Section 7.2.5, “One-frame does not learn restoration; two-frame eventually does”1F stays near 0.075–0.081; 2F rises to 0.99998 test accuracy by epoch 178 (one saved run)
Empirical result Supported by saved experimental runs, not by a proof.
1 Paper text
One-frame does not learn restoration; two-frame eventually does. Figure 5 shows validation restoration accuracy as a function of epoch. The 1F model has low test restoration accuracy throughout training, approximately 0.075–0.081 across saved checkpoints. We do not identify this range with a measured random-legal-move baseline, which is not supplied in the reported evaluation. In contrast, the 2F model learns the restoration mapping and reaches essentially perfect accuracy with enough training. On the test set (Figure 6), the 2F checkpoints achieve 0.209 at epoch 1, 0.538 at epoch 23, 0.876 at epoch 45, 0.989 at epoch 67, and 0.99998 by epoch 178 (remaining at epoch 200).
Generated from paper_jmlr.tex, lines 1104–1108.
2 Explain this
The result. In the saved run ( heaps, bits, examples), the 1F model stays at low test restoration accuracy throughout (about 0.075–0.081 across checkpoints). The 2F model learns: 0.209 at epoch 1, 0.538 at epoch 23, 0.876 at epoch 45, 0.989 at epoch 67 and 0.99998 by epoch 178 (still at least 0.99993 at epoch 200).
What it shows. Under this supervised protocol, transition information makes restoration much easier to learn.
What it does not show. That 1F is at a random-move baseline (no such baseline was measured), worst-case mastery, self-play learning, robustness across seeds, or anything about the asymptotic lower bound.
3 How this is formalised
How this is covered. Empirical, from the saved checkpoint-accuracy file of run 20260214_120801_cd18fc9f. The
independent audit's correspondence row for this section notes that the numerical baseline theorems do not certify
checkpoints or the cause of the plateau. Lean proves nothing about training runs.
| Paper anchor | par-one-frame-does-not-learn-restoration-two-frame-eventually-do-p1 (website anchor; the manuscript has no label here) |
|---|---|
| Displayed as | Section 7.2.5, “One-frame does not learn restoration; two-frame eventually does” |
| Claim | 1F stays near 0.075–0.081; 2F rises to 0.99998 test accuracy by epoch 178 (one saved run) |
| Coverage category | Empirical result |
| Statement source | paper_jmlr.tex lines 1104–1108; SHA-256 of the normalised block source d8e7638567036e61… |
| Audited correspondence rows | sec:supervised-restoration |
| Explanations | explain/c-exp-result-1f2f, formal/c-exp-result-1f2f |
| 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).