Research
TASUMER MAF is the independent research house I founded in 2026. It applies classical cybernetics (Wiener, Ashby, Beer) to how AI systems are trained, steered and handed off between people and machines.
- Pre-registered. Protocols and predictions are fixed before the data exists.
- Audited. Fresh-context agents attack the work, and headline numbers are independently re-derived.
- Reported whole. Failed predictions, nulls and corrections stay in the record.
October 2026 · Release candidate 0.1.0
Crossroads
The same reading can appear twice while the underlying state has changed. Crossroads keeps the full field, the order in which it is read and its reference address together, so an exact return can be checked rather than inferred from appearances.
It comprises a reversible state model, a methods paper and an exact-arithmetic laboratory that runs entirely in the browser. In 512 seeded journeys across 16 synthetic fixtures, the complete state came back every time.
July 2026 · Asset-1 · With Araminta K
The Gauge Is the Obstacle
Task identity, cross-family transfer and merge-conflict prediction from adapter weights alone.
A pre-registered bank of 480 LoRA adapters, two model families by six tasks by forty seeds, was read out by weights alone. Raw weights fail to identify the training task. Features made canonical under the GL(r) gauge identify every one. The obstacle is the parameterization, not the information.
Two of five pre-registered outcomes went against the prediction, including a no-transfer prediction refuted by its own triviality control. A pilot correlation of 0.888 shrank to 0.300 at bank scale. A 117-finding adversarial ladder audited the paper.
July 2026 · XR-001
Typed State Beats Prose
A pre-registered measurement of numeric corruption under agent context compaction.
When an agent compacts its own context, the format of the compacted state is usually chosen by convention. Measured at matched token budgets, the format decides how many numbers survive, and most of the loss happens at the moment of writing.
Paired exact McNemar p = 3.524 × 10−21, direction unanimous across three models. The finding now shapes how my teams and agents pass work to one another.
March 2026 onward · Open source
Lattice topology
Computation defaults to the cubic lattice, and nobody chose it. rhombic measures the
alternative: the face-centred cubic lattice, whose Voronoi cell is the rhombic dodecahedron. That
cell contains the cube, with 8 trivalent vertices at its corners and 6 tetravalent vertices as bridges.
| FCC against cubic | Measured |
|---|---|
| Average shortest path | 30% shorter |
| Algebraic connectivity | 2.3 to 2.5× higher |
| Flood fill reach | 55% more nodes |
| Signal reconstruction | 4 to 10× lower MSE |
| Edge cost | about 2× more |
The ratios hold across 125 to 8,000 nodes and compare spatially embeddable lattices only. A pre-registered control shows a random rewire of the same edges beating FCC, at wire lengths no local physical layout can carry. Paper 1 is complete but neither published nor submitted; Papers 2 and 3, including TeLoRA and the 6.1× result under structured weights, are working drafts.
Tools
| Crossroads Laboratory | Exact rational arithmetic in the browser: catalogue, instruction journal, probes and 15 geometry scenes. | Research license |
| Crossroads State | A Python environment that checks recovery of a whole state, using only the standard library. | Research license |
| rhombic | Lattice topology benchmarks and the TeLoRA module, with 827 tests. | MPL-2.0 |
| The Adversary | Adversarial review that reproduces each finding against the artifact before it is reported. | MPL-2.0 |
| LIOTHIL | A research environment scaffold for Claude Code: knowledge injection, orchestration and an epistemic charter. | MIT |
Both released papers are published from the repository and internally audited. Neither has been through a formal venue. More at tasumermaf.com.