MirrorEthic / Research
Systems that watch themselves work
One question across very different systems: can a system be built so that its own workings are legible — to itself and to us — and does that legibility actually govern how it behaves? The programs below attack that from three angles: an architecture built to be read, a measurement of whether retrieved context truly controls behavior, and the cognitive infrastructure that carries it. Open weights, open receipts, nulls on the record.
Programs
Lines of work, one thesis
Program · cl33-opLMCurrent line
A model you can read
A language model whose only route from computation to words is a stream of readable Cl(3,3) operators: a 470M base trained to 20B tokens, a memory it can show and delete exactly, procedures learned from experience, and now images entering through the same route (96.2 % against 21.3 % for the wrong image, closed vocabulary). Still open: whether the algebra itself beats a random basis, and memory in natural conversation.
The cl33-opLM program →Program · Context GroundingNew · interim
Retrieval is not grounding
Pre-registered experiments measuring whether retrieved context actually governs model behavior. A frontier model repeated a corrected fact with 100% accuracy and still decided on the contradicted value 55.3% of the time. Across five models and three providers: only deterministic transformation reliably grounds behavior — annotation and self-report are variable. Nulls and boundary conditions on the record.
Read the result →Program · HologramNew · published
Five pilots, one failed
Before shipping Hologram 0.5, five preregistered pilots on our own repositories: does an adjudicated causal record change what a coding agent builds? Two honest nulls, one decisive separation when the history is absent from the tree (9/18 → 0/18), one failed pilot that changed the product, and the preregistered rerun: ungrounded sessions found the design that held 0 of 12 times; grounded, 10 of 12.
Read the pilots →Program · CAT-4BLive · seven-day challenge
Where does an assistant's behaviour live?
A 4-billion-parameter model fine-tuned on text alone to deliver ordinary help through a cat frame, then measured on a frozen 490-item evaluation under four conditions: base and tuned, with and without the cat prompt. The base put a cat in 5% of replies; the tuned model, 77% with the cat prompt and 78% without it; its image prompts named a cat first 97% of the time. The public challenge serves the same weights with every control visible: the prompt it composed, the render, its own review of the picture, and a counter showing how often the wrapper had to ask again. Retention cost, judge disagreement and the safety finding are on the record.
NO CATS? TOO BAD. →Program · HologramDeployed · live
Self-learning cognitive infrastructure
The runtime the research runs on: learned context routing, unified memory, and a temporal grounding layer that improves with every interaction — the emit-and-check machinery the Context Grounding results are built from.
Hologram Pro↗Cross-substratePrototype
Beyond language
The same instruments pointed where the question demands: satellite anomaly detection across five sensing modalities, cellular-automata substrates for controlled experiments, and vision transfer validated against DINOv2.
Applied systems →Program · T³Earlier line · peer review in progress
Transparent architectures, first generation
The lab's first architecture: a transformer augmented with a geometric-algebra state space (Cl(3,3)) and a per-head regulatory ecology, built to be read like an instrument panel. 260 public traces, open weights and a reproducibility paper under review. cl33-opLM grew out of it, and the architecture work continues there.
The T³ program →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