Long-Document Questions in Independent Wording / M2.1–M2.3
Exact answers to long-document questions had depended on the documents’ own wording: under fixed paraphrases they were given for 0 of 240 development questions. After a conservative rewording step, two sealed sets of 120 paraphrases written independently of the laboratory’s generator were answered 110 (91.7%) and 115 (95.8%) right, with no wrong answer; every miss was withheld.
SCOPE + LIMITATIONS
- One document generator and six question kinds on the exact path. The rewording step has a vocabulary, so unfamiliar phrasings are withheld. It went on by default in the freeze-candidate configuration after the V124.1 rehearsal showed no regression.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Score questions about long synthetic machine-learning dossiers in six kinds: multiple facts, aggregation, multi-step references, conflicting entries, variations between runs, and first occurrences in logs. M2.1 uses 240 development questions in fixed template paraphrases. M2.2 and M2.3 each use 120 paraphrases written by an independent code agent and frozen before scoring; M2.3 covers a wider range of styles. The rewording step acts only when exactly one question intent is recognised.
M2.2 and M2.3 passed all three locked gates on sets frozen before scoring. Misses were withheld rather than answered wrongly: in M2.2, ten variation questions phrased only as “differs only in”; in M2.3, phrasings such as “sole difference”. No gate was changed.
- M2.1 · development, fixed paraphrases (240)
- 0 exact answers before the change
- M2.2 · sealed (120)
- 110 right (91.7%); 0 wrong; 10 withheld or passed on
- M2.2 · question intent misread
- 0
- M2.3 · sealed, wider styles (120)
- 115 right (95.8%); 0 wrong; 5 withheld
- M2.3 · mean recall of the needed evidence
- 98.3%
Question and method
Score questions about long synthetic machine-learning dossiers in six kinds: multiple facts, aggregation, multi-step references, conflicting entries, variations between runs, and first occurrences in logs. M2.1 uses 240 development questions in fixed template paraphrases. M2.2 and M2.3 each use 120 paraphrases written by an independent code agent and frozen before scoring; M2.3 covers a wider range of styles. The rewording step acts only when exactly one question intent is recognised.
Recorded results
| Condition | Recorded result |
|---|---|
| M2.1 · development, fixed paraphrases (240) | 0 exact answers before the change |
| M2.2 · sealed (120) | 110 right (91.7%); 0 wrong; 10 withheld or passed on |
| M2.2 · question intent misread | 0 |
| M2.3 · sealed, wider styles (120) | 115 right (95.8%); 0 wrong; 5 withheld |
| M2.3 · mean recall of the needed evidence | 98.3% |
Comparative standing and locked gates
M2.2 and M2.3 passed all three locked gates on sets frozen before scoring. Misses were withheld rather than answered wrongly: in M2.2, ten variation questions phrased only as “differs only in”; in M2.3, phrasings such as “sole difference”. No gate was changed.
Interpretation and limitations
One document generator and six question kinds on the exact path. The rewording step has a vocabulary, so unfamiliar phrasings are withheld. It went on by default in the freeze-candidate configuration after the V124.1 rehearsal showed no regression.
SOURCE PROVENANCE
Long-Document Questions in Independent Wording / M2.1–M2.3
LABORATORY REPORT / 2026-10-28SOURCE CHECKSUM / SHA-256
0c1f57d40c6a246be4a215e5fa259bc9c49904b6a043df12dd6750d6e4eb4b7fPublic 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.