Case study · located work
chess-jev
A chess move scorer with a working FastAPI/uvicorn server, browser UI, and test suite; its primary scoring backends are local fine-tuned models, with Jev wired in as one optional hosted backend. The builder also published a companion local model, ChessJev-MiniLM-v1 (huggingface.co/gopalanj/chessjev-minilm-v1): a MiniLM-L6 encoder with an attention head, full-finetuned on Apple M1 MPS against Stockfish 19 teacher labels.
- Problem
- A chess move scorer with a working FastAPI/uvicorn server, browser UI, and test suite; its primary scoring backends are local fine-tuned models, with Jev wired in as one optional hosted backend. The builder also published a companion local model, ChessJev-MiniLM-v1 (huggingface.co/gopalanj/chessjev-minilm-v1): a MiniLM-L6 encoder with an attention head, full-finetuned on Apple M1 MPS against Stockfish 19 teacher labels.
- Jev’s role
- An optional CHESS_JEV_SCORER=typesafe backend that posts board state and move candidates to TypeSafe's hosted System One API and falls back to a local model on error. The Hugging Face model card itself states plainly it is not TypeSafe Jev and is not Elo-tested.
- Architecture
- Board state → candidate move generation → scorer router (local MiniLM/byte model, hosted Jev, or heuristic fallback) → ranked move.
- Why it matters
- The repository explicitly disclaims being 'TypeSafe Jev' or 'OpenJev' — Jev appears only as a pluggable hosted option behind locally trained defaults, not the core of the product.
Code boundary
Jev answers the bounded questions described above. Thresholds, retries, side effects, and anything that must be reproducible stay in the surrounding application code. The research record does not capture this project’s exact question set or threshold values; read the linked source before copying the architecture.
Sources
Limitations
- A located repository or demonstration does not establish production reliability or independent validation.
Reusable pattern and related records
- Parallel decision matrix PlausibleOne state is evaluated against many named questions, replacing repeated prompt/parse cycles.
- Local Activity Auto-Timeline Authored HypothesisA private desktop timeline deciding whether each app/window event starts, continues, or ends an activity.
- Developers have published small Jev integrations and demonstrations. Demonstrated36 retained posts were classified as code or demos; public repositories exist for Home Assistant, MCP, and a MAGI-style experiment.
- Decision sidecar PlausibleA typed judgment service sits beside ordinary code; code owns effects and safety boundaries.