While extracting formula names from a classical medical compilation, I ran into a problem that a cleanly recognized name could not solve: which book did the passage come from? I had collected quotations attributed to Yanxia Shengxiaofang in a public transcription of Uibang Yuchwi. When I checked the 47 candidates previously labelled high confidence against stricter source boundaries, 20 remained inside those boundaries and 27 fell outside them.
The original aim was to use quotations preserved in another book to recover something of Yanxia Shengxiaofang, then investigate whether the material supported a new formula or textual discovery. Narrowing a list did not make its attributions reliable. In the results dated 18 August 2026, I closed that discovery hypothesis and retained the corrected attribution record and material for further comparison.
Where does one quoted source end?
Uibang Yuchwi brings together passages from many medical books. A source title can introduce a run of text, but the next source may begin before a simple extraction rule notices it. A clearly recognizable formula name can therefore receive the wrong attribution, changing the comparisons built on it.
The initial extraction collected paragraphs after the target title and stopped when it encountered a short paragraph ending in 方. That was used as a sign of the next source heading. Actual headings did not always follow that pattern, and transcription errors made the stopping rule more fragile.
I then checked the candidates against boundaries defined using a separate list of source-title prefixes. Each candidate retained its source URL, paragraph identifier and position, alongside the start and end of the attributed passage. The useful test was whether I could trace a candidate back to a defined stretch of text.
What the high-confidence label missed
The initial results record a reduction from 333 raw candidates to 63 curated entries, including 47 labelled high confidence. That label belonged to the extraction procedure. It did not mean that a specialist had established a formula’s identity or source. The boundary check changed the interpretation of the list.
| Boundary check | Candidates | Meaning |
|---|---|---|
| Inside the explicit source-title boundary | 20 | The candidate paragraph lies within the applied boundary |
| Outside that boundary | 27 | The original attribution cannot simply be retained |
| Original high-confidence list | 47 | The complete set examined in this check |
Removing the candidates outside the boundary did not establish the remaining 20 as new formulas. The same name can occur elsewhere, and different names can conceal overlapping ingredients or contexts. This table answers the narrower question of whether a candidate falls within the source span defined by the rule.
The accompanying catalogue contains 34 source blocks comprising 381 paragraphs and 12,239 characters. A block is a stretch of text grouped under a heading. This catalogue includes short index passages and title-only fragments as well as substantive text, so its block count is not a formula count. Regular-expression matches also include dosing phrases and names run together in the transcription. I kept the material as a starting point for reading the text.
A catalogued manuscript changed the next question
The assumption that the book had disappeared entirely also needed revisiting. CiNii Books records a two-volume manuscript titled 煙霞聖效方 held by Kyoto University Library. That bibliographic record identifies a surviving manuscript; it does not support a claim that I discovered a previously unknown book. CiNii Books record
The catalogue alone cannot establish its relationship to the quotations in Uibang Yuchwi. It could represent an independent transmission or a later reconstruction from quotations; deciding between those possibilities requires comparison of the texts. Formula order, omissions and additions, and differences in ingredient names and quantities all belong in that examination.
What remained from this attempt was a candidate table whose attributions could be checked. The 18 August results confirmed no new medical text or formula and closed the claim of a first reconstruction. Further research needs to begin with the relationship between the surviving manuscript and the quotations.