FINDINGS FROM THE COMPLETED STUDY · public presentation of [preserved evidence reference]

Companion experiment: visual context versus proposed hand

Question. Are recurring, visually similar Voynich candidates better predicted by proposed writer or by their visual surroundings and page setting? This is an exploratory comparison of geometric candidates, not a test of glyph identity or writer count.

Result

Across four usable anchor-pair contrasts, measured context generally predicted the selected visual variant better than the Davis-derived hand tag, but one stacked-loop contrast retained a proposed-hand association even within the Herbal section. The result does not establish that the hand tags identify writers. Section, manuscript order, image condition, content, and hand tags remain entangled. No robust contrast showed independent gains from both hand and structural context after visibly faint support was excluded.

The table reports mean held-out log loss (lower is better). Each of five folds withheld complete numbered folios, including their recto/verso surfaces. “Base” uses section, catalogue order, page darkness, stroke thickness, threshold response, and scale. “Hand” adds the proposed hand to base. “Context” adds provisional line/cluster position, nearest left/right broad shape families, and gaps to base. “Both” adds both sets. Page/folio ID is never a predictor.

The held-out folio bootstrap interval for context gain over base is .056–.097 for curved loop/open curve and .001–.025 for repeated run/terminal curve. The interval for hand gain over base in stacked loops is .003–.028. The horizontal contrast's intervals include zero for both gains. These intervals resample folios from the fixed cross-validation predictions; they do not account for anchor choice, segmentation errors, or alternative folds. The tall loop/crossbar pair was also screened, but only about 2% of its 994 labeled candidates fell in one variant; it was too imbalanced to interpret.

A separate confirms that section and catalogue order carry some signal, while adding the four measured condition variables to section+order changes held-out log loss by less than .001 for each of these four contrasts. The variables remain in the main comparison to control possible confounding; this tiny incremental prediction is not evidence that actual pen or ink conditions have no effect. Candidate fractions also vary between folios, so the model withholds entire folios rather than learning a page ID. It treats catalogue order as an imperfect position proxy, not a writing date.

Four requested patterns and Yale crops

1. Variants across proposed hands with different contexts: The P01/P02 curved-loop sheet shows both anchor-neighbor classes in Herbal pages tagged hands 1 and 2 (47r, 1r, 26v, 46r). In the manuscript-wide selected set, the P02-nearer class appears in 84/521 provisional isolated clusters and 549/999 cluster interiors. Hand adds no held-out benefit after page setting and context. The P21/P22 terminal-curve sheet gives a second context-dominant example across hands 1 and 3. These are geometric contrasts, not certified alternate renderings of one sign.
2. Equivalent measured context with hand separation: The P03/P04 stacked-loop sheet shows four Herbal candidates at provisional cluster interior positions on pages tagged hands 1 and 2 (49r, 22r, 39v, 43v). Within Herbal alone, the P04-nearer class is 416/1,308 (31.8%) for hand 1 and 142/260 (54.6%) for hand 2, across 47 and 10 folios respectively. On a separate five-fold folio holdout, adding hand reduces log loss from .6626 to .6458; adding measured context yields .6674. This is a proposed-hand association, with only coarse context equivalence. Different neighboring material or image conditions may remain unmeasured.
3. Both effects: No reliable example established. With the original broad darkness cut (8), P03/P04 showed both apparent gains: writer .0058 and context .0131 log-loss improvement over base. Its representative crop selection included paper texture and diagram material. Requiring stronger visible support (25) removed the stable context gain; its folio bootstrap interval crosses zero. The and are retained as sensitivity evidence, not a demonstrated “both” case.
4. Neither explanation: P24/P25 horizontal-ending sheet shows long/short relatives on pages tagged hands 1 and 3 (93r, 9v, 104r, 106r). The small held-out gains have folio bootstrap intervals crossing zero for both hand and context. These shapes remain a useful visual comparison, with neither tested explanation supported here.

Each sheet pairs a native-pixel candidate crop with a wider Yale context. gives Yale TIFF path and SHA-256, candidate ID, native box, context box, and decoded-pixel crop hashes. The 16 displayed source TIFF hashes were checked during generation. Some candidates still include attached neighboring strokes; the sheets do not settle segmentation or identity.

Method and limits

The experiment reuses the full_v001 neutral candidate features and 16 broad, unsupervised shape prototypes from the existing blind handwriting screen. Five pairs of independently source-linked f1r visual anchors from the 28-card internal index define the compared outcomes; four are interpretable enough to report in the table. A candidate must satisfy the existing size/writing-neighborhood screen, have support darkness at least 25, lie within a fixed DCT distance of an anchor, and favor one anchor by at least .035. The outcome is which anchor contour is nearer, not a linguistic class. All 204 manuscript surfaces were screened; the four evaluated pairs have hand tags and span 198 images/102 numbered folios. Mixed or absent hand tags abstain from model comparison. Exact inclusion counts and fixed cuts are in and .

Provisional lines come from same-height local neighbors; clusters from normalized horizontal box gaps. These can be wrong in diagrams, labels, foldouts and curved text. Section labels and proposed hands come from the preserved Archetype archive's derivative Davis assignments; they are comparison variables, not ground truth. Catalogue position is not established writing chronology. The page-level pen/ink/condition measures are imperfect proxies and may reflect imaging or candidate mix. Models use regularized logistic regression with fixed predictors and no page ID; losses are out-of-folio predictions, not in-sample fit. The outcome and examples are exploratory choices from the visual index, so numerical advantages should guide closer human review, not certify scribal habits.

No linguistic interpretation, existing transcription meaning, or decipherment hypothesis entered the analysis. This report and its code/data/crops are confined to this directory. It did not read the active Davis-vs-Timm experiment or change canonical project-state files.

Review images

Embedded from the preserved project files (display copies; originals remain in the project folder).