Recursive Reasoning at Small Scale / V106a
A small recursive reasoning model trained on 6x6 Sudoku solved 0.0% of 1,000 sealed puzzles, against 0.9% for a single-pass control, and its stopping signal never fired. All three gates failed. The causes were implementation faults against the published reference, so the result is a defect record, not evidence against the method.
SCOPE + LIMITATIONS
- A diverging loss is a defect, not a result. The lessons recorded: reproduce a reference’s normalisation and schedule before judging a mechanism, and time every arm before locking a budget. The working planner came from a different design (V119).
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Train a recursive arm and a single-pass control of the same size on generated puzzles, and score once on 1,000 sealed puzzles, with and without early stopping.
Failed the recursion-gain, step-saving and stopping-calibration gates. The training loss diverged because the implementation omitted the reference’s normalisation between steps, and its training schedule cost eight times the reference’s. An evaluation bug was fixed and the saved weights were scored without retraining; the failed weights are not kept.
- Recursive arm
- 0.0% solved; 3.4 h of training
- Single-pass control
- 0.9% solved; 2 min of training
- Early stopping
- Never fired; 8 of 8 steps used
Question and method
Train a recursive arm and a single-pass control of the same size on generated puzzles, and score once on 1,000 sealed puzzles, with and without early stopping.
Recorded results
| Condition | Recorded result |
|---|---|
| Recursive arm | 0.0% solved; 3.4 h of training |
| Single-pass control | 0.9% solved; 2 min of training |
| Early stopping | Never fired; 8 of 8 steps used |
Comparative standing and locked gates
Failed the recursion-gain, step-saving and stopping-calibration gates. The training loss diverged because the implementation omitted the reference’s normalisation between steps, and its training schedule cost eight times the reference’s. An evaluation bug was fixed and the saved weights were scored without retraining; the failed weights are not kept.
Interpretation and limitations
A diverging loss is a defect, not a result. The lessons recorded: reproduce a reference’s normalisation and schedule before judging a mechanism, and time every arm before locking a budget. The working planner came from a different design (V119).
SOURCE PROVENANCE
Recursive Reasoning at Small Scale / V106a
LABORATORY REPORT / 2026-10-26SOURCE CHECKSUM / SHA-256
900c1bc6861e83779c663d2e4be13bb7965ffd38454438aef6532e670d79c70bPublic journal edition reviewed 2026-10-28. Source documents and saved evidence were inspected; experiments were not rerun for this edition. Proprietary implementation code, model binaries, private infrastructure, and detailed machine records are not published here. Journal identifiers are editorial references. Catalog inclusion does not imply qualification or runtime promotion.