The research program behind WOMO.

Zero-Forget Continual Learning

Enterprise models must learn daily without erasing what they knew. We present a consolidation architecture that achieves backward transfer of approximately 0% across 14 months of live market updates. The method requires no replay of raw customer data, preserving the zero-egress deployment contract.

Measurement protocol described in the paper; full text under briefing.

Causal Inference at Enterprise Scale

We operationalize Pearl's do-calculus over a 119K-edge causal graph, answering interventional and counterfactual queries in under 150 ms. Identification is checked mechanically before any estimate is returned, so unanswerable queries fail loudly rather than plausibly. We report calibration against 4,100 held-out natural experiments.

Calibration protocol described in the paper; full text under briefing.

MIRAGE: A Benchmark for World Models

Existing benchmarks grade how convincing output looks; MIRAGE measures coherence: whether a system holds one consistent internal world, a state that moves under rules and reacts to intervention, rather than a fluent description of one. Thirty-one task types across thirteen domains, procedurally generated from private seeds. Answers are computed by a simulator and independently re-derived by a separate solver, so no model and no judge sits in the grading loop, and there is nothing to memorize. Because it grades states rather than text or pixels, language models, state-space models, generative world models, and energy-based scorers are measured on the same items.