ARCHIVED RESEARCH PROJECT · completed 2026-09-25 · Which visual shapes and combinations recur across all 213 Yale images, without assigning identities? · Back to Home
PROJECT SUMMARY
QuestionWhich visual shapes and combinations recur across all 213 Yale images, without assigning identities?
MethodThe source-linked extraction contains 502,286 provisional geometric regions. A size-only retrieval index covers 496,071 of them; the 6,215 outside its broad size range remain in the original candidate and raw-component records. A narrower search stratum of 75,614 regions used native…
Important numbers28 pattern family · 4 r-group observation
Finding and conclusionCurved and looped structures recur. A rounded upper loop with descending tail (P01), an open upper curve with an angular lower stroke (P02), a stacked-loop form (P03), and a curved upper stroke with hooked base (P06) have visually close examples on distinct folios. Loop closure, tail direction and stroke joins vary; the cards keep them separate.Tall paired-stem configurations recur. Upper loops, crossing bars and lower horizontal attachments can be seen in different combinations (P08, P09, P11, P12, P14). These are visible structure families, not a merged “gallows” class. Some apparent components may belong to neighboring material.Repeated short-stroke structures recur at more than one scale. Compact runs (P19), strokes beneath an arch (P18, P20), and wider runs ending in a curved structure (P21, P22) appear on multiple folios. The images do not settle how many separate strokes or units each contains.Some longer visual combinations recur in the same visible order. A loop-like beginning with short strokes and a rounded ending (P23), horizontal structures ending in a closed loop (P24, P25), and a terminal rounded form (P26) are particularly easy to compare in the contact sheets. “Combination” describes the crop; no cluster, word, reading order or accepted identity is inferred.
Limits- Wide tall crops (P07), lower horizontal/tall joins (P13) and mixed multi-line regions (P27) yielded visual neighbors, but some include adjacent material or have modest similarity scores. They are search leads only. - The four f1r detached groups remain separate provisional layout…
MeaningScope: all 213 verified Yale originals, using the completed full_v001 neutral extraction. This is a visual search and contact-sheet review, not a transcription or identification of glyphs. Open the pattern browser.
The one-line dashboard note used while this project was current was not preserved; the cards, categories and controls below are regenerated unchanged from the saved review records.
HUMAN REVIEW · AS PRESENTED WHEN CURRENT
Voynich internal visual patterns
28 source-linked cards. Display images are embedded; original TIFFs and scientific records remain authoritative in the project folder.
P01 · 1 of 28HIGH CONFIDENCE OBSERVATION
Loop with descending curved tail
hook/loop
A rounded upper loop and long falling stroke recur across separate folios.
First manuscript-wide internal visual-pattern survey
Scope: all 213 verified Yale originals, using the completed full_v001 neutral extraction. This is a visual search and contact-sheet review, not a transcription or identification of glyphs. Open the pattern browser.
What was measured
The source-linked extraction contains 502,286 provisional geometric regions. A size-only retrieval index covers 496,071 of them; the 6,215 outside its broad size range remain in the original candidate and raw-component records. A narrower search stratum of 75,614 regions used native geometric size, mask area and foreground/background darkness to make near-shape retrieval practical. Neither stratum is asserted to be “writing.” Raw masks, original TIFF pixels, component alternatives and candidate IDs remain unchanged.
Each indexed region has a strict hash of its native binary support and a separate aspect-preserving shape descriptor. Among 496,071 indexed regions, zero native-support hashes repeated exactly. This is an exact-pixel result, not evidence that visual forms fail to recur: handwriting, boundaries, threshold behavior and background variation easily defeat bitwise equality. Near-shape scores are retrieval rankings, not identity probabilities. Every displayed neighbor retains its own folio, candidate ID, native box, source TIFF link, crop and wider context.
The 28 inspectable cards were selected from f1r visual anchors and searched across other images. They are provisional analyst-curated examples: 18 high-confidence observations of visible recurrence, seven interesting patterns, and three possible leads. “High confidence” refers only to visually observable similarity among the shown crops, never to one accepted glyph class. The cards deliberately show several forms that look related but need not be the same.
High confidence observations
Curved and looped structures recur. A rounded upper loop with descending tail (P01), an open upper curve with an angular lower stroke (P02), a stacked-loop form (P03), and a curved upper stroke with hooked base (P06) have visually close examples on distinct folios. Loop closure, tail direction and stroke joins vary; the cards keep them separate.
Tall paired-stem configurations recur. Upper loops, crossing bars and lower horizontal attachments can be seen in different combinations (P08, P09, P11, P12, P14). These are visible structure families, not a merged “gallows” class. Some apparent components may belong to neighboring material.
Repeated short-stroke structures recur at more than one scale. Compact runs (P19), strokes beneath an arch (P18, P20), and wider runs ending in a curved structure (P21, P22) appear on multiple folios. The images do not settle how many separate strokes or units each contains.
Some longer visual combinations recur in the same visible order. A loop-like beginning with short strokes and a rounded ending (P23), horizontal structures ending in a closed loop (P24, P25), and a terminal rounded form (P26) are particularly easy to compare in the contact sheets. “Combination” describes the crop; no cluster, word, reading order or accepted identity is inferred.
Interesting patterns
Narrower stacked loops (P04) and partially open curves (P05) look related to the first group, but visible differences argue for keeping them separate.
Tall structures vary in apparent stem count, upper-loop closure, crossbar length and lower attachments (P10, P15, P16). The add/remove-component question is visible, but component ownership is not resolved.
Short strokes beside a sweeping curve (P17) and a longer loop-started combination (P28) recur approximately; candidate boundaries differ enough to require source-context review.
Candidate density varies greatly by page: 13 images have more than 5,000 form-scale indexed regions, including foldouts and illustration-heavy layouts. Counts are strongly affected by page area, drawing, texture and segmentation, so they are not normalized glyph frequencies or section-specific writing rates.
A separate found a candidate beginning to the right within 120 native pixels for 52,949 of the 75,614 darker, form-scale regions, under a half-height vertical-overlap rule. Nearest gaps span 0–120 px (median 25 px; 90th percentile 82 px). This is geometric proximity, not a word, cluster or line sequence. The broad spread supplies no justified universal space threshold.
Possible leads and limits
Wide tall crops (P07), lower horizontal/tall joins (P13) and mixed multi-line regions (P27) yielded visual neighbors, but some include adjacent material or have modest similarity scores. They are search leads only.
The four f1r detached groups remain separate provisional layout objects. Their annotated gaps from the main-line boxes are R1 341 px, R2 509 px, R3 443 px and R4 450 px. P23 shows relatives of a loop-started combination visible near R4, not a confirmed match to the whole R4 group or its placement. No full-group recurrence is established for R1–R4.
The full neutral collection does not have accepted line, cluster, label, detached-group or word boundaries outside reviewed f1r. Therefore this survey does not claim cluster-start/end associations, line-start/end effects, label vocabulary or sequence n-gram frequencies. The longer crop recurrences above are whole-crop visual observations. A later uncertainty-aware layout/sequence model must test positional claims without silently imposing f1r's paragraph model on diagrams and foldouts.
Candidate size and darkness strata may underrepresent faint ink, unusual large structures and tiny modifiers. The original TIFFs and all original extraction artifacts preserve those cases. Some visually compelling neighbors may be illustration or threshold artifacts. Human inspection of the crop and wider context is necessary before any recurring class or sequence is accepted.
Provenance and validation
The , , , , per-page and reproducible scripts ([deterministic method], [deterministic method], [deterministic method], [deterministic method]) identify source paths, hashes, coordinates and thresholds. This derivative survey made zero accepted identity, adjacency or linguistic assignments and did not alter full_v001 or earlier datasets.
Recommended next action: Jeremy and ChatGPT inspect the 28 cards, especially the wider contexts for the tall, minim and longer-combination cases. Record which visual distinctions remain useful before designing an uncertainty-aware manuscript-wide positional/sequence analysis. No external corpus or language interpretation was used here.