I work on continual learning for frozen models: installing new skills and facts into a pretrained model as small modules that switch on only for their own domain and can be removed exactly. Based in Thessaloniki, Greece.
- assay-research: the research record, with preregistered results, controls, and the bound on every claim, failures included.
- assay-gate: a write-path gate for agent memory that rejects poisoned, contradictory and stale writes (
pip install assay-gate). - assay-memory: memory for local models that folds facts into the weights and revokes them exactly.
Selected results (frozen Qwen3-4B)
- One ~10k-parameter module installs a reasoning operation: 0.73 → 0.94 on held-out domains, above 10-shot prompting (0.78), removed bit-for-bit.
- Five operations run side by side at 0.82–0.96, built from the model's own verified answers, with unrelated skills unchanged.
- An installed fact queried as an isolated lookup inside reasoning is recalled 93% of the time, against 71% when read in place.