Machine Learning & AI / Independent exploration
Project Janus
Classical and neural chess-engine research that keeps search correctness, evaluator behavior, and playing strength separate.
Problem
A more accurate static evaluator does not necessarily produce stronger search. Janus investigates how learned evaluation interacts with tree size, position type, and fixed compute budgets.
Architecture
The repository includes a Python baseline and later native C++ research engines. Native NNUE components support accumulator refresh, quantized evaluation, and optional AVX2 paths. The latest inspected native engine exposes neural evaluation as an option while retaining a classical default.
Implementation
The native v0.64 source implements phase-dependent blending between classical and neural scores. Its default disables neural evaluation. The v0.65 report documents factorial comparisons, endgame evaluation, numerical parity checks, and a distinction between frozen classical and neural research baselines.
The inspected adaptive blend scales its base neural weight by material phase: . This is an implemented weighting rule, not evidence of a strength gain.
Engineering decisions
Keeping the classical control available makes neural changes falsifiable. Fixed-depth runtime, fixed-time behavior, static error, and tournament outcomes answer different questions and should not be collapsed into one strength claim.
Results
VERIFIED by source inspection: optional neural evaluation, accumulator-based native evaluation, and phase-tapered blending exist. DOCUMENTED only: tournament estimates, latency numbers, and manuscript-readiness conclusions in repository reports. No official Elo rating or accepted publication is claimed.
Limitations
The report's exploratory comparison against the classical baseline includes uncertainty spanning no improvement. An older README describes only the initial baseline and is not a complete account of the pinned tree. Reproduction must identify a specific native version, model, corpus, and budget.
Reproducibility
Reviewed at commit 12fd92f43618 on 2026-11-09. Source inspection establishes implementation structure; repository-authored results are not independent measurements. The upstream application and benchmark suite were not rerun during this review.
See search and learned-evaluation research for the experimental question.
Next investigation
Resolve the limitations above, preserve the reviewed commit and exact experiment configuration, and reproduce correctness checks before promoting performance or deployment claims.