Long Documents in the Whole Demonstration / V118.2–V118.4
With long documents on, the whole demonstration first failed its memory gate at 19.3 GB against 13.6 GB, because a second resident language-model process did not fit. Answering exact conclusions directly and loading the reader only when needed then answered 36 of 36 questions at a median of 0.01 seconds, and the whole-demonstration rehearsal passed every gate at 13.3 GB.
SCOPE + LIMITATIONS
- One generator and six question kinds in the generator’s wording; independent wording is tested in M2.2 and M2.3. One rehearsal on one desktop computer.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
V118.2 runs the whole-demonstration rehearsal with documents on. V118.3 answers exact conclusions, such as first occurrences, current values, resolved references and counts, directly, and loads the reader only for the rest; it is scored on six fresh dossiers. V118.4 repeats the whole rehearsal.
V118.2 failed the memory gate and the baseline gate with it; V118.3 and V118.4 passed all their gates. For questions with an exact answer, the language-model reader added only error, time and memory. Long documents went on by default after V118.4.
- V118.2 · documents
- 6 of 6 right, routed and cited
- V118.2 · peak memory
- 19.3 GB against 13.6 GB; failed
- V118.3 · fresh dossiers
- 36 of 36 right; 6 of 6 log first occurrences
- V118.3 · median per question
- 0.01 s, against 7.9 s with the reader
- V118.4 · whole rehearsal
- Every gate passed; peak 13.3 GB; mean turn 2.8 s
Question and method
V118.2 runs the whole-demonstration rehearsal with documents on. V118.3 answers exact conclusions, such as first occurrences, current values, resolved references and counts, directly, and loads the reader only for the rest; it is scored on six fresh dossiers. V118.4 repeats the whole rehearsal.
Recorded results
| Condition | Recorded result |
|---|---|
| V118.2 · documents | 6 of 6 right, routed and cited |
| V118.2 · peak memory | 19.3 GB against 13.6 GB; failed |
| V118.3 · fresh dossiers | 36 of 36 right; 6 of 6 log first occurrences |
| V118.3 · median per question | 0.01 s, against 7.9 s with the reader |
| V118.4 · whole rehearsal | Every gate passed; peak 13.3 GB; mean turn 2.8 s |
Comparative standing and locked gates
V118.2 failed the memory gate and the baseline gate with it; V118.3 and V118.4 passed all their gates. For questions with an exact answer, the language-model reader added only error, time and memory. Long documents went on by default after V118.4.
Interpretation and limitations
One generator and six question kinds in the generator’s wording; independent wording is tested in M2.2 and M2.3. One rehearsal on one desktop computer.
SOURCE PROVENANCE
Long Documents in the Whole Demonstration / V118.2–V118.4
LABORATORY REPORT / 2026-10-26SOURCE CHECKSUM / SHA-256
1a3a6c6d4bacf30df48ec20ce5f398591ec653f366724768cdfe32a4234b01dePublic 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.