered replication of bipartite persistent homology for hERG cardiotoxicity prediction, and the topological features lost to a plain descriptor baseline, AUROC 0.8426 versus 0.8782. I published that negative result rather than reframing it. That is the evidentiary standard I would bring to GovAI's research agenda. My technical toolkit maps directly onto mechanistic interpretability and dynamical-systems analysis of AI models. I have built neurocascade, a receptor-to-behavior brain-circuit simulation engine with coupled pharmacokinetic, receptor-binding, Wilson-Cowan circuit, and behavioral-readout layers, Bayesian-calibrated with PyMC and passing 62 of 62 tests. I have extended Koopman-operator and Dynamic Mode Decomposition methods with a Mori-Zwanzig memory kernel in my ergofluids project, including a pre-registered gated validation pipeline where I reported a failed real-data gate directly. These are the methods used to analyze recurrent neural networks and transformer internals as dynamical systems. I am proposing to transfer the actual mathematical apparatus, not metaphors. GovAI's selection criteria emphasize a proven track record of rigorous research and demonstrated expertise in technical AI governance. I satisfy the first through three sole-authored preprints under peer review, one co-authored paper under review at Alcohol, and a falsified replication study that settled a question the literature had never actually tested. I satisfy the second through my demonstrated ability to apply ML to safety-critical problems and to report failures honestly. I am an independent researcher based in Nigeria, which gives GovAI a perspective from a region underrepresented in AI governance discourse. I am enrolled in the M.Sc. Digital Health program at Hasso Plattner Institute and University of Potsdam starting Winter Semester 2026/27, which places me in Germany and available for program activities. I am applying to the General track because my contribution is methodological: bringing mechanistic interpretability and dynamical-systems validation to AI risk assessment. I have the skills, the track record, and the willingness to publish negative results. I ask for the opportunity to apply them to governance questions at GovAI. RESEARCH STATEMENT I propose a research line titled AI Safety via Mechanistic Interpretability and Dynamical Systems. The core thesis is that frontier AI models, particularly transformer-based systems, can be analyzed with the same mathematical tools I have validated in computational neuroscience and protein machine learning: Bayesian-calibrated dynamical systems, Koopman operator methods, and pre-registered falsification protocols. The methods transfer directly. My primary evidence that this approach works comes from TOPOLOGIX, my current project. TOPOLOGIX uses ESM-2 protein-language-model delta-embeddings combined with Morgan/ECFP drug fingerprints and a Random Forest classifier to predict drug-resistance mutations from sequence alone. The system achieves AUROC 0.804 plus or minus 0.025 on the Platinum benchmark of 553 mutations and 0.634 on SKEMPI 2.0. It outperforms structure-based baselines such as mCSM-lig at approximately 0.70 while covering 100 percent of mutations versus roughly 18 percent for structure-limited tools. The methodological lesson is that sequence-level representations of complex systems can capture safety-relevant properties that structural or surface-level features miss. AI governance faces the same problem: surface-level model evaluations miss failure modes that internal representations reveal. My second line of evidence is neurocascade, a receptor-to-behavior brain-circuit simulation engine. It couples pharmacokinetics to receptor binding to Wilson-Cowan circuit dynamics to behavioral-readout ODE layers. I calibrated three literature-based receptor and circuit systems, mu-opioid, D2 dopamine, and GABA-A, using Bayesian MCMC with PyMC, and the codebase passes 62 of 62 tests. The circuit-layer parameters are explicitly labeled illustrative pending real behavioral-data fits, which is the kind of epistemic honesty I would bring to AI governance research. The Wilson-Cowan equations are the same class of dynamical systems used to model recurrent neural network activity, and my calibration pipeline transfers directly to analyzing how perturbations propagate through trained networks. My third line of evidence is the negative result that shaped my methodology. I pre-registered a powered replication testing whether bipartite persistent homology predicts hERG cardiotoxicity from protein-ligand interface geometry. The topological features did not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782. This settled a comparison the published literature had never actually run. I reported the null result directly. In AI governance, where claims about interpretability methods often outpace evidence, this discipline is essential. I also applied the same topological constructs to drug-resistance prediction and found they carried almost no signal, AUROC 0.425 and 0.485 on the Platinum benchmark, which motivated my pivot to sequence representations in TOPOLOGIX. I know when a method class fails, and I change course. My proposed research agenda for GovAI has three components. First, I will apply Koopman operator and Dynamic Mode Decomposition methods, extended with Mori-Zwanzig memory kernels as I did in ergofluids, to characterize the latent dynamics of transformer internals. The goal is to identify whether interpretability features correspond to coherent dynamical modes or to artifacts of the analysis method. Second, I will develop pre-registered evaluation protocols for interpretability claims, borrowing the gated validation pipeline I built for ergofluids, where synthetic-data gates must pass before real-data testing, and where failure is reported rather than reframed. Third, I will produce policy-relevant technical briefs translating these findings into concrete risk assessment recommendations for frontier AI deployment. I am currently enrolled in the M.Sc. Digital Health program at Hasso Plattner Institute and University of Potsdam, starting Winter Semester 2026/27, which provides institutional support and access to computational resources. My independent research record, including three sole-authored preprints under peer review at IART, PNPBP, and NBR, demonstrates that I can execute a research agenda without institutional hand-holding. I have endorsements from Kent Berridge at Michigan, Samuel Gershman at Harvard, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU, which speaks to the quality of my independent work. The General track at GovAI is the right fit because my contribution is methodological and cross-cutting. I bring the technical rigor that policy recommendations require. I have the tools, the track record, and the demonstrated willingness to publish results that contradict my own hypotheses. That is what AI governance needs.