ARCHIVED RESEARCH PROJECT · completed 2026-09-30 · Can the fragmented v005 visual families be consolidated into a defensible neutral inventory while preserving variants, rare forms and uncertainty? · Back to Home

PROJECT SUMMARY

QuestionCan the fragmented v005 visual families be consolidated into a defensible neutral inventory while preserving variants, rare forms and uncertainty?
MethodStatus: SEALED as a provisional visual inventory (2026-09-30). v006 is an overlay on the unchanged v005 ledger. It replaces only the family, variant, singleton and segmentation layers for the 114,551 v005 writing forms. It does not assert glyph, sound, word or meaning identity. No EVA, transcription, language hypothesis, translation or phonetic value was used.
Important numbers42,762 visual families · 42,883 variants · 114,551 provisional writing forms
Finding and conclusionStatus: SEALED as a provisional visual inventory (2026-09-30). v006 is an overlay on the unchanged v005 ledger. It replaces only the family, variant, singleton and segmentation layers for the 114,551 v005 writing forms. It does not assert glyph, sound, word or meaning identity. No EVA, transcription, language hypothesis, translation or phonetic value was used.
Limits- QC labels come from the coordinating reviewer (Claude) viewing native crops. They are blind to decisions, but they are not an independent human second rater. A 21–35% unsure rate reflects small, faint and composite crops. - Families, shape modes and singleton categories are provisional…
MeaningStatus: SEALED as a provisional visual inventory (2026-09-30). v006 is an overlay on the unchanged v005 ledger. It replaces only the family, variant, singleton and segmentation layers for the 114,551 v005 writing forms. It does not assert glyph, sound, word or meaning identity. No EVA, transcription, language hypothesis, translation or phonetic value was used.

Read the findings preserved from this completed study.

EXPLORE ALL RESULTS · VISUAL FAMILY INVENTORY

42,762 active provisional families, 42,883 variants, 114,551 provisional writing occurrences. Every family and variant is searchable here. 958 reproducibly selected families have portable Yale-crop sheet previews; every family has a Mac-folder preview and the complete writing occurrence drill-down. Browse all source material classes and observations in the Mac project folder. Family resemblance and counts are machine proposals, not accepted glyph identities.

203,403 unresolved material candidates are outside the assigned-family inventory. 56,561 writing forms have recorded segmentation alternatives; uncertainty remains explicit in the authoritative ledger.

REVIEW NEEDED

6 items need review in the preserved curated card set.

This curated view is separate from the complete meaningful results listed above.

VOYNICH PROJECT

Archived research project · completed 2026-09-30

ARCHIVED RESEARCH PROJECT

Visual Transcription Pass 3 (v006) — sealed provisional visual inventory

Can the fragmented v005 visual families be consolidated into a defensible neutral inventory while preserving variants, rare forms and uncertainty?

OPEN THE REVIEW FINDINGS AS COMPLETED

HUMAN REVIEW · AS PRESENTED WHEN CURRENT

Visual Transcription Pass 3 (v006) — sealed provisional visual inventory

6 source-linked cards. Display images are embedded; original TIFFs and scientific records remain authoritative in the project folder.

Pass 3 consolidated v005 fragmentation without targeting an alphabet size: 42,762 provisional families (was 46,955), 39,213 singletons (was 43,296) and 42,883 shape modes. Fresh blind QC found 92.7% of singleton attachments and 93.9% of within-family pairs the same form; failed family-to-family merges were reverted. Families remain provisional visual groupings that partition continuous variation. Six cards need review; every family and occurrence is in Explore All Results.

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

Visual Transcription Pass 3 — v006 provisional visual inventory

Status: SEALED as a provisional visual inventory (2026-09-30). v006 is an overlay on the unchanged v005 ledger. It replaces only the family, variant, singleton and segmentation layers for the 114,551 v005 writing forms. It does not assert glyph, sound, word or meaning identity. No EVA, transcription, language hypothesis, translation or phonetic value was used.

What we were trying to find out

v005 left 43,296 of its 46,955 visual families as singletons (92%). Pass 3 set out to consolidate that fragmentation into the most defensible language-neutral inventory. It had to preserve genuine variants, rare forms and unresolved ambiguity, without targeting any alphabet size.

What we did

  1. Corrected a schema defect. Pass 3 had found that v005's family graph read two page-feature columns one position off (complexity/holes20 instead of holes12/n_raw). Recomputing the saved calibration features reproduced v005 exactly (max difference 0.0). The pair verifier was refitted with the correct columns on the same 208 blind S/D labels; its 20×5-fold CV precision at P ≥ 0.90 is 96.6% (legacy 97.4%).
  2. Re-ran the unchanged v005 family linkage on corrected inputs. It gave 46,926 families and 43,377 singletons, so the defect was not the cause of the fragmentation.
  3. Diagnosed the mechanism. 8,734 v005 singletons had a verified link (P ≥ 0.90) to an established family, but v005 blocked the link with a fixed centroid constant designed for group-to-group merges.
  4. Froze the rules before any decision or QC (). A singleton joins an established family only when:
  5. at least two of its five nearest members verify at P ≥ 0.90,
  6. it lies inside that family's own measured handwriting-variation envelope (95th-percentile member-to-centroid distance), and
  7. no second family qualifies within a score margin of 0.05.
  8. Blind held-out QC on native Yale crops, with strata and decisions hidden.
  9. Round 1 failed the frozen criterion (within-family "different" ≤ 10%): it measured 13.9%. Pairs joined only by the new family-to-family merges were different 31% of the time. Per the frozen rule, that stage was reverted, and round 1 is preserved as a failed attempt ().
  10. Round 2 used fresh forms, the same criteria and no retuning, and passed.
  11. Re-ran over-broad detection and variant modes with the v005 stage-9 null test, and re-ran segmentation with the unchanged stage-10/10b rules on corrected inputs. Then built, replayed and sealed the ledger.

What we found

Measurev005v006
Writing occurrences (default scope)114,551 (98,240)114,551 (98,240) — unchanged
Families46,95542,762
Singleton families43,29639,213
Variants105,488 near-duplicate groups (distance ≤ 0.33)42,883 null-tested shape modes (41 families have more than one)
v005 singletons consolidated—5,236 (4,795 into established families, 441 into new rare recurring families; 1,660 by the corrected rebuild, 3,576 by envelope attachment)
Retained rare/distinct—6,663 v005 singletons (6,685 v006 singletons)
Unresolved v005 singletons—31,397: 17,485 weak evidence, 9,984 faint/uncertain material, 3,573 segmentation-dependent compounds, 355 ambiguous between families
Over-broad family candidatesreview-only15 flagged, none split automatically
Open or both-attested writing segmentation hypotheses79,17676,664 (3,104 newly resolved, 592 reopened)

Under the corrected rebuild, 1,153 forms that sat in v005 multi-member families became singletons. The net singleton reduction is therefore 4,083.

Blind QC (round 2, fresh forms, reviewer = coordinating session). Rates are shares of decided S+D judgements, with Wilson 95% intervals; U is reported separately.

StratumSameDifferentUnsure
Singleton attachments51 (92.7%, 82.7–97.1)4 (7.3%)29/84
Random within-family pairs62 (93.9%)4 (6.1%, 2.4–14.6)18/84
Retained rare/distinct vs nearest family2 (6.3%)30 (93.8%)10/42
Weak-evidence singleton vs nearest27 (87.1%)4 (12.9%)11/42
Nearest pair across two families28 (93.3%)2 (6.7%)12/42

In round 1, rare/distinct was 23 D of 33 decided; pooled over both rounds, 53 of 65 (81.5%) were different. Segmentation: newly resolved splits and detachments were separate pen forms in 28 of 29 decided blind cells (1 unsure). A still-open sample was 8 separate, 2 one construction and 2 unsure. ,

Handwriting variation (descriptive). The median within-family descriptor distance is 0.622 on the same folio, 0.694 across folios and 0.705 across conventional sections. Page conditions add variation, so family envelopes must be family-specific rather than one constant.

What it means

The v005 fragmentation was mainly an artifact of the linkage mechanism, not evidence of 43,000 distinct forms. About a tenth of the singletons join families with measured ~93% agreement. A further 6,663 forms are genuinely different from any established family. The largest remaining group, 17,485 weak-evidence singletons, mostly resembles an established family by eye (87%), but lacks the verified support the frozen rule requires. They stay separate rather than being merged by a threshold chosen after seeing QC.

The failed round 1 shows that family boundaries are not natural category edges. Merging whole neighbouring families produced wrong members, yet the closest forms across two families usually look alike. The inventory is best read as a stable, auditable partition of a visual continuum, not as a letter count.

Limitations

Readiness for downstream experiments

v006 is sealed and reproducible: stage replay is exact, v005 and all 213 Yale TIFFs are verified unchanged, and the ledger is lineage-complete (). It is stable enough to be the input for character-neighbourhood, recurring-sequence and morphology experiments, with three conditions:

None of those experiments was started.

Data: . Code: [deterministic method], [deterministic method].