# Scientific Writing 1.1.0 qualification

Qualification date: 5 October 2026. AI-assisted source inspection and bounded
helper execution; no independent human scientific review is claimed.

| Check | Observed result |
|---|---|
| Entry point | Skill validator passed; version 1.1.0 |
| Inventory and links | 44 resources, 77 local links and 38 caller relationships passed static coverage checks |
| Helper behavior | 25 tests passed on staged source; the same 25 passed from a clean ZIP extraction |
| Archive | 44 files; 472,303 unpacked bytes; 182,599 archive bytes; path safety and source-to-archive equality verified |
| Baseline replay | 1.0.1 and 1.1.0 assembled the same explicit synthetic formatting input |
| New visual example | Three DOCX pages rendered and visually inspected; table, headings, units and arithmetic checked |
| Privacy and provenance | Full text and DOCX XML/relationships inspected; no private workstation source path or credential match; component inventory retained |

The 25 tests cover identity-only versus claim-support records, exact title
matching, legacy migration, duplicate citation keys, failed refreshes, non-DOI
records, escaped editorial views, attributed decisions, source preservation,
hidden/tracked content, numerical signs and tiny values, clean DOCX assembly,
unverified citation rejection and overwrite/missing-image failures.

Python 3.10 with python-docx 1.2.0, PyYAML 6.0.2, lxml, openpyxl and pandoc was
used for helper execution. Rendering used the bundled LibreOffice inside an
existing Ubuntu 24.04 compatibility container because the host's older glibc
could not load that bundled binary. All three rendered pages were inspected.
This is alternative-renderer evidence, not native Microsoft Word qualification.

Tests were designed from the inspected source, without independent or blinded
evaluation. They do not establish independent prose generation, broad writing
quality, scientific validity, author satisfaction or submission readiness.
Native agent discovery has not been retested for this edition. Earlier native
checks and the older 16-page example retain their original scope.

Numerical screening matches tokens, so an unrelated identical value can pass.
Publication screening is heuristic and supports selected file types and hidden
structures; figures, scientific meaning and unsupported formats require separate
review. Offline non-DOI attestation checks validate record structure and do not
fetch or authenticate the claimed source. Editorial views reproduce recorded
human responses and cannot create or authenticate approval.

The demonstrated arithmetic uses two explicit synthetic series and no
inferential test. It supports the formatting demonstration only.