Draft — MATS Autumn 2026 Scholar - AI in Society

MOTIVATION LETTER

The CCT model, a tripartite pharmacological framework for reward-memory encoding prevention in addiction, directly addresses a core AI safety problem: how to prevent a learning system from locking onto a single, destructive reward signal. My work formalizes this as a three-axis ODE system coupling dopaminergic reward-prediction error, NMDAR-dependent long-term potentiation, and affective contrast. Bayesian MCMC calibration with PyMC DEMetropolisZ across 14 free parameters, priors drawn from a systematic screen of 1,847 records, confirmed all five pre-registered hypotheses with posterior super-additivity of 13 to 22 percentage points across model versions. This is a mathematical structure that can be ported directly into reinforcement learning architectures to test whether the same dynamical principles that prevent pathological memory consolidation in biological circuits can prevent reward hacking and goal misgeneralization in artificial agents.

The MATS Autumn 2026 Scholar programme at AI in Society Hub, with its focus on AI alignment and governance, is the correct environment to develop this translation. My independent research record demonstrates the output orientation the programme requires: three sole-authored preprints under peer review at IART, PNPBP, and NBR, plus a co-authored paper under review at Alcohol (Elsevier). I have built neurocascade, a receptor-to-behavior brain-circuit simulation engine that couples pharmacokinetics to receptor binding to Wilson-Cowan circuit dynamics to behavioral readout, with 62 of 62 tests passing. This engine provides a testbed for simulating how pharmacological interventions alter circuit-level dynamics, which maps directly onto the problem of designing corrigible AI systems.

My technical competence spans ODE modeling, Bayesian calibration, protein language models, topological data analysis, and production systems engineering. I have built four independent DuckDB-based ingest-to-analyze pipelines across life sciences, tech/AI security, and social science domains. I self-host local LLM serving with llama.cpp and manage production Linux infrastructure. These skills mean I can produce tangible research outputs within the 10-week intensive phase, not just plan them.

The Africa angle is structural, not biographical. Nigerian pharmaceutical regulation and clinical practice confront addiction without the computational tools available in high-income settings. A validated CCT framework, translated into a deployable screening tool, could change how addiction risk is assessed in primary care across West Africa. MATS provides the mentorship and peer environment to pressure-test this translation against the AI safety literature at Institut Polytechnique de Paris and HEC Paris.

I am 29, enrolled in an M.Sc. in Digital Health at Hasso Plattner Institute starting Winter 2026/27, and have no PhD yet. This is the moment to embed AI safety thinking into my research trajectory before it solidifies. MATS is the programme for that embedding.

RESEARCH STATEMENT

My independent research program addresses a single question: can the dynamical principles that prevent pathological reward-memory consolidation in biological neural circuits inform the design of safe artificial reward systems? The CCT model is the current answer.

The model formalizes addiction as a failure of conjunctive consolidation thresholding. Three coupled axes interact: dopaminergic reward-prediction error drives salience, NMDAR-dependent LTP encodes the memory trace, and affective contrast provides a gating signal that determines whether the trace consolidates. I implemented this as a system of ODEs solved with RK45, calibrated against literature data using Bayesian MCMC with PyMC DEMetropolisZ. The 14 free parameters were assigned priors from a systematic literature screen of 1,847 records. All five pre-registered hypotheses (H1-H5) were confirmed. Posterior super-additivity ranged from 13 to 22 percentage points across model versions, meaning the interaction of the three axes produces effects larger than the sum of their individual contributions. This is the core finding: the threshold is not a fixed value but an emergent property of coupled dynamics.

The neurocascade simulation engine extends this approach to whole circuits. It couples four layers: pharmacokinetics, receptor binding, Wilson-Cowan circuit dynamics, and behavioral readout. Three literature-calibrated receptor/circuit systems are implemented: mu-opioid, D2 dopamine, and GABA-A. Bayesian calibration with PyMC confirmed 62 of 62 tests passing. The circuit-layer parameters are explicitly labeled illustrative pending real behavioral-data fits, but the architecture is validated and ready for data.

The TOPOLOGIX project demonstrates my ability to switch between methods when evidence demands it. A pre-registered replication of bipartite persistent homology for hERG cardiotoxicity prediction found that topological features do not beat a plain descriptor baseline (AUROC 0.8426 vs 0.8782), settling a comparison the published literature had never actually run. Applying the same topological constructs to drug-resistance prediction produced AUROC values of 0.425 and 0.485 on the Platinum benchmark, ruling out interface geometry as the driver. I then built TOPOLOGIX using ESM-2 protein-language-model delta-embeddings plus Morgan/ECFP drug fingerprints with a Random Forest classifier, achieving AUROC 0.804 plus or minus 0.025 on the Platinum benchmark (553 mutations) and 0.634 on SKEMPI 2.0. This beats structure-based baselines like mCSM-lig at approximately 0.70 while covering 100 percent of mutations versus approximately 18 percent for structure-limited tools.

For the MATS 10-week phase, I propose to formalize the CCT model as a constraint on reward function design in deep reinforcement learning. The three-axis coupling can be implemented as a regularization term that penalizes runaway consolidation of any single reward signal. I will test this on standard RL benchmarks (Atari, MuJoCo) and on a custom addiction-mimicking environment where a single action yields high but diminishing returns. Success would be a published preprint demonstrating that CCT-regularized agents show reduced reward-hacking behavior compared to baseline PPO and SAC agents, with analysis of the learned representations using the same topological tools I have already validated.

TECHNICAL STATEMENT

My computational toolkit is built for independent, reproducible research across domains. I write production-grade Python using scipy, numpy, pandas, and PyMC for Bayesian inference. ODE systems are solved with RK45. Topological data analysis uses Ripser and GUDHI. Protein language models use ESM-2 with custom embedding extraction. Drug chemistry uses RDKit for Morgan/ECFP fingerprints. Neural simulation uses NEURON and Brian2. Molecular dynamics uses GROMACS. Docking uses AutoDock. Workflow management uses Nextflow on SLURM HPC clusters.

I maintain four independent DuckDB-based ingest-to-analyze pipelines. Each pipeline ingests raw data from domain-specific sources, normalizes schemas, runs analysis, and produces publication-ready figures. The life-sciences pipeline handles protein sequences, drug structures, and pharmacological assay data. The tech/AI security pipeline tracks vulnerability disclosures and model card metadata. The social-science pipeline ingests survey and behavioral data. The fourth pipeline is a general-purpose RAG system for literature review, using local LLM serving with llama.cpp and on-demand model swapping.

My production systems experience includes Linux VPS administration with systemd, Caddy TLS termination, CI/CD pipelines, and automated backup and disaster-recovery procedures. This infrastructure supports all my research workflows and ensures reproducibility: every analysis is containerized, version-controlled, and executable from a single command.

For the MATS programme, I will use this infrastructure to run the CCT-RL experiments. The DuckDB pipeline will ingest benchmark results. The local LLM serving will support literature review and hypothesis generation. The CI/CD system will ensure that every experiment is reproducible from the raw data to the final figure. This is how I already work.

SHORT ESSAY: RELEVANCE TO AI SAFETY

Addiction is a failure mode of a learning system. A biological agent encounters a reward signal, encodes it, and then cannot stop pursuing it even when the signal becomes destructive. This is structurally identical to reward hacking in reinforcement learning agents, where a proxy reward is optimized to the exclusion of the true objective. The CCT model shows that the biological solution to this problem is a conjunctive threshold: consolidation requires simultaneous activation of three coupled axes, not just one. No single signal can hijack the system.

This principle is directly applicable to AI safety. Current RL agents use a single scalar reward signal. If that signal is misspecified or hackable, the agent optimizes it pathologically. A CCT-regularized agent would require conjunctive activation across multiple channels before any reward signal is consolidated into a stable policy. This provides a structural defense against reward misspecification, not just a heuristic patch.

The neurocascade simulation engine extends this to circuit-level analysis. By simulating how pharmacological interventions alter the dynamics of Wilson-Cowan circuits, I can test whether the same interventions that prevent addiction in biological systems can prevent reward hacking in artificial systems. This is a concrete research program.

MATS provides the mentorship to formalize this connection rigorously. The labs at Institut Polytechnique de Paris and HEC Paris have relevant expertise in reinforcement learning theory and multi-agent systems. My independent preprint record and technical infrastructure mean I can execute the experiments within the 10-week phase.

SHORT ESSAY: INDEPENDENT RESEARCH CAPACITY

I have produced three sole-authored preprints, each under peer review at a different journal, without any institutional affiliation or PhD supervision. The CCT model preprints are on OSF and Zenodo, under review at IART, PNPBP, and NBR. The co-authored paper under review at Alcohol (Elsevier) was produced through remote collaboration. I have endorsements from Kent Berridge at Michigan, Samuel Gershman at Harvard, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU. These are not co-authors. They are scientists who have read my work and vouched for its quality.

My research output is not dependent on lab resources. I run Bayesian MCMC calibrations on my own HPC allocations. I maintain my own DuckDB pipelines. I self-host my own LLM serving. I have built a complete research infrastructure from scratch, documented it, and made it reproducible. The TOPOLOGIX project, which beats structure-based baselines while covering 100 percent of mutations, was executed entirely independently.

The ergofluids project demonstrates my approach to negative results. The pre-registered gated validation pipeline specified that the first real-data gate would test against digitized published figures. It did not meet the primary pre-registered criterion. I reported this directly rather than reframing it. The paper is a methods-validation study. No IP or product claims are made. This is how independent research should work: pre-register, execute, report honestly.

For MATS, this independence means I will not need hand-holding. I will arrive with a clear research plan, the infrastructure to execute it, and the discipline to report results honestly. The 10-week phase will be productive from day one.

CHECKLIST

- [ ] Complete MATS Autumn 2026 Scholar application form at aiinsocietyhub.com
- [ ] Upload motivation letter (this document)
- [ ] Upload research statement (this document)
- [ ] Upload technical statement (this document)
- [ ] Upload short essay on relevance to AI safety (this document)
- [ ] Upload short essay on independent research capacity (this document)
- [ ] Provide ORCID: 0009-0001-9272-6735
- [ ] Provide GitHub: github.com/AmunRaPtah
- [ ] Provide personal site: zyco.org
- [ ] Provide current affiliation: Independent researcher
- [ ] Provide nationality: Nigerian
- [ ] Provide degree information: B.Pharm, University of Ibadan; enrolled M.Sc. Digital Health, HPI/Potsdam
- [ ] Provide links to three sole-authored preprints on OSF/Zenodo
- [ ] Provide link to co-authored paper under review at Alcohol (Elsevier)
- [ ] Provide links to endorsements from Berridge, Gershman, Daw, Mattar if available
- [ ] Confirm eligibility for early-career, pre-PhD, LMIC-track programmes
- [ ] Confirm MATS Autumn 2026 deadline from programme website

EDITOR NOTES

- Eligibility risk: MATS may require PhD enrollment or completion. Eniola has a B.Pharm and is enrolled in an M.Sc. Verify whether the programme accepts pre-PhD applicants. If not, the application may need to emphasize the M.Sc. enrollment as current graduate-level training.
- Endorsement letters: Berridge, Gershman, Daw, and Mattar are named as endorsers. Confirm that each is willing to provide a letter or be contacted. The application may require formal letters of recommendation.
- Preprint status: Three sole-authored preprints are under review at IART, PNPBP, and NBR. Verify current review status. If any have been accepted or rejected, update the application accordingly.
- M.Sc. start date: Enrolled for Winter Semester 2026/27 at HPI/Potsdam. Confirm exact start month. If the MATS programme overlaps with the first semester, clarify how time will be allocated.
- Africa angle: The application mentions Nigerian pharmaceutical regulation and clinical practice. Eniola should insert specific examples of addiction prevalence in Nigeria or West Africa, and any personal clinical experience with addiction patients, to ground this claim.

MOTIVATION LETTER

The CCT model, a tripartite pharmacological framework for reward-memory encoding prevention in addiction, directly addresses a core AI safety problem: how to prevent a learning system from locking onto a single, destructive reward signal. My work formalizes this as a three-axis ODE system coupling dopaminergic reward-prediction error, NMDAR-dependent long-term potentiation, and affective contrast. Bayesian MCMC calibration with PyMC DEMetropolisZ across 14 free parameters, priors drawn from a systematic screen of 1,847 records, confirmed all five pre-registered hypotheses with posterior super-additivity of 13 to 22 percentage points across model versions. This is a mathematical structure that can be ported directly into reinforcement learning architectures to test whether the same dynamical principles that prevent pathological memory consolidation in biological circuits can prevent reward hacking and goal misgeneralization in artificial agents.

The MATS Autumn 2026 Scholar programme at AI in Society Hub, with its focus on AI alignment and governance, is the correct environment to develop this translation. My independent research record demonstrates the output orientation the programme requires: three sole-authored preprints under peer review at IART, PNPBP, and NBR, plus a co-authored paper under review at Alcohol (Elsevier). I have built neurocascade, a receptor-to-behavior brain-circuit simulation engine that couples pharmacokinetics to receptor binding to Wilson-Cowan circuit dynamics to behavioral readout, with 62 of 62 tests passing. This engine provides a testbed for simulating how pharmacological interventions alter circuit-level dynamics, which maps directly onto the problem of designing corrigible AI systems.

My technical competence spans ODE modeling, Bayesian calibration, protein language models, topological data analysis, and production systems engineering. I have built four independent DuckDB-based ingest-to-analyze pipelines across life sciences, tech/AI security, and social science domains. I self-host local LLM serving with llama.cpp and manage production Linux infrastructure. These skills mean I can produce tangible research outputs within the 10-week intensive phase, not just plan them.

The Africa angle is structural, not biographical. Nigerian pharmaceutical regulation and clinical practice confront addiction without the computational tools available in high-income settings. A validated CCT framework, translated into a deployable screening tool, could change how addiction risk is assessed in primary care across West Africa. MATS provides the mentorship and peer environment to pressure-test this translation against the AI safety literature at Institut Polytechnique de Paris and HEC Paris.

I am 29, enrolled in an M.Sc. in Digital Health at Hasso Plattner Institute starting Winter 2026/27, and have no PhD yet. This is the moment to embed AI safety thinking into my research trajectory before it solidifies. MATS is the programme for that embedding.

RESEARCH STATEMENT

My independent research program addresses a single question: can the dynamical principles that prevent pathological reward-memory consolidation in biological neural circuits inform the design of safe artificial reward systems? The CCT model is the current answer.

The model formalizes addiction as a failure of conjunctive consolidation thresholding. Three coupled axes interact: dopaminergic reward-prediction error drives salience, NMDAR-dependent LTP encodes the memory trace, and affective contrast provides a gating signal that determines whether the trace consolidates. I implemented this as a system of ODEs solved with RK45, calibrated against literature data using Bayesian MCMC with PyMC DEMetropolisZ. The 14 free parameters were assigned priors from a systematic literature screen of 1,847 records. All five pre-registered hypotheses (H1-H5) were confirmed. Posterior super-additivity ranged from 13 to 22 percentage points across model versions, meaning the interaction of the three axes produces effects larger than the sum of their individual contributions. This is the core finding: the threshold is not a fixed value but an emergent property of coupled dynamics.

The neurocascade simulation engine extends this approach to whole circuits. It couples four layers: pharmacokinetics, receptor binding, Wilson-Cowan circuit dynamics, and behavioral readout. Three literature-calibrated receptor/circuit systems are implemented: mu-opioid, D2 dopamine, and GABA-A. Bayesian calibration with PyMC confirmed 62 of 62 tests passing. The circuit-layer parameters are explicitly labeled illustrative pending real behavioral-data fits, but the architecture is validated and ready for data.

The TOPOLOGIX project demonstrates my ability to switch between methods when evidence demands it. A pre-registered replication of bipartite persistent homology for hERG cardiotoxicity prediction found that topological features do not beat a plain descriptor baseline (AUROC 0.8426 vs 0.8782), settling a comparison the published literature had never actually run. Applying the same topological constructs to drug-resistance prediction produced AUROC values of 0.425 and 0.485 on the Platinum benchmark, ruling out interface geometry as the driver. I then built TOPOLOGIX using ESM-2 protein-language-model delta-embeddings plus Morgan/ECFP drug fingerprints with a Random Forest classifier, achieving AUROC 0.804 plus or minus 0.025 on the Platinum benchmark (553 mutations) and 0.634 on SKEMPI 2.0. This beats structure-based baselines like mCSM-lig at approximately 0.70 while covering 100 percent of mutations versus approximately 18 percent for structure-limited tools.

For the MATS 10-week phase, I propose to formalize the CCT model as a constraint on reward function design in deep reinforcement learning. The three-axis coupling can be implemented as a regularization term that penalizes runaway consolidation of any single reward signal. I will test this on standard RL benchmarks (Atari, MuJoCo) and on a custom addiction-mimicking environment where a single action yields high but diminishing returns. Success would be a published preprint demonstrating that CCT-regularized agents show reduced reward-hacking behavior compared to baseline PPO and SAC agents, with analysis of the learned representations using the same topological tools I have already validated.

TECHNICAL STATEMENT

My computational toolkit is built for independent, reproducible research across domains. I write production-grade Python using scipy, numpy, pandas, and PyMC for Bayesian inference. ODE systems are solved with RK45. Topological data analysis uses Ripser and GUDHI. Protein language models use ESM-2 with custom embedding extraction. Drug chemistry uses RDKit for Morgan/ECFP fingerprints. Neural simulation uses NEURON and Brian2. Molecular dynamics uses GROMACS. Docking uses AutoDock. Workflow management uses Nextflow on SLURM HPC clusters.

I maintain four independent DuckDB-based ingest-to-analyze pipelines. Each pipeline ingests raw data from domain-specific sources, normalizes schemas, runs analysis, and produces publication-ready figures. The life-sciences pipeline handles protein sequences, drug structures, and pharmacological assay data. The tech/AI security pipeline tracks vulnerability disclosures and model card metadata. The social-science pipeline ingests survey and behavioral data. The fourth pipeline is a general-purpose RAG system for literature review, using local LLM serving with llama.cpp and on-demand model swapping.

My production systems experience includes Linux VPS administration with systemd, Caddy TLS termination, CI/CD pipelines, and automated backup and disaster-recovery procedures. This infrastructure supports all my research workflows and ensures reproducibility: every analysis is containerized, version-controlled, and executable from a single command.

For the MATS programme, I will use this infrastructure to run the CCT-RL experiments. The DuckDB pipeline will ingest benchmark results. The local LLM serving will support literature review and hypothesis generation. The CI/CD system will ensure that every experiment is reproducible from the raw data to the final figure. This is how I already work.

SHORT ESSAY: RELEVANCE TO AI SAFETY

Addiction is a failure mode of a learning system. A biological agent encounters a reward signal, encodes it, and then cannot stop pursuing it even when the signal becomes destructive. This is structurally identical to reward hacking in reinforcement learning agents, where a proxy reward is optimized to the exclusion of the true objective. The CCT model shows that the biological solution to this problem is a conjunctive threshold: consolidation requires simultaneous activation of three coupled axes, not just one. No single signal can hijack the system.

This principle is directly applicable to AI safety. Current RL agents use a single scalar reward signal. If that signal is misspecified or hackable, the agent optimizes it pathologically. A CCT-regularized agent would require conjunctive activation across multiple channels before any reward signal is consolidated into a stable policy. This provides a structural defense against reward misspecification, not just a heuristic patch.

The neurocascade simulation engine extends this to circuit-level analysis. By simulating how pharmacological interventions alter the dynamics of Wilson-Cowan circuits, I can test whether the same interventions that prevent addiction in biological systems can prevent reward hacking in artificial systems. This is a concrete research program.

MATS provides the mentorship to formalize this connection rigorously. The labs at Institut Polytechnique de Paris and HEC Paris have relevant expertise in reinforcement learning theory and multi-agent systems. My independent preprint record and technical infrastructure mean I can execute the experiments within the 10-week phase.

SHORT ESSAY: INDEPENDENT RESEARCH CAPACITY

I have produced three sole-authored preprints, each under peer review at a different journal, without any institutional affiliation or PhD supervision. The CCT model preprints are on OSF and Zenodo, under review at IART, PNPBP, and NBR. The co-authored paper under review at Alcohol (Elsevier) was produced through remote collaboration. I have endorsements from Kent Berridge at Michigan, Samuel Gershman at Harvard, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU. These are not co-authors. They are scientists who have read my work and vouched for its quality.

My research output is not dependent on lab resources. I run Bayesian MCMC calibrations on my own HPC allocations. I maintain my own DuckDB pipelines. I self-host my own LLM serving. I have built a complete research infrastructure from scratch, documented it, and made it reproducible. The TOPOLOGIX project, which beats structure-based baselines while covering 100 percent of mutations, was executed entirely independently.

The ergofluids project demonstrates my approach to negative results. The pre-registered gated validation pipeline specified that the first real-data gate would test against digitized published figures. It did not meet the primary pre-registered criterion. I reported this directly rather than reframing it. The paper is a methods-validation study. No IP or product claims are made. This is how independent research should work: pre-register, execute, report honestly.

For MATS, this independence means I will not need hand-holding. I will arrive with a clear research plan, the infrastructure to execute it, and the discipline to report results honestly. The 10-week phase will be productive from day one.

CHECKLIST

- [ ] Complete MATS Autumn 2026 Scholar application form at aiinsocietyhub.com
- [ ] Upload motivation letter (this document)
- [ ] Upload research statement (this document)
- [ ] Upload technical statement (this document)
- [ ] Upload short essay on relevance to AI safety (this document)
- [ ] Upload short essay on independent research capacity (this document)
- [ ] Provide ORCID: 0009-0001-9272-6735
- [ ] Provide GitHub: github.com/AmunRaPtah
- [ ] Provide personal site: zyco.org
- [ ] Provide current affiliation: Independent researcher
- [ ] Provide nationality: Nigerian
- [ ] Provide degree information: B.Pharm, University of Ibadan; enrolled M.Sc. Digital Health, HPI/Potsdam
- [ ] Provide links to three sole-authored preprints on OSF/Zenodo
- [ ] Provide link to co-authored paper under review at Alcohol (Elsevier)
- [ ] Provide links to endorsements from Berridge, Gershman, Daw, Mattar if available
- [ ] Confirm eligibility for early-career, pre-PhD, LMIC-track programmes
- [ ] Confirm MATS Autumn 2026 deadline from programme website

EDITOR NOTES

- Eligibility risk: MATS may require PhD enrollment or completion. Eniola has a B.Pharm and is enrolled in an M.Sc. Verify whether the programme accepts pre-PhD applicants. If not, the application may need to emphasize the M.Sc. enrollment as current graduate-level training.
- Endorsement letters: Berridge, Gershman, Daw, and Mattar are named as endorsers. Confirm that each is willing to provide a letter or be contacted. The application may require formal letters of recommendation.
- Preprint status: Three sole-authored preprints are under review at IART, PNPBP, and NBR. Verify current review status. If any have been accepted or rejected, update the application accordingly.
- M.Sc. start date: Enrolled for Winter Semester 2026/27 at HPI/Potsdam. Confirm exact start month. If the MATS programme overlaps with the first semester, clarify how time will be allocated.
- Africa angle: The application mentions Nigerian pharmaceutical regulation and clinical practice. Eniola should insert specific examples of addiction prevalence in Nigeria or West Africa, and any personal clinical experience with addiction patients, to ground this claim.

AI Draft — MATS Autumn 2026 Scholar - AI in Society