About the archive
Editorial standards
The encyclopedia works under these rules, several of them written after something went wrong.
- Never fabricate a citation, a page number, a date or a quotation. This is the one rule with no exceptions: a thin sourced article beats a rich invented one. If it was not read, it is not cited.
- A quality score is not evidence that a transcription is faithful. A vision model reading a ruined scan does not reliably fail — it sometimes invents fluent, period-plausible prose, which scores better on every statistical measure than honest garbled OCR. Statistics can show that text is not garbage; only a comparison against the page image can show that it is true, and only that comparison is called verification here. It is why archive search results link to the source file rather than paraphrasing it.
- Cite the printed folio, not the PDF page. Two sources held here have offsets between the two — Scott's 1902 Coast Pilot runs about fifty-one pages ahead of its printed folios, and Sawyer's Volume III continues the pagination of Volumes I–II, so its PDF page 434 is printed page 1363. The folio is what a reader with any copy can find.
- Report failure counts beside result counts. A retriever that turns an exception into an empty result must say how many requests failed. "Zero, with errors" and "zero, with none" are different findings, and a retrieval failure must never overwrite a good earlier result.
- Verify a zero before believing it. Several layers legitimately return nothing for this archipelago. Each was checked against the layer's statewide count first, because a bounding-box mistake looks exactly like a genuine absence.
- Withhold sensitive detail, and count what is withheld. 22 entries in the parish register are held back: cause-of-death detail about named children, and one baptism recent enough that the person is alive. The deaths themselves and their dates stay in the register; only the particulars are removed, and the removal is published as a figure rather than left invisible. A reader who cannot see what was taken out cannot judge what is left.
- Corrections go in the documentation, with their evidence. Claims in this project have been wrong and were corrected in place, in writing, rather than quietly edited away. How to report one.