230 accessible result rows from 230 authoritative meaningful result rows in the declared scope. Review cards are separate.
All rows from the stated saved result sets are accessible; detailed machine arrays remain in the linked source files.
5 items need review in the preserved curated card set.
This curated view is separate from the complete meaningful results listed above.
Archived research project · completed 2026-09-26
ARCHIVED RESEARCH PROJECT
Does the measured handwriting fall into stable writer groups when compared blind across the manuscript?
OPEN THE REVIEW FINDINGS AS COMPLETEDThe 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
5 source-linked cards. Display images are embedded; original TIFFs and scientific records remain authoritative in the project folder.
source-linked Yale visual comparison
A falling-loop contour remains recognizable on separate folios despite substantial changes in source contrast. Loop closure and tail curvature still vary; no writer or glyph identity is assigned.
source-linked Yale visual comparison
The broader and narrower stacked forms both occur among the f1r proposals. Their loop junctions and proportions differ visibly. Their symbol relationship and structural positions remain unresolved.
source-linked Yale visual comparison
A paired-stem and upper-loop configuration appears in lighter and darker source regions. Compare stem separation, upper closure and attachments before attributing the thickness difference to a hand.
source-linked Yale visual comparison
The arch and short strokes recur, while the number, join and end curve differ among these crops. Candidate boundaries and minim counts remain provisional.
source-linked Yale visual comparison
Longer and shorter whole-crop contours recur. The extra horizontal material is visible, but no word, sequence or writer-specific variant has been accepted.
FINDINGS FROM THE COMPLETED STUDY · public presentation of [preserved evidence reference]
Answer from this automated visual screen: The measured handwriting behaves more like a broad, locally changing population than five clean, stable page groups. There is a modest catalogue-order gradient, much reduced when page darkness, measured stroke thickness, threshold response and character scale are accounted for. This does not establish one writer. The current candidates are too noisy to distinguish reliably between one evolving writer, changing tools/conditions, and multiple writers using a shared system; it also cannot exclude additional writers. See the five Yale-crop review cards or the portable START HERE.html.
The experiment started from all 213 previously verified native Yale TIFFs and the source-linked full_v001 neutral extraction. There are 204 manuscript image surfaces after removing physical covers, flyleaves and exposed edges. Of these, 191 supplied enough compact, locally writing-like candidates for page comparison; 13 abstained rather than receiving guessed page styles. A total of 129,557 provisional regions entered the local writing-like sample; 21,611 native-mask candidate samples supplied usable stroke-thickness estimates. The filter requires plausible region size/darkness and at least two neighboring regions with similar vertical placement. It can still admit illustration and texture and can miss faint writing.
Each candidate retained its independent ID and native box. Sixteen unsupervised, broad shape prototypes were learned from page-balanced contour descriptors. Sixteen is a measurement codebook size, not a writer count or glyph inventory. Within each prototype, page medians captured normalized contour construction, pixel-column slant, width/height proportion, occupancy, height and enclosed-background count. The contour descriptor indirectly reflects loop, tall-stem, crossbar/bench, repeated-stroke, curvature and hook geometry. The experiment did not reliably isolate pen-lift order, exact minim count, hook ownership or symbol identity. Additional page measures used the original-extraction masks and TIFF-derived foreground/background color: local horizontal box gaps, neighboring box-bottom offsets, native-pixel stroke thickness and dispersion, darkness, and strict-threshold response. Their operational definitions are in . Measurements are of candidate regions, not validated characters or lines.
The page comparison used medians within broad shape families, not the frequency of those families. This reduces, but cannot eliminate, content and section effects. Candidate IDs were deterministically divided in half for a holdout check. Hierarchical cluster cuts at 2–8 were inspected without asking for five writers. Admission required silhouette ≥0.12, split-half adjusted Rand index ≥0.55 and at least five pages per group.
| Result | Evidence | Reading |
|---|---|---|
| Natural clusters | Weak two-way split: silhouette 0.129, split-half ARI 0.533 (59/132 pages). Cuts 3–8 had silhouette 0.026–0.028. No cut passed the rule specified in the blind script before execution. | A weak two-way tendency exists; no stable discrete writer groups were established. |
| Independent subsets | Contour, construction, and slant/proportion subsets admitted no stable split; removing obvious material variables did not reveal a robust group. | The weak split is feature dependent. |
| Measurement reliability | Same-page half distance median 1.590, versus between-page median 1.463; only 33% of same-page halves beat the between-page median. | Page style estimates are noisy. This sharply limits any writer-count conclusion. |
| Nonsequential recurrence | Some distant-order pages are close in the measured space, while some adjacent surfaces differ sharply. | There is no reliable discrete style label to demonstrate that a writer style disappears and returns. |
Distances in this table are normalized feature distances, not probabilities of common authorship. The half-page result is a key limitation: weak clustering may reflect poor measurement rather than a truly single population. The page_groups value 1 in the frozen JSON is an undivided placeholder because admitted_cluster_count is null; it is not a one-writer assignment.
Pages within two Yale catalog positions have median style distance 1.367, compared with 1.485 for separations of 40 or more. Pairwise distance rises modestly with catalog separation (Spearman ρ=0.209, page-label permutation p<0.001). Construction-only features show ρ=0.184. The sequence is catalog/manuscript order, not proven writing chronology; disordered production, foldouts, section changes and page reuse could break that proxy.
After regressing measured native stroke thickness, foreground/background darkness, strict-threshold fraction and scale from page vectors, median pair distance falls from 1.463 to 1.350 and the order-distance association falls to ρ=0.083 (permutation p≈0.001). This is a sensitivity analysis, not a causal estimate of pen changes. The four condition measures themselves change appreciably with catalog order (rank correlations 0.44–0.58); imaging, parchment, ink, threshold behavior and candidate mix can all contribute. Stroke width alone was never allowed to create a writer cluster.
The within-page, shape-conditioned execution-dispersion measure shows no directional trend (ρ≈0.001, p=0.989; first/last order quartile medians 0.0761/0.0757). Provisional local gap dispersion and box-bottom offsets become smaller across order, but these depend on layout, section, scale and segmentation; they do not demonstrate learning, simplification or fluency. Stroke-thickness dispersion rises and is strongly quantized by the extraction mask. No supported trend establishes improvement or deterioration of a writer's hand. Individual loop, gallows, minim or terminal-hook trajectories were not accepted: the broad shape prototypes retain different candidate structures, and apparent per-form trajectories would first need human-validated comparable forms.
The comparison uses the public Archetype palaeography archive as a derivative Davis/Currier label source. Its hand labels follow Davis's section-based attributions, and some folios have mixed or absent assignments. The archive bytes are preserved . For background, see Davis's five-hand study, Currier's original papers, and Timm's one-hand critique. Those arguments played no role in making the blind groups.
| Revealed label | Matched images | Between minus within median distance | After condition regression |
|---|---|---|---|
| Davis-derived five hand tags | 185 | 0.107 (permutation p<0.001) | 0.016, p=0.106 |
| Currier A/B tags | 170 | 0.109 (p<0.001) | 0.020, p<0.001, a very small effect |
| Illustration/section tags | 175 | 0.128 (p<0.001) | not used to assign hands |
The raw external associations are real associations in this measurement space; nearby-order comparisons also retain a raw association. They do not supply independent, reproducible five-way handwriting clusters. The archive's hand tags, Currier tags, section and order are heavily coupled: among images with both tags, 94 of 95 hand-1 pages are Currier A; hands 2 and 5 are all B; 25 of 26 hand-3 pages are B. Section/page design affects what shapes and layouts the neutral extraction samples. Currier A/B is a textual/statistical distinction and must not be relabeled as a handwriting result.
The review cards show five concise, source-linked comparisons: a repeated falling loop under changing contrast; two stacked-loop constructions; paired tall stems; arches and short-stroke runs; and longer/shorter horizontal combinations ending in loops. Their crops and contexts come from the preserved , with Yale source hashes and native boxes. For this review, 13 Yale TIFF hashes and all 30 crop/context images were checked against the original decoded pixels. They are intentionally examples to inspect, not statistically selected proof of a writer group. The broad and narrow stacked forms both occur in f1r candidates; neither structural equivalence nor a distinct symbol is established. The full extraction lacks accepted line/cluster positions outside reviewed f1r, and no stable blind writer group exists here. Therefore no form was accepted as a writer-specific variant and nothing was merged.
The strongest supported description is modest, condition-sensitive visual drift with substantial overlap and noisy candidate execution. It does not choose between one changing writer, one writer under varying working conditions, or several similarly trained users of the same system. Several stable writers also remain possible because the measurement layer lacks reliable held-out page discrimination. Human comparison of the five source-context cards is the next check before a more selective, validated repeated-form authorship test. No translation or linguistic interpretation was performed.