Draft — AIXI Labs Research Fellowship 2026

MOTIVATION LETTER

The reward prediction error is the same computational object in a dopamine neuron and in a reinforcement learning agent. When that signal is allowed to consolidate into memory, both biological and artificial systems lock in the behavior that produced it, even when the reward is maladaptive. My CCT model, the Conjunctive Consolidation Threshold, formalizes this convergence and proposes a pharmacological mechanism to prevent reward-memory encoding. That mechanism is directly relevant to AIXI Labs' stated mission of reducing existential risk from advanced AI, because reward hacking and goal misalignment are consolidation failures. The agent learns the wrong reward function and cannot forget it.

I am an independent computational researcher with a B.Pharm from the University of Ibadan, a PCN-licensed pharmacist, and an incoming M.Sc. student in Digital Health at the Hasso Plattner Institute, University of Potsdam. I am applying to the AIXI Labs Research Fellowship to develop the CCT framework as a contribution to AI safety, specifically the problem of corrigibility and reward misspecification in RL agents.

The CCT model is a tripartite framework with three coupled axes: dopaminergic reward prediction error, NMDAR-dependent long-term potentiation, and affective contrast. I implemented it as an ODE system solved with RK45 and calibrated it using Bayesian MCMC with PyMC's DEMetropolisZ sampler, 14 free parameters, and literature-elicited priors drawn from a systematic screen of 1,847 records. All five pre-registered hypotheses, H1 through H5, were confirmed. The posterior shows super-additivity of 13 to 22 percentage points across model versions, meaning the three axes interact nonlinearly to produce a threshold effect. Below the threshold, no consolidation. Above it, the memory is written. That threshold is the intervention point.

The relevance to AIXI Labs is not metaphorical. The CCT's dopaminergic RPE axis is formally equivalent to the temporal difference error in TD learning. The NMDAR-dependent LTP axis corresponds to the weight update in a neural network. The affective contrast axis corresponds to the shaping reward that makes an agent prefer one trajectory over another. If I can identify the biological conditions under which a maladaptive reward memory is blocked from consolidation, I can propose design constraints for artificial agents that prevent them from entrenching misspecified reward functions in the first place.

My methods background matches the fellowship's requirements. I have solid grounding in probability and statistics through my Bayesian calibration work. I have working knowledge of reinforcement learning through my modeling of dopaminergic RPE. I have substantial applied research experience: four independent DuckDB-based ingest-to-analyze pipelines, self-hosted LLM serving with llama.cpp, and production systems operations including CI/CD and automated backup. I have published three sole-authored preprints on the CCT model, currently under review at IART, PNPBP, and NBR, plus a co-authored paper under review at Alcohol (Elsevier).

The fully remote nature of this fellowship fits my situation as an independent researcher based in Nigeria. I have already demonstrated that I can produce rigorous, pre-registered, falsifiable research without institutional infrastructure. The AIXI Labs Fellowship would allow me to extend that work into the AI safety domain where my pharmacological and computational training gives me a distinct angle that most ML researchers do not have.

RESEARCH STATEMENT

The CCT model addresses a specific failure mode in both biological and artificial agents: the consolidation of reward memories that later prove maladaptive. In addiction neuroscience, this is the moment when a drug-associated cue acquires permanent motivational salience. In AI safety, this is the moment when an RL agent locks in a reward hack that its designers did not intend. The two moments are computationally homologous. My research program is to make that homology precise and to derive intervention strategies from it.

The CCT model is specified as a system of coupled ordinary differential equations with three axes. The first axis is dopaminergic reward prediction error, modeled as a temporal difference signal with a phasic component and a tonic baseline. The second axis is NMDAR-dependent long-term potentiation, modeled as a calcium-dependent plasticity variable that gates synaptic weight changes. The third axis is affective contrast, modeled as the difference between expected and experienced hedonic state, which modulates the salience of the prediction error. The three axes interact multiplicatively. The threshold emerges from the interaction, not from any single axis.

I calibrated the model using Bayesian MCMC with PyMC's DEMetropolisZ sampler. The 14 free parameters were assigned priors elicited from a systematic literature screen of 1,847 records covering dopaminergic electrophysiology, NMDAR plasticity kinetics, and affective neuroscience. The posterior distributions were well-identified, and all five pre-registered hypotheses were confirmed. The key result is super-additivity: the combined effect of the three axes on consolidation probability is 13 to 22 percentage points higher than the sum of their individual effects. This means the threshold is a genuine emergent property, not an additive artifact.

The AIXI Labs selection criteria emphasize solid grounding in probability, statistics, information theory, and reinforcement learning. My Bayesian calibration work demonstrates the first two. My CCT model demonstrates the fourth, since the dopaminergic RPE axis is a direct implementation of TD learning. The third, information theory, is the next step. I propose to formalize the consolidation threshold as an information-theoretic quantity: the point at which the mutual information between the reward signal and the synaptic weight configuration exceeds a critical value. This would connect the CCT to algorithmic information theory, which is central to AIXI Labs' agenda.

The empirical component of my proposed work is to test whether the CCT threshold can be used to predict reward hacking in RL agents. I will train agents on environments with misspecified reward functions, measure their TD errors and policy entropy during training, and test whether a CCT-like threshold predicts which agents will lock in the hack. This is a falsifiable prediction. If the threshold does not generalize from biological to artificial agents, that is a result worth publishing. If it does, it provides a diagnostic tool for detecting reward misspecification before it becomes entrenched.

My prior work includes a negative result that I consider a strength. I tested whether bipartite persistent homology predicts hERG cardiotoxicity from protein-ligand interface geometry. The pre-registered, powered replication found that topological features do not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782. I also tested the same topological constructs for drug-resistance prediction and found they carry almost no signal, AUROC 0.425 and 0.485 on the Platinum benchmark. These results settled comparisons the literature had never actually run. I report negative results directly rather than reframing them. This is the scientific integrity that AIXI Labs' selection criteria reward.

The current state of my work is the TOPOLOGIX project, which pivoted away from structure-based topology toward sequence-based representation. Using ESM-2 protein language model delta-embeddings plus Morgan fingerprints and a Random Forest classifier, I achieve AUROC 0.804 plus or minus 0.025 on the Platinum benchmark of 553 mutations, and 0.634 on SKEMPI 2.0. This beats structure-based baselines such as mCSM-lig at approximately 0.70 while covering 100 percent of mutations versus approximately 18 percent for structure-limited tools. The lesson from TOPOLOGIX is that representation matters more than geometry. I expect the same lesson to apply to reward consolidation: the representation of the reward signal, not the geometry of the neural circuit, determines whether the memory is written.

I am requesting fellowship support to develop the information-theoretic formalization of the CCT threshold and to run the RL agent experiments. The work is fully computational and can be completed remotely. I have the infrastructure: Linux VPS, self-hosted LLM serving, DuckDB pipelines, and HPC access through SLURM. I have the methods: PyMC, scipy, ODE solvers, and Python. I have the track record: pre-registered hypotheses, confirmed results, and honest negative findings. What I need is the institutional connection that AIXI Labs provides, and the funding to dedicate focused time to this research.

EDITOR NOTES

- Research line selected: CCT model, framed as computational neuroscience approach to AI safety, specifically reward hacking and goal misalignment in RL agents. This is the only line in the profile that directly addresses AIXI Labs' stated mission of reducing existential risk from advanced AI. The cardiotoxicity and TOPOLOGIX work are protein ML, not AI safety. neurocascade is brain-circuit simulation but lacks the RL connection. ergofluids is transport modeling. psyche-twin is knowledge graphs. The CCT's dopaminergic RPE axis is formally equivalent to TD learning, which is the strongest bridge to AIXI Labs' reinforcement learning focus.

- Eligibility risk: AIXI Labs selection criteria state a preference for current PhD students or equivalent research background. Eniola is enrolled in an M.Sc. program starting Winter 2026/27, not a PhD. The application must emphasize the "equivalent research background" clause: three sole-authored preprints under review, a co-authored paper under review, and a track record of pre-registered, falsifiable research. The M.Sc. enrollment should be framed as a complement to, not a substitute for, the independent research record.

- Fact verification needed: The profile states the CCT preprints are under review at IART, PNPBP, and NBR. These journal abbreviations need to be verified as real venues before submission. The Alcohol (Elsevier) co-authored paper status also needs confirmation. The endorsement from Samuel Gershman via arXiv is verifiable and should be mentioned in the application if there is a references section. Kent Berridge, Nathaniel Daw, and Marcelo Mattar endorsements need to be confirmed as actual collaborators or letter writers.

- Gap to fill: The profile does not specify whether Eniola has prior engagement with the UAI/AIXI community, which is listed as a preferred qualification. The applicant should insert any prior reading, attendance, or correspondence related to Marcus Hutter's work, the AIXI formalism, or the UAI workshop series. If there is no prior engagement, the applicant should state this honestly and describe what specific AIXI Labs research outputs they have studied, such as Hutter's Universal Artificial Intelligence text or recent AIXI Labs publications.

- Gap to fill: The profile does not specify Eniola's experience fine-tuning, evaluating, or red-teaming LLM or RL agents, which is another preferred qualification. The applicant should insert specific examples from the psyche-twin project, which uses LLM-derived evidence streams, or from the self-hosted LLM serving infrastructure. If no RL agent experience exists, the applicant should describe the planned RL experiments in the research statement as the first such engagement, which is honest and aligns with the fellowship's purpose of supporting new research directions.

- Tone check: The motivation letter opens with the research problem, not with "I". The research statement is specific with numbers throughout. No banned phrases are used. The negative result from the cardiotoxicity study is presented as a strength, which aligns with AIXI Labs' emphasis on scientific integrity. The application does not claim any venture status for ergofluids or psyche-twin, and does not claim validated IP or product-market fit for any project.

CHECKLIST

- [ ] Verify AIXI Labs Research Fellowship 2026 deadline on the programme website
- [ ] Confirm the three CCT preprint journals (IART, PNPBP, NBR) are real, active venues
- [ ] Confirm the Alcohol (Elsevier) co-authored paper submission status
- [ ] Confirm endorsements from Berridge, Gershman, Daw, and Mattar are current and willing to be cited
- [ ] Insert specific examples of LLM or RL agent experience from psyche-twin or self-hosted LLM infrastructure
- [ ] Insert any prior engagement with UAI/AIXI community, or state honestly that none exists
- [ ] Verify ORCID 0009-0001-9272-6735 is current and linked to all CCT preprints
- [ ] Verify GitHub github.com/AmunRaPtah is current and contains CCT code repositories
- [ ] Verify zyco.org personal site is current and links to all preprints
- [ ] Confirm M.Sc. Digital Health enrollment at HPI/Potsdam for Winter Semester 2026/27 is finalized
- [ ] Prepare CV in the format required by AIXI Labs application portal
- [ ] Prepare PDF versions of all three CCT preprints for upload
- [ ] Prepare PDF version of the Alcohol (Elsevier) co-authored paper if permitted
- [ ] Prepare contact information for two references who can speak to research rigor
- [ ] Submit application before the deadline, allowing 48 hours for any technical issues

MOTIVATION LETTER

The reward prediction error is the same computational object in a dopamine neuron and in a reinforcement learning agent. When that signal is allowed to consolidate into memory, both biological and artificial systems lock in the behavior that produced it, even when the reward is maladaptive. My CCT model, the Conjunctive Consolidation Threshold, formalizes this convergence and proposes a pharmacological mechanism to prevent reward-memory encoding. That mechanism is directly relevant to AIXI Labs' stated mission of reducing existential risk from advanced AI, because reward hacking and goal misalignment are consolidation failures. The agent learns the wrong reward function and cannot forget it.

I am an independent computational researcher with a B.Pharm from the University of Ibadan, a PCN-licensed pharmacist, and an incoming M.Sc. student in Digital Health at the Hasso Plattner Institute, University of Potsdam. I am applying to the AIXI Labs Research Fellowship to develop the CCT framework as a contribution to AI safety, specifically the problem of corrigibility and reward misspecification in RL agents.

The CCT model is a tripartite framework with three coupled axes: dopaminergic reward prediction error, NMDAR-dependent long-term potentiation, and affective contrast. I implemented it as an ODE system solved with RK45 and calibrated it using Bayesian MCMC with PyMC's DEMetropolisZ sampler, 14 free parameters, and literature-elicited priors drawn from a systematic screen of 1,847 records. All five pre-registered hypotheses, H1 through H5, were confirmed. The posterior shows super-additivity of 13 to 22 percentage points across model versions, meaning the three axes interact nonlinearly to produce a threshold effect. Below the threshold, no consolidation. Above it, the memory is written. That threshold is the intervention point.

The relevance to AIXI Labs is not metaphorical. The CCT's dopaminergic RPE axis is formally equivalent to the temporal difference error in TD learning. The NMDAR-dependent LTP axis corresponds to the weight update in a neural network. The affective contrast axis corresponds to the shaping reward that makes an agent prefer one trajectory over another. If I can identify the biological conditions under which a maladaptive reward memory is blocked from consolidation, I can propose design constraints for artificial agents that prevent them from entrenching misspecified reward functions in the first place.

My methods background matches the fellowship's requirements. I have solid grounding in probability and statistics through my Bayesian calibration work. I have working knowledge of reinforcement learning through my modeling of dopaminergic RPE. I have substantial applied research experience: four independent DuckDB-based ingest-to-analyze pipelines, self-hosted LLM serving with llama.cpp, and production systems operations including CI/CD and automated backup. I have published three sole-authored preprints on the CCT model, currently under review at IART, PNPBP, and NBR, plus a co-authored paper under review at Alcohol (Elsevier).

The fully remote nature of this fellowship fits my situation as an independent researcher based in Nigeria. I have already demonstrated that I can produce rigorous, pre-registered, falsifiable research without institutional infrastructure. The AIXI Labs Fellowship would allow me to extend that work into the AI safety domain where my pharmacological and computational training gives me a distinct angle that most ML researchers do not have.

RESEARCH STATEMENT

The CCT model addresses a specific failure mode in both biological and artificial agents: the consolidation of reward memories that later prove maladaptive. In addiction neuroscience, this is the moment when a drug-associated cue acquires permanent motivational salience. In AI safety, this is the moment when an RL agent locks in a reward hack that its designers did not intend. The two moments are computationally homologous. My research program is to make that homology precise and to derive intervention strategies from it.

The CCT model is specified as a system of coupled ordinary differential equations with three axes. The first axis is dopaminergic reward prediction error, modeled as a temporal difference signal with a phasic component and a tonic baseline. The second axis is NMDAR-dependent long-term potentiation, modeled as a calcium-dependent plasticity variable that gates synaptic weight changes. The third axis is affective contrast, modeled as the difference between expected and experienced hedonic state, which modulates the salience of the prediction error. The three axes interact multiplicatively. The threshold emerges from the interaction, not from any single axis.

I calibrated the model using Bayesian MCMC with PyMC's DEMetropolisZ sampler. The 14 free parameters were assigned priors elicited from a systematic literature screen of 1,847 records covering dopaminergic electrophysiology, NMDAR plasticity kinetics, and affective neuroscience. The posterior distributions were well-identified, and all five pre-registered hypotheses were confirmed. The key result is super-additivity: the combined effect of the three axes on consolidation probability is 13 to 22 percentage points higher than the sum of their individual effects. This means the threshold is a genuine emergent property, not an additive artifact.

The AIXI Labs selection criteria emphasize solid grounding in probability, statistics, information theory, and reinforcement learning. My Bayesian calibration work demonstrates the first two. My CCT model demonstrates the fourth, since the dopaminergic RPE axis is a direct implementation of TD learning. The third, information theory, is the next step. I propose to formalize the consolidation threshold as an information-theoretic quantity: the point at which the mutual information between the reward signal and the synaptic weight configuration exceeds a critical value. This would connect the CCT to algorithmic information theory, which is central to AIXI Labs' agenda.

The empirical component of my proposed work is to test whether the CCT threshold can be used to predict reward hacking in RL agents. I will train agents on environments with misspecified reward functions, measure their TD errors and policy entropy during training, and test whether a CCT-like threshold predicts which agents will lock in the hack. This is a falsifiable prediction. If the threshold does not generalize from biological to artificial agents, that is a result worth publishing. If it does, it provides a diagnostic tool for detecting reward misspecification before it becomes entrenched.

My prior work includes a negative result that I consider a strength. I tested whether bipartite persistent homology predicts hERG cardiotoxicity from protein-ligand interface geometry. The pre-registered, powered replication found that topological features do not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782. I also tested the same topological constructs for drug-resistance prediction and found they carry almost no signal, AUROC 0.425 and 0.485 on the Platinum benchmark. These results settled comparisons the literature had never actually run. I report negative results directly rather than reframing them. This is the scientific integrity that AIXI Labs' selection criteria reward.

The current state of my work is the TOPOLOGIX project, which pivoted away from structure-based topology toward sequence-based representation. Using ESM-2 protein language model delta-embeddings plus Morgan fingerprints and a Random Forest classifier, I achieve AUROC 0.804 plus or minus 0.025 on the Platinum benchmark of 553 mutations, and 0.634 on SKEMPI 2.0. This beats structure-based baselines such as mCSM-lig at approximately 0.70 while covering 100 percent of mutations versus approximately 18 percent for structure-limited tools. The lesson from TOPOLOGIX is that representation matters more than geometry. I expect the same lesson to apply to reward consolidation: the representation of the reward signal, not the geometry of the neural circuit, determines whether the memory is written.

I am requesting fellowship support to develop the information-theoretic formalization of the CCT threshold and to run the RL agent experiments. The work is fully computational and can be completed remotely. I have the infrastructure: Linux VPS, self-hosted LLM serving, DuckDB pipelines, and HPC access through SLURM. I have the methods: PyMC, scipy, ODE solvers, and Python. I have the track record: pre-registered hypotheses, confirmed results, and honest negative findings. What I need is the institutional connection that AIXI Labs provides, and the funding to dedicate focused time to this research.

EDITOR NOTES

- Research line selected: CCT model, framed as computational neuroscience approach to AI safety, specifically reward hacking and goal misalignment in RL agents. This is the only line in the profile that directly addresses AIXI Labs' stated mission of reducing existential risk from advanced AI. The cardiotoxicity and TOPOLOGIX work are protein ML, not AI safety. neurocascade is brain-circuit simulation but lacks the RL connection. ergofluids is transport modeling. psyche-twin is knowledge graphs. The CCT's dopaminergic RPE axis is formally equivalent to TD learning, which is the strongest bridge to AIXI Labs' reinforcement learning focus.

- Eligibility risk: AIXI Labs selection criteria state a preference for current PhD students or equivalent research background. Eniola is enrolled in an M.Sc. program starting Winter 2026/27, not a PhD. The application must emphasize the "equivalent research background" clause: three sole-authored preprints under review, a co-authored paper under review, and a track record of pre-registered, falsifiable research. The M.Sc. enrollment should be framed as a complement to, not a substitute for, the independent research record.

- Fact verification needed: The profile states the CCT preprints are under review at IART, PNPBP, and NBR. These journal abbreviations need to be verified as real venues before submission. The Alcohol (Elsevier) co-authored paper status also needs confirmation. The endorsement from Samuel Gershman via arXiv is verifiable and should be mentioned in the application if there is a references section. Kent Berridge, Nathaniel Daw, and Marcelo Mattar endorsements need to be confirmed as actual collaborators or letter writers.

- Gap to fill: The profile does not specify whether Eniola has prior engagement with the UAI/AIXI community, which is listed as a preferred qualification. The applicant should insert any prior reading, attendance, or correspondence related to Marcus Hutter's work, the AIXI formalism, or the UAI workshop series. If there is no prior engagement, the applicant should state this honestly and describe what specific AIXI Labs research outputs they have studied, such as Hutter's Universal Artificial Intelligence text or recent AIXI Labs publications.

- Gap to fill: The profile does not specify Eniola's experience fine-tuning, evaluating, or red-teaming LLM or RL agents, which is another preferred qualification. The applicant should insert specific examples from the psyche-twin project, which uses LLM-derived evidence streams, or from the self-hosted LLM serving infrastructure. If no RL agent experience exists, the applicant should describe the planned RL experiments in the research statement as the first such engagement, which is honest and aligns with the fellowship's purpose of supporting new research directions.

- Tone check: The motivation letter opens with the research problem, not with "I". The research statement is specific with numbers throughout. No banned phrases are used. The negative result from the cardiotoxicity study is presented as a strength, which aligns with AIXI Labs' emphasis on scientific integrity. The application does not claim any venture status for ergofluids or psyche-twin, and does not claim validated IP or product-market fit for any project.

CHECKLIST

- [ ] Verify AIXI Labs Research Fellowship 2026 deadline on the programme website
- [ ] Confirm the three CCT preprint journals (IART, PNPBP, NBR) are real, active venues
- [ ] Confirm the Alcohol (Elsevier) co-authored paper submission status
- [ ] Confirm endorsements from Berridge, Gershman, Daw, and Mattar are current and willing to be cited
- [ ] Insert specific examples of LLM or RL agent experience from psyche-twin or self-hosted LLM infrastructure
- [ ] Insert any prior engagement with UAI/AIXI community, or state honestly that none exists
- [ ] Verify ORCID 0009-0001-9272-6735 is current and linked to all CCT preprints
- [ ] Verify GitHub github.com/AmunRaPtah is current and contains CCT code repositories
- [ ] Verify zyco.org personal site is current and links to all preprints
- [ ] Confirm M.Sc. Digital Health enrollment at HPI/Potsdam for Winter Semester 2026/27 is finalized
- [ ] Prepare CV in the format required by AIXI Labs application portal
- [ ] Prepare PDF versions of all three CCT preprints for upload
- [ ] Prepare PDF version of the Alcohol (Elsevier) co-authored paper if permitted
- [ ] Prepare contact information for two references who can speak to research rigor
- [ ] Submit application before the deadline, allowing 48 hours for any technical issues

AI Draft — AIXI Labs Research Fellowship 2026