Computational models of human and machine cognition. Aetherya publishes the models, contracts and validation work behind every engagement and behind Thesia, so that cognitive simulation becomes a measurable discipline rather than a plausible one.
The benchmark for cognitive simulation
Survey benchmarks check the answers synthetic people give. Umbra checks what they do: choices, response times, mood, confidence, whether one person stays the same across tasks, and whether many add up to a population. Nine tasks on public data, scored against free baselines, a language model and the human ceiling, with our own model under the same rules.
Dan Andrei · October 2026 · Benchmark, v0.1
It scored 0.065.
We ran Thesia on the seven sources of the Artificial Societies Benchmark, pre-registered and scored once. It averaged 0.065, closer to real answers than the figure Artificial Societies report for their own system on six of seven sources. The losses, the method checks, and a run kit that is on its way.
Dan Andrei · October 2026 · Lab note, pre-registered
Pushing the frontier of simulated cognition
Myrias is the cognition model behind every simulated person on Aetherya. This is what it gets right, measured against real people: predicted versus actual shares across three public datasets, every figure scored on held-out participants the model was never fitted to, and calibration reported alongside accuracy. Where a simpler method does as well, it says so.
Dan Andrei · September 2026 · Model card
An authority-preserving runtime contract for persistent AI memory
Transcripts preserve events but not currency, vector indexes find similar passages but not authority, and summaries save tokens while hiding uncertainty and lineage. Anamnesis separates source authority, epistemic status, retrieval utility, visibility authority and executable authority, and forbids authority inflation through retrieval or summarization.
Dan Andrei · August 2026 · Preprint, v1.0
Each line of work feeds the others. The cognition model proposes a mechanism, grounding gives it a population, validation says where it holds, and memory keeps the audience honest over time.
Appraisal, memory, friction and confidence: the mechanism every simulation runs on.
A runtime contract for persistent memory that cannot inflate its own authority.
Population grounding
Sampling personas from published marginals so audiences are drawn, not written.
Human-versus-synthetic studies that report where the model fails, not only where it agrees.
Replications, datasets and critique are welcome.
