The Tail-Preserving Alternative is a design proposal by Lee Sharks for language models intended to preserve more distributional variation.
It surveys interventions in training objectives, decoding, corpus weighting, model routing, evaluation, and provenance. It argues that present systems favor predictable and conventionally preferred outputs because those align with common benchmarks and deployment economics.
Rare continuations may be valuable, erroneous, unsafe, or incoherent, so practical systems would need quality and safety criteria. The work does not report a completed tail-preserving frontier model.