How this reader was produced

Validator lineage

Each validator in the chain was built on the outputs of the one before it, and each retains the version, the data slice it was fitted on, and the date it was closed to further change. The lineage matters because a validator that has seen a slice cannot be used to judge performance on that slice.

Only the terminal validators in the lineage were used for the locked tests. Earlier validators remain in the record for auditing, not for scoring.

Frozen Quire O history

Quire O was the development quire. Its history is frozen: the transliteration slice, the fitting runs, and the parameter choices made against it are recorded as they stood at the freeze point and are not revised retroactively.

Because the model saw Quire O during development, no result measured on Quire O is treated as a test result. It is development history and is labelled that way wherever it appears.

Locked Quire F final test

Quire F was held out entirely and opened once. The test protocol, the metrics, and the acceptance criteria were written down and sealed before the quire was opened, so the result could not be renegotiated afterwards.

There was one run. No tuning, no re-selection of metrics, and no second pass. Whatever the single run produced is what is reported.

Semantic holdout H

Holdout H tests whether apparent structure survives when the surface form changes but the underlying arrangement does not. It is a separate axis from the quire holdouts: those test unseen material, H tests unseen representation of comparable material.

H is reported separately and is never merged into the headline figures.

Exploratory versus locked results

Exploratory results come from material the model or the analyst had access to during development. They are hypothesis-generating and are labelled exploratory everywhere they appear.

Locked results come from pre-registered, single-run tests on sealed material. Only locked results carry evidential weight, and they carry it only for the narrow question the protocol specified.

Mixing the two is the most common way work of this kind overstates itself. They are kept apart here by construction.

Representation-sensitivity controls

Transliteration is a choice, and results that depend on which transliteration was used are results about the transliteration. Controls re-run the same analyses over alternative representations and alternative tokenisations to see what survives.

STA1 is an internal representational label used by these controls. It is not a phonemic inventory, not an alphabet, and not a proposed sound value for any glyph. Reading it as phonemes or letters misstates what it is.

What the candidate translation is

The English text in the reader is a candidate translation produced by the Imago method: a continuous reading of the manuscript offered so that others can test it. Read it against the page, the layout, the labels, the images and the paragraph structure, and against independent historical, botanical and pharmaceutical evidence.

Confirmation depends on independent readers and external validation, including serious attempts to find errors. Agreement or fluency alone confirms nothing, and the reading will not be called established or proven until it meets this project's independent validation standard.

In technical provenance the same text is recorded as machine-generated output, not as evidence or ground truth: earlier project history calls it a synthetic reader projection, and every export still carries content_origin generative_ai, generated_output_is_evidence false and a non-factual claim status. Where it reads fluently, that fluency is the generator's, not the source's, until evidence says otherwise.

The reader also makes no claim that Currier A and Currier B are different languages. Those labels are scribal or statistical groupings and are carried here as descriptive metadata only.

The visual glyph layer, source images and provenance

The visual glyph layer records written forms as they appear on the page and where each one occurs. Its identifiers are arbitrary labels for shapes: they are not letters, sounds, alphabet positions or words, and no meaning is assigned to any of them. Where no crop or registered outline exists for an occurrence, the identifier is shown and the asset is declared unavailable; no shape is ever drawn to fill the gap.

Source images are the digitised pages held by Yale University for Beinecke MS 408. They are linked at the source and are never copied or altered here. Crop and coordinate status records whether a location on the source image has been registered; where it has not, the full source page is shown instead.

Grouping of visually similar forms into variant families is a human proposal about appearance only. It is never applied automatically, the default state is ungrouped, and every proposal, approval, rejection and reversal is appended to a history that is never edited or deleted. Reversal adds a new version rather than removing the old one.

Source descriptions are plain observations of what a page looks like. They are kept apart from the candidate translation and are never presented as a reading. Anything produced by a model rather than observed in the source is labelled as generated output and is never evidence or ground truth.