MirrorEthic / Research / T³
The T³ program
T³ augments the transformer with a geometric-algebra state space (Cl(3,3)) and a per-head regulatory ecology — and asks one question across substrates: is there a common coordinate system for adaptive systems? Language models, cellular automata, satellite telemetry, and quantum hardware, measured with the same instruments.
Plain terms
What this program is actually making
Most AI models are black boxes: you can measure what goes in and what comes out, but not what's happening inside. T³ is a transformer built to be read like an instrument panel. As it runs, every layer reports a small set of gauge readings — how spread out its attention is, how fast its internal state is moving, how close it sits to instability — and the model uses those same readings to regulate itself, the way a thermostat uses temperature.
The bet is that those gauges aren't specific to language models. Point the same instruments at cellular automata, satellite telemetry, or quantum hardware, and if the readings behave the same way across systems that different, that's evidence for a common coordinate system for adaptive systems — one set of dials for anything that senses, acts, and corrects itself.
Where it stands: 260 public traces, an open checkpoint anyone can run, and a reproducibility paper under review at TMLR. And the program has forked: the newest branch stops adding instruments to a standard transformer and instead builds a model that can only compute through channels you can read. Everything below is one of those artifacts or that fork.
t3atlas.devLive · public
The public research atlas
An open interpretability library: 260 traces across 3 substrates, a research-grade benchmark suite with parameter-efficiency and compute-frontier panels, and two working demonstrations — a Cl(3,3) Turing CPU and the T³ Logic Machine, a neuro-symbolic system sharing one geometric substrate end-to-end. The trace library is the citeable artifact; the viewer is one consumer.
t3atlas.dev →Reference releaseApache-2.0
Open weights & code
A public inference implementation with a trained checkpoint (t3-124m-v36) on Hugging Face. Not a demo repo — the same architecture the research runs on.
huggingface.co/mirrorethic →Peer reviewIn progress
Reproducibility, on the record
A reproducibility study (Watson & Sutherland) is under review at TMLR, with shot-level data from real QuEra Aquila quantum-hardware runs included in the reviewer package.
The benchmark record behind it↗Cross-substratePrototype
Beyond language
The same architecture applied where the question demands it: satellite anomaly detection across five sensing modalities, cellular-automata substrates for controlled experiments, and vision transfer validated against DINOv2 features.
Applied systems →The forkIn preparation
cl33-opLM — transparent by construction
T³ reads gauges on a standard transformer. The fork goes further: a language model whose only path from computation to output is a stream of readable Cl(3,3) operators — so interpretation isn't something you do to the model afterward, it's the only way the model can compute at all. Same substrate as the Turing CPU and Logic Machine demos, taken to a full LM. First paper in preparation.
A re-run of one of our own paper-era results reversed it — statistically significant, the other direction. It's documented in the public record, not buried. That is the ethic in the name.
Lab policy: internal ambition, external rigor