How the team turns manuscript photographs into a checked diplomatic text. The same 13 stages run for every archive; each project keeps its own rules file, sign key and project log. Nothing here is AI-only: every spot without consensus goes to a human, and the whole text is read through against the images by human readers.
Diplomatic text. The transcription records what is on the leaf — struck text, insertions, underlining, spelling — not a corrected or modernised version.
Arithmetic may choose, never supply. Where a figure is unclear, a calculation may choose between readings the ink allows; it never produces a reading the ink does not support (Parrish rules T16/T17).
| # | Stage | What happens |
|---|---|---|
| 1 | Load config | The project's rules file, sign key, folder map and engine roles are loaded. Nothing is carried over from another project. |
| 2 | Intake | Photographs are listed and copied to working copies; the originals are never edited. Every image gets a row in the shelfmark workbook. |
| 3 | Item map | Many leaves are photographed several times (overview, full shot, details). Each photo is mapped to its item; the sharpest full view becomes the reading copy; the others are cross-checks. |
| 4 | Rotation check | Top-priority rule: an image is right only if the text reads left to right, right side up. Every image is checked by eye with the 4-angle test and signed off on a contact sheet (green OK, blue fixed, yellow flag, red error, purple recheck). |
| 5 | Detail crops (prism) | Text that runs in several directions on one sheet is split into an upright whole-leaf file (_a) and rotated details (_b, _c…), each read separately and reassembled. Nothing is read sideways. |
| 6 | First pass | The primary reader writes the base text, leaf by leaf, with a Word file per image. Unreadable letters are never left blank: the letters seen are given, with ? for each letter not read, e.g. [c??dence]. Low confidence means pause and re-test, never stop. |
| 7 | Transkribus | Images are uploaded to Transkribus (HTR). Only its exported files are used (PAGE XML line outlines and text); there is no live link. Its letter-by-letter reading guards against word-level guessing. |
| 8 | Other readers | Blind readers from other model families (ChatGPT/Codex, Gemini) read the same images with anti-invention gates: lines first, letters before words, no outside knowledge, refuse rather than guess. They sort disagreements; they never decide a reading. |
| 9 | Decision Table | Fixed in advance: which reader outcomes change the text, which go to a second reader, which go to the PI. Two models agreeing is never proof. Every name, number and date gets a letter-by-letter check even when all readers agree. |
| 10 | Blind re-reads / panel | Disputed spots get blind letter-by-letter re-reads (per-letter CLEAR / CLASS / NOT SEEN), then a two-reader blind panel with an adjudicator. Every change is checked on the full-resolution crop before it is accepted. |
| 11 | PI review | Everything without consensus goes to the PI in the viewer: one spot at a time, with the close-up, the readers' readings and a blank to type. Then the whole document is read through against the images by at least two human readers. |
| 12 | Apply to DOCX | Rulings are applied as new versions of the Word files (never edited in place), with a comment per change: ink evidence, other readers, confidence. |
| 13 | Logs | Every step, time and reader goes into the project's Excel log: images, derivatives, leaves, sheets, questions, answers, readers (engine, exact model, settings), changes, decisions, sessions. Nothing is deleted; superseded files move to a holding folder the PI empties. |
Any engine can be the primary (base text), secondary (blind reader) or adjudicator; the roles are set per archive in the viewer's Engines tab and the comparison scripts take any assignment. Current engines: Claude Opus (primary in all three archives), Claude Sonnet (mechanical steps), ChatGPT/Codex and Gemini (blind readers run by the PI), Transkribus (HTR, exports only). The exact model, settings and dates of every reader are recorded in each project log.
Each archive works under its own written rules, kept in its Box folder and dated: the Ada Lovelace Edition's process rules (the origin of the method, 2026-07/08), the Mary Somerville Papers' working rules (v14, 3 Nov 2026, §2a–§2m), the Parrish archive's rules (T1–T17) and the cross-project working rules. Where rules differ, the newer PI ruling wins. Names, facts and collaborators are never carried from one project to another.
Photographs, transcriptions, manifests and answers stay in the project's Box folder. This website holds only the viewer. Each archive has a IIIF Presentation 3 manifest (one canvas per leaf, images pointing at the Box working copies, one annotation page per leaf for the lines and spots) so the data can later be served by any IIIF viewer without rebuilding it.
Sources: ALE READ ME and generalized process rules; MSP Working Rules v14; cross-project working rules (3 Nov 2026); Parrish Leaves Workbench v21 (Process, Rules, Verification tabs); Collatorium feature list v2 (4 Nov 2026).