Gambit
Post-mortems for people who lose the same way twice
- Runs
- Fully local
- Batch
- ~90 s / month
- Output
- PGN comments
Engines are excellent at telling you a move was wrong and useless at telling you why you played it. Gambit sits between the two: Stockfish provides the evaluation swing and the refutation, and a local model turns that structure into two sentences of plain English.
The rule the whole thing is built around is that the model never gets to disagree with the engine. It explains, it does not evaluate. That single constraint removes almost every category of hallucination that would otherwise make the output untrustworthy.
Everything runs on the workstation, which means a decade of my games never leaves the house. Output is written back as standard PGN comments, so it opens in whatever board I was already using.