01 / Control & forgetting
What should information retain, and what should it lose?
TSDR & semantic lifecycles
Research foundation · Design & implementation
Typed Semantic Decay plus Staged Key Retirement (TSDR) is a foundation of the lab’s information-lifecycle research. It studies how sensitive meaning can lose specificity over time while preserving permitted utility, including the lifecycle of derived information.
Subject of U.S. provisional patent application 64/057,190, filed May 4, 2026 by James Francis Clarke, Jr.; patent pending.
Verifiable machine unlearning
Research implementation · Measured experiments
Work on retraining, deterministic computation, and model-artifact verification asks what evidence can support a claim that selected training contributions have been removed. Studies distinguish counterfactual artifact agreement from measured forgetting behavior.
The evaluated scope covers governed adapter training and counterfactual retraining with a fixed base model. Foundation-model pretraining is outside that scope.
Governed memory & temporal reasoning
Research prototype · Measured experiments
The governed-memory program studies source-grounded records, changing facts, claim lineage, retrieval, faithful quotation, and deletion completeness through controlled experiments and error analysis.
Source faithfulness and deletion are evaluated within the governed system.
Explore the governed-memory project ↗02 / Cryptographic assurance
What can a proof or attestation establish?
Zero-knowledge semantic transitions
Research implementation · Adversarial evaluation
Scoped proof systems examine whether selected transformations of hidden values were performed as specified. The work includes proof packaging, binding, and checks against replay or substitution.
Current implementation covers selected deterministic transformations of committed values.
Verifiable differential privacy
Theory & design · Early tooling
Design research into proof-carrying database queries: could a noisy answer arrive with evidence that a specified differential-privacy mechanism was applied correctly?
The project is at an early tooling and scaffold stage. Performance figures remain design targets.
Erasure assurance & attested execution
Formal models · Implementations & scoped experiments
Work on erasure receipts, reproducible artifacts, and attested execution asks how evidence can connect a governance operation to the system that actually performed it.
Cryptographic assumptions, hardware trust, surviving copies, backups, and deployment obligations remain explicit parts of the claim.
Post-quantum analysis & formal methods
Mathematical analysis · Experiments & finite formal models
Research examines security arguments, verification procedures, source fidelity, and the relationship between mathematical assumptions and implemented tests. Machine-checked finite models help make selected claims precise.
Machine-checked results apply to the stated finite models. The checked models do not cover the full security game.
03 / Adversarial privacy
Where do privacy protections fail under pressure?
Privacy evaluation & attack suites
Implemented test suites · Measured experiments
Adversarial work examines inference leakage, linkage, deletion residue, proof integrity, and governance failures through targeted tests and defect injection.
Evaluation combines project-run test suites and AI-assisted review, with negative controls and patch-and-retest cycles.
Harvest Now, Re-Identify Later
Synthetic experiments · Aggregate-data analysis
HNRL studies how retained information may become more revealing when auxiliary sources, models, or computational capabilities change. The program examines mosaic privacy, linkage risk, and the cost of proposed inference methods.
Identity-linkage results concern synthetic populations. Aggregate population analysis is reported separately. Quantum components use bounded models rather than hardware demonstrations.
Restricted quantum inference & reliability
Simulation experiments · Implementation cross-checks
Small-model studies ask whether permitted observations can reveal deliberately hidden structure, and how false positives change when the assumed noise model is wrong.
Results characterize the simulated systems and noise assumptions. Cross-checks assess implementation consistency. Hardware feasibility requires separate evaluation.
04 / Privacy foundations
What should privacy engineering protect?
Harm reduction, governance & equitable design
Essays · Literature synthesis & design frameworks
The earlier privacy-theory work examines contextual integrity, post-consent governance, collective privacy, and equity. It connects technical design choices to the harms and relationships those choices affect.
Find public resources ↗Research and implementation
Mochi, Shomei Memory, and Shomei Unlearn connect the research program to concrete implementations. Research maturity and implementation status are identified separately.
Maturity labels distinguish designs, implementations, measured experiments, and formal models.
Historical graded and predictive decay concepts remain part of the research lineage. The current governed-memory prototype uses sharp deletion.
Read the methods framework ↗