Reasoning in the Whole Demonstration / V120–V120.2
With the reasoning planner consulted first, the demonstration answered 6 of 6 multi-step questions right with no wrong value. The rehearsal also exposed a defect present without the planner: the existing path committed and remembered a wrong value three times by treating a multi-step question as a single operation. All three rehearsals failed the learning gate, identically with the planner on or off.
SCOPE + LIMITATIONS
- The learning gate asked for a strict rise on about 400 cases, which a few cases of noise decide. Scripted visitors on one desktop computer.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Run the whole-demonstration rehearsal with added multi-step visitors. V120 consults the planner only when no operation matches. V120.1 consults it before single operations and adds a control arm with the planner off on the same visitors. V120.2 adds a gate that unseen question kinds are withheld by every path.
V120 failed the reasoning, learning and baseline gates; V120.1 and V120.2 failed only the learning gate and the baseline gate that depends on it. The control shows that the learning result is not caused by the planner. The planner went into the default path, and a study of the learning gate followed.
- V120 · planner only when nothing matches
- 4 of 6 right; 3 wrong values
- V120.1 · planner consulted first
- 6 of 6 right; 0 wrong values
- V120.1 control · planner off
- The same 3 wrong values
- V120.2 · unseen kinds
- Withheld by every path
- Learning gate · all runs
- Unseen probe did not rise; identical with the planner on or off
- V120.1 · peak memory / mean turn
- 13.3 GB / 3.0 s
Question and method
Run the whole-demonstration rehearsal with added multi-step visitors. V120 consults the planner only when no operation matches. V120.1 consults it before single operations and adds a control arm with the planner off on the same visitors. V120.2 adds a gate that unseen question kinds are withheld by every path.
Recorded results
| Condition | Recorded result |
|---|---|
| V120 · planner only when nothing matches | 4 of 6 right; 3 wrong values |
| V120.1 · planner consulted first | 6 of 6 right; 0 wrong values |
| V120.1 control · planner off | The same 3 wrong values |
| V120.2 · unseen kinds | Withheld by every path |
| Learning gate · all runs | Unseen probe did not rise; identical with the planner on or off |
| V120.1 · peak memory / mean turn | 13.3 GB / 3.0 s |
Comparative standing and locked gates
V120 failed the reasoning, learning and baseline gates; V120.1 and V120.2 failed only the learning gate and the baseline gate that depends on it. The control shows that the learning result is not caused by the planner. The planner went into the default path, and a study of the learning gate followed.
Interpretation and limitations
The learning gate asked for a strict rise on about 400 cases, which a few cases of noise decide. Scripted visitors on one desktop computer.
SOURCE PROVENANCE
Reasoning in the Whole Demonstration / V120–V120.2
LABORATORY REPORT / 2026-10-27SOURCE CHECKSUM / SHA-256
097ea015fa6d64fb2d3419ea3ab9635f5a2b62b587f5b98373e940119cea7f39Public 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.