FINDINGS FROM THE COMPLETED STUDY · public presentation of [preserved evidence reference]
Davis–Timm diagnostic-form audit · v001 · 2026-09-26
Narrow answer
Neither interpretation is established. The Yale examples support a reproducible tendency in the execution of two comparable constructions between proposed Hands 1 and 2, even within botanical pages. They do not support a categorical five-way handwriting partition: the proposed Hands 2 and 3 are not reliably separated by the same held-out shape test, including on f115r, where Davis proposes a change after line 12. Timm's objection that the diagnostics overlap is materially supported; his stronger claims that each exact variant occurs on nearly every page, or that one evolving writer made the manuscript, are not demonstrated by this audit. Confidence: moderate about the sampled 1/2 tendency and 2/3 nonseparation; low about actual writer count.
The current 12-pair blind Yale review is embedded in the single portable START HERE.html. All measurements below use original Yale pixels. Published claims, archived search annotations and our measurements remain distinct.
Primary published positions checked
Davis. Her 2022 primary keynote, Voynich Paleography, pp. 2–4, says she initially annotated 225 examples of the one-loop/two-leg gallows on 74 representative pages. For this [k] construction, she asks whether stems have feet or lean, whether the pen appears lifted, whether the crossbar is bowed or level, and whether the loop is round/oval and large/small. Her proposed Hand 1 is bowed with a round loop and slight second-leg foot; Hand 2 is level with an oval loop and rising tick; Hand 3 resembles a compact Hand 1; Hand 4 has a level bar, large loop and marked foot; Hand 5 is tall and narrow with a bowed bar, round loop and tiny tick. Davis explicitly treats a bowed bar as suggestive of continuous movement and a level bar as suggestive of pen lift. Pixels cannot prove movement order.
For the terminal [n] construction, Davis describes the angle, height, length and end position of the returning flourish in [in]/[iin] combinations (keynote p. 3). She describes Hand 1 as high and returning left; Hand 2 as shorter and lower; Hand 3 as tight and nearly closed; Hand 4 as short, open and ending right; Hand 5 as long and low. She reports that long low examples on f66r caused her to reassign it to Hand 5. Her third check, [f], compares the upper junction and loop cant: Hand 1 has a slight left pull and about 45° loop, Hand 2 a second crossbar stroke and more upright loop, Hand 3 a modest bow, Hand 4 a deeper bow, Hand 5 a small round loop (keynote pp. 3–4). Her Figure 1 is a five-hand writing montage, Figure 2 maps hands to sections/quires, and Figure 3 illustrates the alternating Quire 4 bifolia. The keynote specifically assigns ff25/32 and 27–30 to Hand 1, ff26/31 to Hand 2; f57r, f66r and f66v to Hand 5; f57v to Hand 1; ff58/65 to Hand 3; Quire 13 to Hand 2; and f115r lines 1–12 to Hand 2, lines 13–45 to Hand 3 (pp. 5–6). Its figures do not give native coordinates for individual diagnostic glyphs. Davis's 2020 journal article was identified as the earlier primary publication, but its publisher PDF redirected to an access page here; claims and figures reported above were checked against her accessible 2022 primary keynote, not inferred from the inaccessible 2020 figure pages.
The gives the precise page labels used in this experiment. It is transcribed from the archived Archetype project, a Timm-hosted derivative of Davis assignments, not Davis's own complete folio table. It covers 204 Yale manuscript surfaces: 105 tagged 1, 43 tagged 2, 29 tagged 3, 14 tagged 4, 7 tagged 5, two mixed (85v/86r and 115r), and four without a single mapped tag. Foldout panel labels and mixed lines must not be flattened into whole-page hand claims. The archive places Hand 5 on 41r/v, 48r/v, 57r and 66r/v; these entries are derivative and should be checked against Davis if her original table becomes accessible.
Timm. His 2026 primary preprint, One Hand, Five Labels, §§4–5, argues that bowed and level-looking [k] bars, overlap at the stem, and small feet recur across proposed hands and within pages; a visible foot or overlap may contradict a simple one-stroke reading. He says [n] flourish length varies within folios, and specifically compares f2r with f99r, both attributed to Davis Hand 1. He proposes spacing/pace as a cause, and uses the raised plume of [sh] as a further condition-sensitive comparison (§5.3); Davis did not use [sh] as one of her three keynote diagnostics. Timm also objects that hand labels overlap Currier/section distinctions and that a textual statistic cannot independently establish handwriting (§§5–6). His one-writer/continuous-development account is an interpretation, not a pixel observation. His “both variants on nearly every page” assertion is directly image-testable in principle, but needs a validated exhaustive search; the present indexed sample does not supply one.
Yale test and counts
The earlier blind handwriting screen remained frozen. This follow-up used only the verified full_v001 Yale manifest, the existing hand/section comparison code, and the archived Archetype annotations as search coordinates, then read the native Yale TIFF pixels. Archetype's y coordinate is bottom-origin; Yale crop y is image height − archive y. We verified that transformation visually. Archive/Yale dimensions disagreed for 23 annotations on f54v and f90r; those were excluded, not stretched. The resulting selected sample contains 1,427 usable source-linked annotations. It is not an exhaustive glyph census, and source choices may favor the publisher's or critic's expected examples. The 12 review pairs were chosen for comparison, not statistical independence. Nine displayed Yale TIFFs passed file-hash verification; their 48 crop/context images retain native boxes and decoded-pixel hashes in .
| Indexed visible construction | Hand 1 | Hand 2 | Hand 3 | Hand 4 | Hand 5 | Distinct folio labels |
|---|
| [k] two-leg gallows | 202 | 89 | 59 | 15 | 14 | 174 |
| [iin] repeated short strokes + terminal | 258 | 243 | 141 | 27 | 37 | 175 |
| [sh] raised plume + arch | 155 | 78 | 58 | 9 | 10 | 133 |
The table excludes 32 indexed [in] examples that bring the usable archive sample to 1,427. These are counts of selected indexed constructions on Yale pixels, not counts of bowed/level bars, long/short finials or detached/attached plumes. The same broad constructions visibly occur in all five proposed hands. No defensible manuscript-wide count of each specific variant is available from this index. In particular, we cannot verify or reject Timm's “nearly every page” frequency claim. Nor can crossbar shape alone establish the physical number or order of pen strokes.
To test execution instead of form frequency, we normalized each indexed Yale crop to a 24×24 binary shape and used a 16-dimensional regularized shape basis. Folio-held-out folds prevent crop leakage from the same page. This is a targeted two-construction test, not page clustering or writer identification. In botanical pages, Hand 1 versus Hand 2 labels give AUC 0.804 for [k] (236 examples, 108 folios) and 0.860 for [iin] (344 examples, 109 folios). An AUC of 0.5 means no ranking separation; 1 is perfect. Restricting to moderately similar crop contrast (60–100 grayscale units) leaves AUC 0.804 (179 [k]) and 0.841 (257 [iin]). A simple crop-contrast score also associates with the hand labels (direction-adjusted AUC about 0.74–0.76), so this restriction does not isolate a writer cause. Aspect and catalog position also vary. The binary shapes can still reflect nib, pressure, pace, crop placement and page context.
Overlap is substantial: the held-out 1/2 shape score falls on both sides of zero on 46/108 botanical folios with at least two indexed [k] examples and 53/107 such [iin] folios. Zero is a model score threshold, not an adjudicated bowed/level or long/short variant. For proposed Hands 2 versus 3, held-out AUC falls to 0.552 for [k] (148 examples) and 0.529 for [iin] (384 examples). On the especially useful f115r same-page boundary, models trained without f115r have AUC 0.375 for its eight indexed [k] examples and 0.325 for fourteen [iin] examples. Small and selected samples limit precision, but this does not reproduce Davis's proposed 2/3 boundary from the same-form execution tested here.
The review also places Hand 1 f2r against Hand 1 f99r for Timm's terminal-length objection, and two within-f2r gallows examples beside each other. Those visible differences do not require a new hand. The [sh] pairs are an illustrative check of Timm's alternative, not a validated plume-frequency survey. Davis's [f] was researched but not quantitatively reproduced: the archived source index contains no confidently isolated [f] examples, and the neutral candidate boundaries do not establish that identification. We did not relabel an uncertain tall structure to fill this gap.
Context test limit. Botanical restriction partly controls illustration section for Hands 1/2, while f115r holds page, ink surface and much of layout constant for Hands 2/3. Yet Davis hands are still entangled with bifolia, catalog position, Currier textual categories and writing density. The neutral extraction has no validated manuscript-wide line or cluster positions, so a surrounding-structure or cluster-position association would currently be pseudo-precision. No linguistic, semantic or grammatical interpretation was used. Catalog order is not proven writing chronology.
Interpretation and next check
- Several writers: The repeatable Hand 1/2 shape tendency is compatible with more than one writer; it is strongest of the proposed pairwise distinctions measured here. It remains confounded with working regime and page context.
- One changing writer: Same-hand f2r/f99r variation, within-page overlap, and failure to recover the proposed f115r 2/3 boundary are compatible with natural variation. They do not prove one writer or directional development.
- Text/context: Currier and section labels are coupled to Davis tags, and crop contrast/order themselves discriminate labels. This prevents a clean writer-versus-context attribution; it does not establish that a visual variant has linguistic function.
- Unable to discriminate: Exact [k] crossbar and [n] finial variant prevalence, [f] execution, pen-lift order, comprehensive [sh] connectivity and full-manuscript cluster position need validated form-level annotation before strong causal claims.
The current Yale evidence therefore gives more support to Timm's objection to discrete, exclusive diagnostics than to a five-way handwriting separation, while retaining a genuine Hand 1/2 tendency that his one-writer account must explain. Between the complete Davis and Timm interpretations, the appropriate answer is neither, at low-to-moderate confidence. Human blind judgments on the 12 source-context pairs are the immediate next evidence; no form identities were merged.
Reproduce the numeric records with and . and the corrected are versioned. The earlier sensitivity file is retained; its folio-median AUC was invalid because successive contrasts reused mutable row scores, and v002 isolates copies before that calculation. Neither error affected the source crops or the headline folio-held-out AUC values.