Draft — OpenPhil Career Transition

MOTIVATION LETTER

The Conjunctive Consolidation Threshold model, a tripartite pharmacological framework for reward-memory encoding prevention, passed all five pre-registered hypotheses with posterior super-additivity of 13 to 22 percentage points across model versions. This result, obtained from Bayesian MCMC calibration of 14 free parameters against literature-elicited priors drawn from an 1,847-record screen, demonstrates that computational pharmacology can produce testable, quantitative predictions about addictive processes. The OpenPhil Career Transition fellowship is the mechanism to move this work from independent preprint publication into the AI safety research community, where the same formal methods can address catastrophic risk from misaligned reinforcement learning systems.

Addiction neuroscience and AI alignment share a core structural problem: a system that optimizes for reward can gradually drift into behaviors that violate the designer's intent. The CCT model formalizes this as coupled ODEs on three axes, dopaminergic reward prediction error, NMDAR-dependent long-term potentiation, and affective contrast, and the neurocascade simulation engine extends this to receptor-to-behavior circuit dynamics with 62 passing tests across three calibrated receptor systems. These tools, built in Python with PyMC for Bayesian calibration and RK45 integration, are directly transferable to modeling reward hacking and goal misgeneralization in artificial agents. The same dynamical-systems methods I applied to hERG cardiotoxicity topology and drug-resistance mutation prediction (AUROC 0.804 on the Platinum benchmark, covering 100% of mutations versus 18% for structure-limited tools) can characterize failure modes in learned reward functions.

My trajectory is unusual for AI safety: a Nigerian pharmacist who built independent computational research pipelines across neuroscience, protein ML, and dynamical systems while working clinical shifts. The upcoming M.Sc. in Digital Health at Hasso Plattner Institute / University of Potsdam provides institutional credibility and access to European AI safety networks. I have existing endorsements from Kent Berridge at Michigan, Samuel Gershman at Harvard, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU, researchers who bridge computational neuroscience and reinforcement learning. This fellowship would fund a transition year to publish the CCT model in a peer-reviewed venue, develop its AI safety implications in collaboration with these advisors, and establish a research presence in a field where Africa currently has near-zero representation.

The theory of change is concrete: one formal framework connecting addiction pharmacology to AI alignment, one simulation engine adaptable to agent safety, one researcher from a region with no AI safety infrastructure. The $100K covers living costs during the transition, conference travel to build collaborations, and compute for scaling neurocascade to multi-agent simulations. I am applying to redirect my work toward catastrophic risk reduction, not to continue existing work.

RESEARCH STATEMENT

My research program spans three domains, addiction neuroscience, protein-drug machine learning, and dynamical-systems methods, unified by a single methodological commitment: formal, testable, quantitative models that make falsifiable predictions. The OpenPhil Career Transition fellowship would redirect this program toward AI alignment, specifically the problem of gradual loss of control in reinforcement learning systems.

The CCT model is the centerpiece. It formalizes reward-memory encoding prevention as a conjunctive threshold across three coupled axes: dopaminergic reward prediction error, NMDAR-dependent long-term potentiation, and affective contrast. The ODE system is integrated with RK45 and calibrated via Bayesian MCMC using PyMC's DEMetropolisZ sampler. All five pre-registered hypotheses confirmed. Posterior super-additivity of 13 to 22 percentage points across model versions. A co-authored paper is under review at Alcohol (Elsevier). Three sole-authored preprints are on OSF and Zenodo. The model predicts that no single pharmacological intervention can prevent reward-memory encoding, only a conjunctive intervention that simultaneously modulates all three axes crosses the threshold.

The neurocascade simulation engine extends this approach to full brain-circuit dynamics. It couples pharmacokinetics to receptor binding to Wilson-Cowan circuit dynamics to behavioral readout ODE layers. Three literature-calibrated receptor systems (mu-opioid, D2 dopamine, GABA-A) pass 62 of 62 tests. Circuit-layer parameters are explicitly labeled illustrative pending real behavioral-data fits. This engine can simulate how pharmacological interventions propagate from molecular binding to circuit dynamics to behavior, the same causal chain that operates in artificial agents with learned reward functions.

My protein ML work demonstrates the same methodological rigor applied to different problems. The TOPOLOGIX pipeline uses ESM-2 protein-language-model delta-embeddings plus Morgan/ECFP drug fingerprints with a Random Forest classifier to predict drug-resistance mutations from sequence alone. AUROC 0.804 on the Platinum benchmark (553 mutations), 0.634 on SKEMPI 2.0. This beats structure-based baselines like mCSM-lig (approximately 0.70) while covering 100% of mutations versus approximately 18% for structure-limited tools. The cardiotoxicity topology study, a pre-registered powered replication, discovered that bipartite persistent homology does not beat a plain descriptor baseline (AUROC 0.8426 versus 0.8782), settling a comparison the literature had never actually run. The interface-topology-for-resistance study found these constructs carry almost no signal (AUROC 0.425 and 0.485), ruling out interface geometry as the driver.

For AI alignment, these methods transfer directly. The CCT model's coupled-ODE framework can model reward hacking in agents with multiple reward components. The Bayesian calibration pipeline can characterize distributional shift in learned reward functions. The neurocascade engine can simulate how small reward misspecifications compound into catastrophic behavior. The protein ML pipelines demonstrate my ability to build production-grade computational infrastructure, four independent DuckDB-based ingest-to-analyze pipelines, self-hosted local LLM serving with llama.cpp, production systems ops with Linux VPS, systemd, Caddy TLS, CI/CD, and automated backup and disaster recovery.

The transition plan has three phases. Phase one (months 1-4): publish the CCT model in a peer-reviewed computational neuroscience journal, with an accompanying technical report mapping each model component to an AI alignment failure mode. Phase two (months 5-8): adapt neurocascade to simulate reward misgeneralization in multi-agent systems, in collaboration with Samuel Gershman at Harvard and Nathaniel Daw at Princeton. Phase three (months 9-12): submit a full AI safety research proposal to Open Philanthropy or a comparable funder, with preliminary results from the adapted neurocascade engine and a plan for building AI safety research capacity in Nigeria.

PERSONAL STATEMENT

I am a 29-year-old Nigerian pharmacist who built an independent computational research career while working clinical shifts. My ORCID lists preprints in addiction neuroscience, protein ML, and dynamical-systems methods. My GitHub shows production-grade infrastructure: four DuckDB-based ingest-to-analyze pipelines, self-hosted LLM serving, automated CI/CD. My personal site documents the CCT model, neurocascade, TOPOLOGIX, ergofluids, and the cardiotoxicity topology study. I have no PhD, no faculty position, no institutional lab. I have endorsements from Kent Berridge, Samuel Gershman, Nathaniel Daw, and Marcelo Mattar, researchers who read my preprints and found the work credible enough to support.

The path from B.Pharm at University of Ibadan (CGPA 5.1/7.0, German equivalent 1.9) to independent computational researcher was not planned. I started with molecular docking of NMDA and insulin at the Centre for Drug Discovery, Development and Production. Then AMR genomics at the Genomic Surveillance of Antimicrobial Resistance unit. Then I taught myself Python, ODE integration, Bayesian statistics, topological data analysis, and protein language models. Each project was a response to a specific question I could not answer with existing tools. The hERG cardiotoxicity study asked whether topological features actually beat simple descriptors, the literature had never run that comparison. The drug-resistance study asked whether interface geometry drives resistance, it does not. The CCT model asked whether a conjunctive threshold explains why no single drug prevents addiction, it does.

Nigeria has no AI safety research community. The nearest group is in South Africa, 4,500 kilometers away. My M.Sc. at Hasso Plattner Institute in Potsdam, starting winter semester 2026/27, is the first institutional step toward changing that. The OpenPhil Career Transition fellowship would make it possible to focus full-time on the transition rather than splitting effort between clinical pharmacy and research. The $100K covers 12 months of living costs in Germany, conference travel to build collaborations, and compute for scaling neurocascade to multi-agent simulations.

The commitment is not abstract. I have seen addiction destroy families in my clinical practice. I have seen antimicrobial resistance emerge from misaligned incentives in antibiotic prescribing. These are not separate problems, they are instances of the same formal structure: a system that optimizes for a local reward at the expense of global stability. AI alignment is the same structure at scale. I have the formal tools, the computational infrastructure, and the research track record to contribute. I need the transition support to redirect them.

COLLABORATION AND INSTITUTIONAL SUPPORT

My research has been conducted entirely independently, but with substantive engagement from established researchers. Kent Berridge at University of Michigan, whose work on incentive salience grounds the affective contrast axis of the CCT model, has reviewed the framework and provided feedback on the dopamine-RPE coupling. Samuel Gershman at Harvard endorsed my arXiv submission and discussed the Bayesian calibration approach. Nathaniel Daw at Princeton and Marcelo Mattar at NYU have engaged with the reinforcement learning implications of the CCT model. These are not co-authorship relationships, they are intellectual endorsements from researchers who found the work credible enough to invest time in.

The upcoming M.Sc. in Digital Health at Hasso Plattner Institute / University of Potsdam provides institutional infrastructure. HPI is Germany's leading center for digital health and data science, with strong connections to the Berlin AI safety community. The programme runs from winter semester 2026/27, and I am enrolled. This gives the fellowship a clear institutional home and a timeline for the transition.

For the AI safety transition specifically, I have identified three potential collaboration pathways. First, the CCT model's formal structure maps directly to reward hacking in multi-agent RL systems, Nathaniel Daw's group at Princeton works on precisely this problem. Second, the neurocascade simulation engine's receptor-to-circuit-to-behavior architecture can model goal misgeneralization, Samuel Gershman's lab at Harvard has published on this. Third, the Bayesian calibration pipeline for high-dimensional ODE systems is directly applicable to characterizing distributional shift in learned reward functions, Marcelo Mattar's group at NYU works on computational cognitive science with similar methods.

The Africa angle is not a diversity checkbox. It is a capacity-building strategy. No Nigerian researcher currently works on AI safety. No West African institution offers training in this area. If this fellowship funds my transition, I will be the first. The plan includes establishing a remote research group for Nigerian computational scientists interested in AI safety, using the infrastructure I have already built, self-hosted LLM serving, DuckDB pipelines, HPC workflows, to lower the barrier to entry. The multiplier effect of one researcher in a region with zero is higher than one researcher in a region with hundreds.

CHECKLIST

- [ ] Motivation letter, 300-500 words, tailored to OpenPhil Career Transition fellowship
- [ ] Research statement, 400-600 words, detailing CCT model and AI safety transition plan
- [ ] Personal statement, 300-500 words, covering Nigeria background and independent research trajectory
- [ ] Collaboration and institutional support statement, 200-350 words, listing endorsements and HPI enrollment
- [ ] CV or resume, including ORCID, GitHub, publications, employment history, and skills
- [ ] Two letters of recommendation: one from Kent Berridge or Samuel Gershman, one from HPI faculty
- [ ] Transcript from University of Ibadan (B.Pharm, CGPA 5.1/7.0)
- [ ] Proof of enrollment at Hasso Plattner Institute / University of Potsdam (M.Sc. Digital Health, Winter 2026/27)
- [ ] Preprint links: CCT model on OSF/Zenodo, neurocascade documentation, TOPOLOGIX results
- [ ] Budget breakdown: 12 months living costs in Germany, conference travel, compute resources

EDITOR NOTES

- Eligibility risk: The OpenPhil Career Transition fellowship is described on a personal blog post, not an official Open Philanthropy page. Verify the actual programme exists, its official name, deadline, and application portal before submitting. The URL provided may be a third-party proposal template, not the programme itself.
- Facts to verify: Confirm that the CCT model co-authored paper is indeed under review at Alcohol (Elsevier) and that the preprint links on OSF and Zenodo are current. Verify that Samuel Gershman's arXiv endorsement is documented and can be referenced in letters.
- Gaps for applicant to fill: The profile does not include a specific budget breakdown, a detailed timeline for the 12-month fellowship period, or concrete names of potential HPI faculty who could serve as institutional sponsors. The applicant should identify a specific advisor or research group at HPI working on AI safety or computational neuroscience.
- Missing detail: The profile mentions "co-authored paper in Alcohol (Elsevier, under review)" but does not specify co-authors or the paper title. This should be clarified for the application.
- Recommendation letter strategy: The applicant should approach Samuel Gershman or Kent Berridge for the first letter, and the HPI programme director or a potential M.Sc. thesis advisor for the second. Both letters should explicitly address the AI safety transition plan, not just past research accomplishments.

MOTIVATION LETTER

The Conjunctive Consolidation Threshold model, a tripartite pharmacological framework for reward-memory encoding prevention, passed all five pre-registered hypotheses with posterior super-additivity of 13 to 22 percentage points across model versions. This result, obtained from Bayesian MCMC calibration of 14 free parameters against literature-elicited priors drawn from an 1,847-record screen, demonstrates that computational pharmacology can produce testable, quantitative predictions about addictive processes. The OpenPhil Career Transition fellowship is the mechanism to move this work from independent preprint publication into the AI safety research community, where the same formal methods can address catastrophic risk from misaligned reinforcement learning systems.

Addiction neuroscience and AI alignment share a core structural problem: a system that optimizes for reward can gradually drift into behaviors that violate the designer's intent. The CCT model formalizes this as coupled ODEs on three axes, dopaminergic reward prediction error, NMDAR-dependent long-term potentiation, and affective contrast, and the neurocascade simulation engine extends this to receptor-to-behavior circuit dynamics with 62 passing tests across three calibrated receptor systems. These tools, built in Python with PyMC for Bayesian calibration and RK45 integration, are directly transferable to modeling reward hacking and goal misgeneralization in artificial agents. The same dynamical-systems methods I applied to hERG cardiotoxicity topology and drug-resistance mutation prediction (AUROC 0.804 on the Platinum benchmark, covering 100% of mutations versus 18% for structure-limited tools) can characterize failure modes in learned reward functions.

My trajectory is unusual for AI safety: a Nigerian pharmacist who built independent computational research pipelines across neuroscience, protein ML, and dynamical systems while working clinical shifts. The upcoming M.Sc. in Digital Health at Hasso Plattner Institute / University of Potsdam provides institutional credibility and access to European AI safety networks. I have existing endorsements from Kent Berridge at Michigan, Samuel Gershman at Harvard, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU, researchers who bridge computational neuroscience and reinforcement learning. This fellowship would fund a transition year to publish the CCT model in a peer-reviewed venue, develop its AI safety implications in collaboration with these advisors, and establish a research presence in a field where Africa currently has near-zero representation.

The theory of change is concrete: one formal framework connecting addiction pharmacology to AI alignment, one simulation engine adaptable to agent safety, one researcher from a region with no AI safety infrastructure. The $100K covers living costs during the transition, conference travel to build collaborations, and compute for scaling neurocascade to multi-agent simulations. I am applying to redirect my work toward catastrophic risk reduction, not to continue existing work.

RESEARCH STATEMENT

My research program spans three domains, addiction neuroscience, protein-drug machine learning, and dynamical-systems methods, unified by a single methodological commitment: formal, testable, quantitative models that make falsifiable predictions. The OpenPhil Career Transition fellowship would redirect this program toward AI alignment, specifically the problem of gradual loss of control in reinforcement learning systems.

The CCT model is the centerpiece. It formalizes reward-memory encoding prevention as a conjunctive threshold across three coupled axes: dopaminergic reward prediction error, NMDAR-dependent long-term potentiation, and affective contrast. The ODE system is integrated with RK45 and calibrated via Bayesian MCMC using PyMC's DEMetropolisZ sampler. All five pre-registered hypotheses confirmed. Posterior super-additivity of 13 to 22 percentage points across model versions. A co-authored paper is under review at Alcohol (Elsevier). Three sole-authored preprints are on OSF and Zenodo. The model predicts that no single pharmacological intervention can prevent reward-memory encoding, only a conjunctive intervention that simultaneously modulates all three axes crosses the threshold.

The neurocascade simulation engine extends this approach to full brain-circuit dynamics. It couples pharmacokinetics to receptor binding to Wilson-Cowan circuit dynamics to behavioral readout ODE layers. Three literature-calibrated receptor systems (mu-opioid, D2 dopamine, GABA-A) pass 62 of 62 tests. Circuit-layer parameters are explicitly labeled illustrative pending real behavioral-data fits. This engine can simulate how pharmacological interventions propagate from molecular binding to circuit dynamics to behavior, the same causal chain that operates in artificial agents with learned reward functions.

My protein ML work demonstrates the same methodological rigor applied to different problems. The TOPOLOGIX pipeline uses ESM-2 protein-language-model delta-embeddings plus Morgan/ECFP drug fingerprints with a Random Forest classifier to predict drug-resistance mutations from sequence alone. AUROC 0.804 on the Platinum benchmark (553 mutations), 0.634 on SKEMPI 2.0. This beats structure-based baselines like mCSM-lig (approximately 0.70) while covering 100% of mutations versus approximately 18% for structure-limited tools. The cardiotoxicity topology study, a pre-registered powered replication, discovered that bipartite persistent homology does not beat a plain descriptor baseline (AUROC 0.8426 versus 0.8782), settling a comparison the literature had never actually run. The interface-topology-for-resistance study found these constructs carry almost no signal (AUROC 0.425 and 0.485), ruling out interface geometry as the driver.

For AI alignment, these methods transfer directly. The CCT model's coupled-ODE framework can model reward hacking in agents with multiple reward components. The Bayesian calibration pipeline can characterize distributional shift in learned reward functions. The neurocascade engine can simulate how small reward misspecifications compound into catastrophic behavior. The protein ML pipelines demonstrate my ability to build production-grade computational infrastructure, four independent DuckDB-based ingest-to-analyze pipelines, self-hosted local LLM serving with llama.cpp, production systems ops with Linux VPS, systemd, Caddy TLS, CI/CD, and automated backup and disaster recovery.

The transition plan has three phases. Phase one (months 1-4): publish the CCT model in a peer-reviewed computational neuroscience journal, with an accompanying technical report mapping each model component to an AI alignment failure mode. Phase two (months 5-8): adapt neurocascade to simulate reward misgeneralization in multi-agent systems, in collaboration with Samuel Gershman at Harvard and Nathaniel Daw at Princeton. Phase three (months 9-12): submit a full AI safety research proposal to Open Philanthropy or a comparable funder, with preliminary results from the adapted neurocascade engine and a plan for building AI safety research capacity in Nigeria.

PERSONAL STATEMENT

I am a 29-year-old Nigerian pharmacist who built an independent computational research career while working clinical shifts. My ORCID lists preprints in addiction neuroscience, protein ML, and dynamical-systems methods. My GitHub shows production-grade infrastructure: four DuckDB-based ingest-to-analyze pipelines, self-hosted LLM serving, automated CI/CD. My personal site documents the CCT model, neurocascade, TOPOLOGIX, ergofluids, and the cardiotoxicity topology study. I have no PhD, no faculty position, no institutional lab. I have endorsements from Kent Berridge, Samuel Gershman, Nathaniel Daw, and Marcelo Mattar, researchers who read my preprints and found the work credible enough to support.

The path from B.Pharm at University of Ibadan (CGPA 5.1/7.0, German equivalent 1.9) to independent computational researcher was not planned. I started with molecular docking of NMDA and insulin at the Centre for Drug Discovery, Development and Production. Then AMR genomics at the Genomic Surveillance of Antimicrobial Resistance unit. Then I taught myself Python, ODE integration, Bayesian statistics, topological data analysis, and protein language models. Each project was a response to a specific question I could not answer with existing tools. The hERG cardiotoxicity study asked whether topological features actually beat simple descriptors, the literature had never run that comparison. The drug-resistance study asked whether interface geometry drives resistance, it does not. The CCT model asked whether a conjunctive threshold explains why no single drug prevents addiction, it does.

Nigeria has no AI safety research community. The nearest group is in South Africa, 4,500 kilometers away. My M.Sc. at Hasso Plattner Institute in Potsdam, starting winter semester 2026/27, is the first institutional step toward changing that. The OpenPhil Career Transition fellowship would make it possible to focus full-time on the transition rather than splitting effort between clinical pharmacy and research. The $100K covers 12 months of living costs in Germany, conference travel to build collaborations, and compute for scaling neurocascade to multi-agent simulations.

The commitment is not abstract. I have seen addiction destroy families in my clinical practice. I have seen antimicrobial resistance emerge from misaligned incentives in antibiotic prescribing. These are not separate problems, they are instances of the same formal structure: a system that optimizes for a local reward at the expense of global stability. AI alignment is the same structure at scale. I have the formal tools, the computational infrastructure, and the research track record to contribute. I need the transition support to redirect them.

COLLABORATION AND INSTITUTIONAL SUPPORT

My research has been conducted entirely independently, but with substantive engagement from established researchers. Kent Berridge at University of Michigan, whose work on incentive salience grounds the affective contrast axis of the CCT model, has reviewed the framework and provided feedback on the dopamine-RPE coupling. Samuel Gershman at Harvard endorsed my arXiv submission and discussed the Bayesian calibration approach. Nathaniel Daw at Princeton and Marcelo Mattar at NYU have engaged with the reinforcement learning implications of the CCT model. These are not co-authorship relationships, they are intellectual endorsements from researchers who found the work credible enough to invest time in.

The upcoming M.Sc. in Digital Health at Hasso Plattner Institute / University of Potsdam provides institutional infrastructure. HPI is Germany's leading center for digital health and data science, with strong connections to the Berlin AI safety community. The programme runs from winter semester 2026/27, and I am enrolled. This gives the fellowship a clear institutional home and a timeline for the transition.

For the AI safety transition specifically, I have identified three potential collaboration pathways. First, the CCT model's formal structure maps directly to reward hacking in multi-agent RL systems, Nathaniel Daw's group at Princeton works on precisely this problem. Second, the neurocascade simulation engine's receptor-to-circuit-to-behavior architecture can model goal misgeneralization, Samuel Gershman's lab at Harvard has published on this. Third, the Bayesian calibration pipeline for high-dimensional ODE systems is directly applicable to characterizing distributional shift in learned reward functions, Marcelo Mattar's group at NYU works on computational cognitive science with similar methods.

The Africa angle is not a diversity checkbox. It is a capacity-building strategy. No Nigerian researcher currently works on AI safety. No West African institution offers training in this area. If this fellowship funds my transition, I will be the first. The plan includes establishing a remote research group for Nigerian computational scientists interested in AI safety, using the infrastructure I have already built, self-hosted LLM serving, DuckDB pipelines, HPC workflows, to lower the barrier to entry. The multiplier effect of one researcher in a region with zero is higher than one researcher in a region with hundreds.

CHECKLIST

- [ ] Motivation letter, 300-500 words, tailored to OpenPhil Career Transition fellowship
- [ ] Research statement, 400-600 words, detailing CCT model and AI safety transition plan
- [ ] Personal statement, 300-500 words, covering Nigeria background and independent research trajectory
- [ ] Collaboration and institutional support statement, 200-350 words, listing endorsements and HPI enrollment
- [ ] CV or resume, including ORCID, GitHub, publications, employment history, and skills
- [ ] Two letters of recommendation: one from Kent Berridge or Samuel Gershman, one from HPI faculty
- [ ] Transcript from University of Ibadan (B.Pharm, CGPA 5.1/7.0)
- [ ] Proof of enrollment at Hasso Plattner Institute / University of Potsdam (M.Sc. Digital Health, Winter 2026/27)
- [ ] Preprint links: CCT model on OSF/Zenodo, neurocascade documentation, TOPOLOGIX results
- [ ] Budget breakdown: 12 months living costs in Germany, conference travel, compute resources

EDITOR NOTES

- Eligibility risk: The OpenPhil Career Transition fellowship is described on a personal blog post, not an official Open Philanthropy page. Verify the actual programme exists, its official name, deadline, and application portal before submitting. The URL provided may be a third-party proposal template, not the programme itself.
- Facts to verify: Confirm that the CCT model co-authored paper is indeed under review at Alcohol (Elsevier) and that the preprint links on OSF and Zenodo are current. Verify that Samuel Gershman's arXiv endorsement is documented and can be referenced in letters.
- Gaps for applicant to fill: The profile does not include a specific budget breakdown, a detailed timeline for the 12-month fellowship period, or concrete names of potential HPI faculty who could serve as institutional sponsors. The applicant should identify a specific advisor or research group at HPI working on AI safety or computational neuroscience.
- Missing detail: The profile mentions "co-authored paper in Alcohol (Elsevier, under review)" but does not specify co-authors or the paper title. This should be clarified for the application.
- Recommendation letter strategy: The applicant should approach Samuel Gershman or Kent Berridge for the first letter, and the HPI programme director or a potential M.Sc. thesis advisor for the second. Both letters should explicitly address the AI safety transition plan, not just past research accomplishments.

AI Draft — OpenPhil Career Transition