Precise Evidence for Long-Document Questions / M2
On 360 sealed questions over 60 fresh long documents, the material passed on to answer each question fell from 32.8 items to 2.2 while recall of the needed evidence stayed at 100%; precision rose from 21.9% to 93.1%. Exact answers were given for 359 of 360 questions, none wrong.
SCOPE + LIMITATIONS
- One generator and six question kinds, all on the exact path; estimates and open questions need their own evidence definitions. Material is counted in lines, not tokens. Questions are in the generator’s wording here; independent wording is tested in M2.2 and M2.3.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Generate 60 fresh dossiers and six kinds of question about them. For each question, measure whether all needed evidence is passed on (recall), the share of passed-on material that is evidence (precision), and the amount passed on, before and after the change. Development used 40 separate dossiers; the sealed set was scored once.
All three locked gates passed: no question lost recall, precision rose far above the baseline audit (M1, 19%), and less material was passed on. The one unanswered question is withheld by design.
- Recall of needed evidence
- 100% before and after
- Precision
- 21.9% to 93.1%
- Items passed on per question
- 32.8 to 2.2
- Exact answers
- 359 of 360; 0 wrong; 1 withheld
- Lowest kind after the change
- Multiple facts · 85.8% precision
Question and method
Generate 60 fresh dossiers and six kinds of question about them. For each question, measure whether all needed evidence is passed on (recall), the share of passed-on material that is evidence (precision), and the amount passed on, before and after the change. Development used 40 separate dossiers; the sealed set was scored once.
Recorded results
| Condition | Recorded result |
|---|---|
| Recall of needed evidence | 100% before and after |
| Precision | 21.9% to 93.1% |
| Items passed on per question | 32.8 to 2.2 |
| Exact answers | 359 of 360; 0 wrong; 1 withheld |
| Lowest kind after the change | Multiple facts · 85.8% precision |
Comparative standing and locked gates
All three locked gates passed: no question lost recall, precision rose far above the baseline audit (M1, 19%), and less material was passed on. The one unanswered question is withheld by design.
Interpretation and limitations
One generator and six question kinds, all on the exact path; estimates and open questions need their own evidence definitions. Material is counted in lines, not tokens. Questions are in the generator’s wording here; independent wording is tested in M2.2 and M2.3.
SOURCE PROVENANCE
Precise Evidence for Long-Document Questions / M2
LABORATORY REPORT / 2026-10-28SOURCE CHECKSUM / SHA-256
a0afc060db89bdacf1ef17b471a4e83a8f1574a81c42d1eac8cc09aae85d5f08Public 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.