FINDINGS FROM THE COMPLETED STUDY · public presentation of [preserved evidence reference]
Voynich 9-like construction and scholastic Latin abbreviation — v001
Completed: 2026-09-28, America/New_York. Verdict: UNRESOLVED. 17-card pixel review · · .
Answer
There are real, recognizable historical parallels. The closest are medieval con/com/cum abbreviation signs, final -us and related ending signs, and compounds such as 9 plus an additional mark for contra. These conventions replaced omitted letter sequences, sometimes an entire word, in otherwise readable writing. They were not medical signs or necessarily a special shorthand alphabet.
The direct comparisons establish visual resemblance, including a particularly close loop-and-descender construction in LJS 484. They do not establish that Voynich uses the same abbreviation operations. Shared scribal machinery remains a credible possibility; simple geometry remains a viable explanation. This study cannot rank their historical probabilities. The requested choice is therefore unresolved, rather than “likely shared machinery” or a demonstrated coincidence.
The convention is widespread medieval Latin writing machinery, also adapted to vernacular writing. These witnesses do not establish a northern Italian or university-specific origin. No Voynich readings, language theories, existing transliterations or proposed translations were used.
1. Voynich observations before historical interpretation
The descriptive layer uses the independent neutral index of 496,071 provisional regions from all 213 Yale source images. Nine already preserved Yale exemplars, three each from internal cards P01, P02 and P06, define the search. Their native boxes and source hashes are in seeds.json. The retrieval, geometric contexts, policy and input-feature hashes were sealed in visual_frozen.json. The analyst knew earlier studies and general paleography: this is a shape-only computational layer, not an independently blind experiment. Historical material did not supply retrieval labels or values.
Variants and relationship to the paired/S family
- Closed upper head + descending tail (P01): rounded, angular, partly filled and imperfectly closed heads; straight-to-curved tails, different tail lengths and sweep directions. Some tails extend under preceding ink.
- Open upper curve + hooked lower end (P02): a broad upper arc ending in a substantial low hook. It is not automatically a broken P01 loop.
- Curved head + thickened/hooked lower end (P06): overlaps the P02 description. The two retrieval neighborhoods are not accepted separate identities, nor accepted variants of one identity.
- Double-loop/S-like constructions: a second lower lobe, enclosed or open, distinguishes their whole contour from a single upper loop with a slender tail. They share reusable curve/loop material. No primary strong P01/P02/P06 region is also a strong target in the earlier paired-form retrieval, under those particular policies. This algorithmic separation proves neither separate characters nor an evolutionary relationship.
Cards N01–N04 preserve these distinctions. The ordinary two-part g in Ferrara's generale is a useful historical control for the double-loop family (N12); being g-like does not make a construction an abbreviation.
Frequency: measured retrieval, not a glyph census
The initial gates admitted 80,197 regions: area ≥130 native pixels; height 24–160; width 10–240; saved mean darkness ≥25; sufficient darker-mask support. DCT similarity uses aspect/area penalties and a normalized-support IoU gate of 0.38. Scores ≥0.88 are “strong”; ≥0.82 are “broad.” These are reproducible search thresholds, not identity confidence.
| Best-matching neighborhood | Strong candidates | Additional broad candidates | Source views with strong candidates |
|---|
| P01, closed head / falling tail | 267 | 308 | 107 |
| P02, open curve / hook | 14 | 136 | 14 |
| P06, thickened lower hook | 11 | 59 | 10 |
| Union | 292 | 503 | — |
Assignment to the best seed is bookkeeping. Membership is nonexclusive: at the broad threshold P02 matches 152 regions and P06 71, with three overlapping memberships. These numbers cannot establish the relative frequency of actual visual variants.
The post-audit sensitivity run retained the same P01 seeds, admitted smaller regions and relaxed IoU to 0.25. It found 576 candidates on 147 views at 0.82; 815/163 at 0.78; 1,153/182 at 0.74; 1,613/192 at 0.70. Lower thresholds increase uncertainty. Fixed-random samples are retained, including a definite non-target in the 0.74–0.82 sample. They do not calibrate recall.
An independent source inspection found a clear missed 9-like form on f001r, native box [686,629,729,692] (N05). Only one P01 region on that source meets the initial policy, despite other visible relatives. Faint heads/tails, differently shaped relatives and motif-containing wider regions are omitted. Thus the actual manuscript frequency is not established. The supported finding is broad recurrence, with explicitly incomplete counts. per_surface_counts.csv records every source view, including zero-hit views. Partial/overlapping views are not independent folios and can repeat material; a zero is not evidence of absence from the source.
Position and neighbors
Native contexts show many P01 examples at the right edge of visibly spaced groups, among otherwise ordinary-sized forms. The random context sheet includes rounded low traces, repeated short strokes, horizontal/bench material, tall loops/crossbars and paired constructions nearby. N17 gives three source-linked examples. These are visible neighbors, not assigned characters or word sequences.
Geometric position is highly sensitive to how sweeping descenders and overlapping boxes are handled. For the 267 strong P01 candidates, using a gap of 0.4 times neighboring height:
| Model | Beginning of proximity group | Middle | End | Isolated |
|---|
| Initial non-overlapping-box model | 75 | 23 | 48 | 121 |
| Alternative admitting overlapping boxes by centroid direction | 53 | 62 | 108 | 44 |
Neither group model establishes words. Thresholds 0.25 and 0.6 are also preserved. The reversal is a substantive limitation, not a result to average away. The horizontal-band proxy calls 224/267 P01 candidates interior, 20 near the left edge and 23 near the right; it is not an accepted line model, especially on diagrams. No reliable paragraph/line-initial frequency or superscript rate follows. In the initial valid-neighbor pairs, candidate bottoms are more often below than above nearby bottoms (285 versus 110; 101 close), but mixed heights and line ambiguity prevent a baseline classification.
Attached versus independent
A separately extracted 9-like region may still touch, overhang or belong with preceding material. Conversely, a wide region may contain a 9-like terminal that this whole-region search misses. N04 shows f113r and f082r wider constructions beside a separately retrieved example. Some source contours touch or sweep under adjacent ink; their component ownership and boundaries remain provisional. Neither extraction connectivity nor a visible space establishes an independent written unit. No exhaustive attached/independent ratio is claimed.
2. What the historical forms actually did
Historical words below are contextual analyst readings, except the explicitly identified National Archives examples. They are not authenticated critical transcriptions of the Penn pages. Square brackets indicate supplied material only on the historical side. Each interpretation retains the native word/line context; the terminal in the syllogism term is deliberately unresolved.
| Requested witness | Direct observation and interpretation | Evidential limit |
|---|
| LJS 56 f.27v, 1420, Padua? | Baseline 9 starts [con]seque[n]ter; a terminal 9-like mark completes cui[us]. Same overall geometry can serve prefix and ending operations in one hand. | The final mark descends conspicuously here, unlike the usual textbook superscript. Height alone is insufficient. N06–N07. |
| LJS 56 f.37r | A 9-like terminal recurs after sillo-, in the syllogism term. Its precise division between retained g, truncation sign and compound remains unresolved. | Do not label it secure -us from silhouette alone. It demonstrates the danger of pooling abbreviation marks and letters. N08. |
| LJS 457 f.62r, Perugia, 1475 | Baseline loop in [con]venit; loop plus additional overbar in [contra]dictoria. | The second is a larger conventional construction: flattening it into plain con or -us loses relevant evidence. N09–N10. |
| LJS 484 f.23r, Ferrara?, 1450–75 | An independently written, full-sized baseline 9 in con seco, in Italian vernacular veterinary prose. Ordinary g in generale is different but relevant to the paired family. | Not a Latin -us ending, not a scholastic-genre token. Ferrara is tentative. N11–N12. |
These pages are available as full-resolution OPenn TIFFs, not merely the medical study's reduced images. URLs, dimensions, foliation and SHA-256 are retained in source_ledger.json; unscaled specimen/word crops are in asset_ledger.json. Ten direct two-column Yale-versus-historical strips put the sign above its wider context. Crops are unretouched; the composite independently fits each side for display and does not assert equal physical scale.
Reference rule: Cappelli distinguishes baseline con/com/cum/cun material (§3.2) from a usually raised final sign for us/os/is/s (§3.3). Added marks can form contra. A similar internal mark can represent omitted er (§§3.32, 4.31). These are substitutions for omitted letters; context and position identify the operation. Isolated cum and joined prefix uses can coexist. Cappelli, Elements, printed pp.14–15, §§3.2–3.32; pp.39–40, §6.2. N16 embeds only the short illustrated examples. The saved 1982 Kansas booklet was acquired from a mirror after Kansas returned an HTML challenge; the challenge and acquisition limitation are preserved. The translators explicitly caution that its practical rules are more dependable than its historical-origin classification.
The prior local Cappelli metadata also records contradictoria constructions (IDs 963, 968, 3513) and consequenter (2782). Those entries are corroborative dictionary examples, not transcriptions of these witnesses or proof of a Voynich expansion. No new all-dictionary ranking was performed.
3. Minimal wider check: broad convention, local rendering
- Germany, LJS 483 f.66v, Ingolstadt c.1480: a baseline prefix loop in contextually read [con]trarie, in university physics notes (N13). This is a dated-period German witness of the same operation. Its place/date are catalogue estimates. OPenn catalogue.
- England, Little Domesday Book, 1086, E 31/1/3 ff.311v–312: the National Archives identifies a raised 9-like ending in edric[us] (N14).
- England, final concord, 1199, CP 25/1/14/4: the National Archives supplies a prefix in [con]cessit and a final ending in ip[s]i[us]. N15 shows the first. The tutorial's initial descriptive reference omits a slash; its transcription gives CP 25/1/14/4. These are administrative/legal controls, not university books. National Archives tutorial, PDF pp.1, 22–23 and 79.
This is enough to reject geographical or university exclusivity. It is not a balanced survey measuring where the signs were most frequent. Scholastic technical vocabulary can make abbreviation conspicuous, but density or stylistic preference in northern Italy was not tested. Perugia is central Italy. The Padua pages are in the second hand, attributed by the catalogue to Johannes de Beylario, a Carmelite from Cologne studying in Padua. Place of copying is not a simple scribe-origin marker. LJS 457 is catalogued as humanistic cursive; LJS 484 as humanistic minuscule, in Italian. Penn LJS 56 orientation, LJS 457 orientation, LJS 484 catalogue.
4. Visual versus functional resemblance
| Comparison | Visual resemblance | Functional resemblance to Voynich |
|---|
| Baseline con sign, especially Ferrara | Close upper loop and descending sweep; independent and joined historical uses | Unresolved: no Voynich omitted sequence is independently known. Apparent group-final Voynich examples need a different explanation from an ordinary prefix, unless it is a different operation. |
| Final Latin -us sign | Clear component/whole-contour relatives; Padua's long descender reduces the size/height contrast | Apparent group-edge use is compatible with many operations, including ordinary characters. Similar final placement does not establish abbreviation. |
| Barred contra construction | 9 component shared | The full construction includes an extra mark and readable context; no corresponding Voynich function shown. |
| Ordinary g and paired/S family | Shared lobe/curve material; sometimes convincing whole contour | Historical g is a letter. No reason to transfer its value or merge the Voynich families. |
Shared machinery would require more than a 9 shape: independently segmented repeatable bases with and without the motif, consistent effects of size/placement/added marks, and evidence that the motif replaces material rather than simply recurring as part of a written sequence. Earlier similar-base leads do not yet supply that evidence. No transformation or omitted content was demonstrated here.
Final choice: UNRESOLVED. The historical lead is now specific and functionally understood; the Voynich relationship is not. The evidence supports recognizable common scribal repertoire as a comparison, while leaving shared operation and coincidental geometry open.