Reasoning Planner for Multi-Step Calculations / V119
EMMA’s own small planner, with no language model in the step, answered all 42 multi-step and all 10 direct calculation questions in unseen wording and declined all 15 questions it should decline, with no wrong answer, at a median of 3.4 seconds on the processor. A nearest-example control under the same checks committed 3 wrong answers.
SCOPE + LIMITATIONS
- Unseen wording of trained kinds only; unseen kinds are tested in V119.1 to V119.4. One family of calculation questions. Paraphrases for the dataset were written by a Source Model; plans were composed and checked by EMMA.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Build a dataset of checked plans for generated machine-learning calculation questions, each admitted only when every check passes on that exact wording. Train the planner on the processor and score once on 67 sealed items written in wording never seen in training and frozen before training. The control retrieves the nearest training question and uses its plan under the same checks.
All seven gates passed: at most one wrong commit (0), at least 85% multi-step (100%), direct and declines complete, under 5 seconds, and better than the control on coverage and wrong commits. An earlier development trial with a released reasoning model as the planner answered about 30% of multi-step questions at about 255 seconds each and was not taken to a sealed run.
- Planner · wrong commits
- 0
- Planner · multi-step
- 42 of 42
- Planner · direct
- 10 of 10
- Planner · declines
- 15 of 15; none committed
- Control · nearest example
- 3 wrong; multi-step 40 of 42; direct 8 of 10
- Median time per item
- 3.4 s, almost all of it the independent check
Question and method
Build a dataset of checked plans for generated machine-learning calculation questions, each admitted only when every check passes on that exact wording. Train the planner on the processor and score once on 67 sealed items written in wording never seen in training and frozen before training. The control retrieves the nearest training question and uses its plan under the same checks.
Recorded results
| Condition | Recorded result |
|---|---|
| Planner · wrong commits | 0 |
| Planner · multi-step | 42 of 42 |
| Planner · direct | 10 of 10 |
| Planner · declines | 15 of 15; none committed |
| Control · nearest example | 3 wrong; multi-step 40 of 42; direct 8 of 10 |
| Median time per item | 3.4 s, almost all of it the independent check |
Comparative standing and locked gates
All seven gates passed: at most one wrong commit (0), at least 85% multi-step (100%), direct and declines complete, under 5 seconds, and better than the control on coverage and wrong commits. An earlier development trial with a released reasoning model as the planner answered about 30% of multi-step questions at about 255 seconds each and was not taken to a sealed run.
Interpretation and limitations
Unseen wording of trained kinds only; unseen kinds are tested in V119.1 to V119.4. One family of calculation questions. Paraphrases for the dataset were written by a Source Model; plans were composed and checked by EMMA.
SOURCE PROVENANCE
Reasoning Planner for Multi-Step Calculations / V119
LABORATORY REPORT / 2026-10-27SOURCE CHECKSUM / SHA-256
8efcbbea2b3cbf994756e54de3ad03545102b81757a73df61859b03fb15efa65Public 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.