Section 7.1.4, paragraph 1Evaluation by functional correctness on nonzero nim-sum positions and P/N classification
Empirical result Supported by saved experimental runs, not by a proof.
1 Paper text
In this version we report two task-level notions of performance rather than full game-level Elo tournaments. Move-level restoration performance is measured by functional correctness: on a test position with nonzero nim-sum, a move is counted as correct if it produces a successor state with nim-sum zero. Positions with zero nim-sum have no winning move (Theorem 4.4); when reporting functional correctness we therefore condition on nonzero nim-sum positions. Position-level evaluation performance is measured by classification accuracy and -position recall in FSM-controlled shootouts (Section 7.3).
Generated from paper_jmlr.tex, lines 1034–1038.
2 Explain this
Two task-level measures are reported instead of full game tournaments: move-level restoration correctness (does the predicted move reach nim-sum zero?), measured only on positions with nonzero nim-sum, since zero-nim-sum positions have no winning move; and position-level classification by the shootout controller.
3 How this is formalised
How this is covered. Empirical definitions. The mathematical fact they rely on, that positions with nim-sum zero have no winning move, is checked (Theorem 4.4).
| Paper anchor | sec-evaluation-criteria-in-this-version-p1 (website anchor; the manuscript has no label here) |
|---|---|
| Displayed as | Section 7.1.4, paragraph 1 |
| Claim | Evaluation by functional correctness on nonzero nim-sum positions and P/N classification |
| Coverage category | Empirical result |
| Statement source | paper_jmlr.tex lines 1034–1038; SHA-256 of the normalised block source 0280938a752c8762… |
| Depends on | Theorem 4.4 |
| Explanations | explain/c-exp-criteria, formal/c-exp-criteria |
| Build evidence | fresh isolated rebuild, audit report |
4 Lean statement and source
No Lean statement corresponds to this claim. This is an empirical statement about saved experimental runs. Lean proofs do not certify training runs, curves or checkpoints.
Build evidence: fresh isolated rebuild of all 38 bundled modules with Lean (version 4.26.0, started 2026-10-10T15:21:56Z, finished 2026-10-10T15:27:00Z: 38/38 modules compiled, 1588 axiom reports, no errors. See Build for the commands and the independent audit of 2026-10-10 (dated evidence).