Long-Document Answers in the Demonstration / V118
Through the demonstration’s route, 35 of 36 questions on six fresh long documents were answered right with cited evidence, none withheld. The run failed its isolation gate because six correctly worded questions would not have entered the route in the live demonstration; the routing rule was fixed and retested in V118.1.
SCOPE + LIMITATIONS
- The routing cue required a question mark, and these questions ended “Answer with the run number.” The probe still scored them by calling the reader directly. The lesson recorded: test a route’s entry rule on the probe’s real wording.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Six fresh dossiers and 36 questions of six kinds through the demonstration route, with the 97 rehearsal visitor messages checked for wrongful entry into the route.
Accuracy, first-occurrence, auditability and responsiveness gates passed; isolation failed. Against the earlier sealed result of 88.9% (V116), accuracy rose with no training.
- Right
- 35 of 36 (97.2%); none withheld
- First occurrences in logs
- 5 of 6
- Median time per question
- 7.9 s
- Answers citing evidence
- All
- Visitor messages wrongly routed
- 0 of 97
- Questions that would not have been routed
- 6 of 6 variation questions
Question and method
Six fresh dossiers and 36 questions of six kinds through the demonstration route, with the 97 rehearsal visitor messages checked for wrongful entry into the route.
Recorded results
| Condition | Recorded result |
|---|---|
| Right | 35 of 36 (97.2%); none withheld |
| First occurrences in logs | 5 of 6 |
| Median time per question | 7.9 s |
| Answers citing evidence | All |
| Visitor messages wrongly routed | 0 of 97 |
| Questions that would not have been routed | 6 of 6 variation questions |
Comparative standing and locked gates
Accuracy, first-occurrence, auditability and responsiveness gates passed; isolation failed. Against the earlier sealed result of 88.9% (V116), accuracy rose with no training.
Interpretation and limitations
The routing cue required a question mark, and these questions ended “Answer with the run number.” The probe still scored them by calling the reader directly. The lesson recorded: test a route’s entry rule on the probe’s real wording.
SOURCE PROVENANCE
Long-Document Answers in the Demonstration / V118
LABORATORY REPORT / 2026-10-26SOURCE CHECKSUM / SHA-256
28fd057adfaa17c88d6006798a8fb9789a2218ad5244c27de60e549b22882ca7Public 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.