The shared scaffold
Every domain we work in has the same skeleton: a content-addressed dataset, a parametric and a non-parametric model run in parallel, an adversarial stress harness, and a public review thread. What changes between surfaces is the adversary — the source of the hardness — and that, more than the model class, is what tells us whether a result is real.
A domain earns its place by having a credible hard distribution.
If the only adversary on a domain is "we held out 20% of the data," it is not a research surface — it is an exam. Every domain below has a defended position on what the worst-case data looks like and why.
01 — Reinforcement learning & agentic systems
We study agents that learn under partial observability and under non-stationary reward, where the environment itself shifts as the policy improves. The flagship project is feature-based Q-learning under spectral drift: the feature basis evolves with the policy, so a stable representation cannot be assumed.
- Open-loop and closed-loop benchmark suites with content-addressed seeds
- Counterfactual replays — what would the policy have done at this state-hash, two checkpoints ago?
- Public adversary catalogue: every reward perturbation we use is published and citeable
02 — Computational biology & cohort analysis
Longitudinal cohorts are the hardest data we touch — sparse, irregular, and with a censoring process that is itself informative. We study non-parametric bounds on observational bias and the price of pretending a cohort is a randomised trial when it is not.
- Hazard-rate models with content-addressed cohort partitions
- Synthetic cohort generators for stress-testing causal claims
- A reproducibility audit pipeline for prior published findings — when a claim fails to replicate, the audit is published with the negative result
03 — Quantum-inspired optimisation
Not quantum hardware — quantum-inspired classical solvers that adopt the structure of variational ansätze for combinatorial problems. We are interested in where the speedup actually comes from: representation, or sampling, or both.
- Benchmarks against simulated annealing on the same problem encodings
- Ablation suites that strip the quantum-inspired layer to see if anything is left
04 — Cryptographic protocol verification
Protocol verification is a domain where a negative result is the publishable result. We focus on machine-checked proofs and the gap between the proven model and the deployed implementation.
- Proof artefacts in Coq and Lean, content-addressed alongside the protocol document
- Side-channel models published with the verification — a clean proof is not the same as a safe implementation
- Joint work with the Cryptography course (CS-6343) at TTU
05 — Adversarial robustness & safety
The catch-all surface for problems where the adversary is the data — prompt injection, distributional shift, emergent misuse. The lab takes an empirical-first stance here: a defence claim is only as strong as the suite of attacks it survives.
- Public attack catalogue with reproducible payloads
- Held-out red team rotation — reviewers run their own attacks at peer-review time
- Disclosure protocol for vulnerabilities found in deployed systems
How a domain joins the atlas
A new domain enters the atlas only when it has (a) a credible adversary, (b) a content-addressed dataset, and (c) a working group of at least three reviewers willing to triage submissions. We review the atlas every semester and retire surfaces that have lost their hard distribution.