Software & Learning Systems / Independent exploration
Search and Learned Chess Evaluation
An experimental study of classical evaluation, neural accumulators, and phase-dependent blending.
Research question
When does blending a learned evaluator with a classical chess evaluator help under a fixed search budget?
Methodology
Project Janus contains a native search implementation and an optional NNUE evaluator with accumulator code. The inspected phase blend scales neural influence with material phase; classical evaluation remains the documented default.
Results
The v0.65 repository report records exploratory comparisons. Its intervals and testing conditions do not establish an official rating or a reliably positive strength gain. The portfolio audit inspected source and the report but did not repeat tournaments.
Limitations
Opening selection, time controls, opponents, seeds, confidence intervals, and independent held-out matches are needed before promoting strength claims. A manuscript described as ready is not a publication.
Reproducibility
Reviewed on 2026-11-09 at commit 12fd92f436186348697a009bc597fe7110a5086d. Evidence is source inspection and repository-authored reporting, not an independent experimental replication.
The Project Janus case study covers implementation trade-offs.