Long-Document Evidence Audit / M1
A baseline audit of how the system gathers evidence from long documents found every needed fact for all 240 development questions (100% coverage and recall), but passed on about 31 lines per question of which only 1 or 2 mattered: 19% precision, while reading about a third of each document.
SCOPE + LIMITATIONS
- One generator, six question kinds and development data. The audit set the targets for M2: keep recall at 100%, raise precision, and pass on less.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Run 240 questions over 40 development dossiers with the long-document path frozen as in the demonstration. For each question, count the needed evidence found, the material passed on, and the share of each document read. Repeat runs are compared for identity.
All three audit gates passed; this is a baseline, not an improvement. The count was corrected once before reporting, because the first version left out one source of evidence.
- Coverage and recall
- 100%
- Precision of passed-on material
- 19%
- Lines passed on per question
- 31
- Share of each document read
- 32%
- Aggregation questions
- 100% precision
- Repeat runs
- Identical
Question and method
Run 240 questions over 40 development dossiers with the long-document path frozen as in the demonstration. For each question, count the needed evidence found, the material passed on, and the share of each document read. Repeat runs are compared for identity.
Recorded results
| Condition | Recorded result |
|---|---|
| Coverage and recall | 100% |
| Precision of passed-on material | 19% |
| Lines passed on per question | 31 |
| Share of each document read | 32% |
| Aggregation questions | 100% precision |
| Repeat runs | Identical |
Comparative standing and locked gates
All three audit gates passed; this is a baseline, not an improvement. The count was corrected once before reporting, because the first version left out one source of evidence.
Interpretation and limitations
One generator, six question kinds and development data. The audit set the targets for M2: keep recall at 100%, raise precision, and pass on less.
SOURCE PROVENANCE
Long-Document Evidence Audit / M1
LABORATORY REPORT / 2026-10-28SOURCE CHECKSUM / SHA-256
79490b2bbc2ca5e6daf5b268a6718a6ed85e91913d7eb44cb069219b64aaf1fePublic 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.