40 public projects in seven lines of research. Everything here orbits one idea: models whose decisions you can read.
Drag to turn the map, scroll to zoom, click any project. Press B to see a black box.
Loading the map
The same map, flattened for reading. Every project is open source on GitHub.
Measured steps toward a language model whose knowledge is a set of readable Tsetlin Machine clauses. Master's research at CAIR, University of Agder.
Three studies so far; the first alone registered 91 hypotheses before the runs. Results are reported as measured, and most of them are negative.Clause-based routing for LLM agents over frozen embeddings. Low-margin queries go to a human, and every decision comes with a SAT-checkable receipt.
94.2% routing accuracy on CLINC150 (logistic regression: 95.6%).BERT lends its attention edges as the graph, then gets uninstalled. A Graph Tsetlin Machine learns Boolean clauses on top.
94.77% mean accuracy on R8 over 5 seeds (the BERT teacher: about 98%).One robustness harness for Tsetlin Machines and transformers, with adversarial attacks and counterfactual data.
Negative result: on counterfactual IMDb the Tsetlin Machine loses about 27 points, BERT about 2.Subword and dependency graphs for text classification with a Graph Tsetlin Machine.
95.05% mean accuracy on R8 over 5 seeds.Attention from a decoder LLM (Qwen2.5-1.5B) as the graph structure for a Graph Tsetlin Machine.
94.53% mean on R8. Work in progress; the differences between teachers are not significant.A graph-walking Tsetlin Machine for Ames mutagenicity that explains each prediction and proposes a minimal valid molecule edit.
Recourse succeeds for 95.5% (191/200) of positives. AUROC 0.685 trails a fingerprint baseline, and the README says so.A library for stacking Graph Tsetlin Machine layers with projected feedback between layers.
Alpha. One-layer parity with the reference GraphTM is verified; joint two-layer training is documented as failing.One Graph Tsetlin Machine tower per input view, fused by readable cross-view clauses.
0.968 vs 0.727 for late fusion on a synthetic cross-view task; on real data the two tie.Weight-tied recurrent message passing for Graph Tsetlin Machines.
Alpha, with the negative results written down.A CPU-only runtime guardrail for AI agents: a Tsetlin content guard with readable clauses, temporal monitors and a tamper-evident certificate ledger.
F1 0.827 on deepset/prompt-injections at 107 µs per check, in 12.5 KiB.An interpretable guardrail whose declared safety properties are proved or refuted by SAT/SMT solvers, with bypasses repaired by construction.
Retraining stalls at 8 provable bypasses; constructive repair reaches 0 while accuracy goes 0.69 to 0.71.Echo-Trace automata and bounded temporal literals for verifying whole agent trajectories, not single steps.
Slow-roll exfiltration (synthetic): attack success 0.687 per step, 0.000 with trajectory monitors.Least-cost repair of unsafe agent tool calls under temporal safety policies, sealed in certificates.
3x the task completion of blocking at the same safety (synthetic benchmark); 226 µs median per step.An audit layer that checks LLM claims against evidence with a dual-graph Tsetlin Machine.
Work in progress: the README lists targets, not measurements.A verifier for AI agent tool calls: ALLOW, DENY or NEEDS-OVERSIGHT, with clauses and signed receipts.
AgentDojo: attack success 0.000 at 0.302 benign utility (fail-closed); 6.18 ms per call on CPU.An explainable hiring-decision prototype with clause explanations, receipts and counterfactual recourse.
Prototype; the current accuracy is weak and documented as such.Interpretability by construction: a Legible Bottleneck routes a layer through named concepts and emits a Faithfulness Certificate on every forward pass.
On synthetic ground truth: δ ≈ 0.002 (99.9% agreement), no measurable accuracy tax, backdoor detection at AUC ≈ 0.90.A glass-box, safety-shielded agent for ARC-AGI-3, with a readable world model and a decision trace for every step.
Completes 2 public ARC-AGI-3 levels (VC33, LP85); its overall score sits in the same sub-1% regime as frontier models.A generative Tsetlin Machine: discrete diffusion over Boolean vectors, readable at every step.
Anomaly ROC AUC 0.677 vs 0.580 for naive Bayes.Certified flip-distance robustness and abstention for Tsetlin Machines.
Certificate in about 0.05 ms per sample (synthetic data).Conformal prediction sets and coverage-guaranteed recourse for Tsetlin Machines.
Adaptive coverage 0.900 at a 0.9 target under drift (synthetic).A federated Tsetlin Machine with differential privacy and Rényi accounting.
0.875 accuracy at ε = 4 (synthetic data).A Tsetlin Machine trained on counterfactual pairs, with recourse receipts.
Demographic parity gap 0.041 vs 0.181 (synthetic data).Continual learning for Tsetlin Machines.
Work in progress.A survey of Boolean clause learners across seven sibling projects.
Survey draft.A wildfire spread simulator with forecast-aware evacuation routing.
Synthetic landscapes, with a technical report.An indoor heat-risk forecaster with equity-aware triage.
Synthetic cities, with a technical report.A browser app that estimates indoor PM2.5 during wildfire smoke and ranks interventions.
Reproduces the indoor/outdoor ratios reported in five published studies.A browser app that estimates when home pipes freeze from forecast lows.
All 6 published time-constant configurations fall within the cited bands.A parked-car cabin heating simulator. The only safe interval is zero minutes.
All 16 published cabin cases reproduce within tolerance.A home energy simulator that ranks upgrades by cost and CO2.
10 published benchmark cases pass.Predicts tonight's sleep from caffeine, alcohol, light, exercise and meals, using published models.
Reproduces all 11 published findings it was checked against. Live in the browser; not a medical device.Personal caffeine half-life and the latest sensible time for one more coffee.
Reproduces published half-life ranges for 11 covariate combinations.A native desktop IDE (Tauri, React, Rust) where Pixie, a pixel-art companion, acts out what your AI coding agent is really doing.
Beta; native installers build today. House rule: the animation never lies.A whole desktop environment inside your phone's browser: windows, a virtual filesystem, a terminal and a code editor.
Offline and installable, and live in the browser.Textbook in, Anki deck out, on local models, so nothing leaves your machine.
Alpha, 156 tests passing.A scroll-driven Three.js storefront for a fictional cookie brand.
The hero cookie is a planet.An AlphaZero-style agent for 2 to 6 player Chinese Checkers, with Gumbel MCTS and three research prototypes: coalition-aware search, per-component values mixed in the tree, and a wreath-equivariant network.
324 tests pass on every push.An epidemiological data warehouse on 473,189 weekly disease reports (1916 to 2011), stored three ways: a PostgreSQL star schema, MongoDB buckets and a Neo4j graph, with Kafka and ksqlDB streaming.
13 decision-support queries run on all three back ends side by side.