Voynich | edkSG Research Volumes · EDKSG-VOY-V021 · Edition 1.0.0 · 5 September 2026, Singapore
Research status: executed fixed-coefficient aggregate-cancellation audit of the three Volume 13 target pages under the split and exclude-touching representations. No model refitting, source correction, independence claim or decipherment is made.
The total barely moves.
The evidence underneath it moves a great deal.
Under the primary split representation, the fixed common model scores approximately +5.693 across f34r, f77r and f105r. Under the stricter policy that excludes pairs touching doubtful seams, the total is approximately +5.734.
The difference is only +0.041.
That small change could easily be read as evidence that the stricter representation leaves the result essentially untouched. Event-level accounting shows otherwise.
f77r falls by approximately -1.199. f105r rises by approximately +1.240. f34r is unchanged. The large opposite movements nearly cancel.
The aggregate is stable because different evidence changes in opposite directions—not because the same evidence survives unchanged.
1. The page-level decomposition
| Page | Split score | Strict score | Change |
|---|---|---|---|
| f34r | +0.500 | +0.500 | +0.000 |
| f77r | +2.545 | +1.345 | -1.199 |
| f105r | +2.649 | +3.889 | +1.240 |
| Batch | +5.693 | +5.734 | +0.041 |
The page changes are not rounding noise. Their absolute magnitudes are roughly thirty times the net change of the batch.
If the only reported number were the batch total, that redistribution would be invisible.
2. f77r loses supportive evidence
The largest movement is inside f77r’s dominant ol/qol family. Under the split representation, that family contributes approximately +2.472. Under the strict filter, it contributes approximately +0.992, a loss of about 1.480.
A second f77r family, okaiin/qokaiin, moves in the other direction by about +0.281. The net page change is therefore approximately −1.199.
The strict filter removes substantial favourable information from f77r.
3. f105r loses adverse evidence
Volume 20 traced f105r’s movement exactly. Its okaiin/qokaiin family contributes approximately −1.240 under the split representation.
The strict filter excludes two adverse q-prefixed observations because their predecessor spans touch doubtful seams. The remaining observation cannot supply current-outcome variation, so the stratum becomes unscored.
The f105r page consequently rises by approximately +1.240. No positive stratum gets stronger.
The same policy is therefore doing two different things on two pages: removing favourable information on f77r and removing unfavourable information on f105r.
4. The cancellation is numerical, not evidential
f77r change ≈ -1.199460
f105r change ≈ +1.240011
f34r change = +0.000000
batch change ≈ +0.040551The final line is the algebraic sum of the first three. It is not evidence that the underlying observations are nearly identical between policies.
The two page changes arise from different current-form families, different loci and opposite evidential directions. Their cancellation is contingent on their magnitudes.
5. Positive and negative mass both shrink
Across the three split-policy pages, positive stratum contributions sum to approximately +8.779, while negative contributions sum to approximately -3.085.
Under the strict filter, positive contributions fall to approximately +7.579, while negative contributions shrink in magnitude to approximately -1.845.
The strict total is slightly higher because the removed negative magnitude exceeds the removed positive magnitude by only about 0.041.
The filter discards meaningful evidence on both sides.
6. What “robust total” can hide
A result can be robust in sign while unstable in composition.
Whenever a sensitivity analysis changes the admitted dataset, the audit should report the aggregate, page or group contributions, positive and negative contribution totals, which strata changed or disappeared, and whether the same observations remain evaluable.
Without that decomposition, a stable headline can conceal a changed evidential basis.
7. This does not decide which representation is correct
The present result does not tell us whether the split or strict policy better reflects the physical manuscript. That requires source evidence about the doubtful boundaries themselves.
It tells us what happens downstream when the stricter rule is chosen: support and counterevidence are both removed, and their effects nearly cancel at batch level.
A preprocessing policy should therefore be justified independently of the final score.
8. Near-equal totals are not independent replication
The split and strict results come from the same manuscript pages and overlapping source observations. They are not two independent tests that happen to agree.
Their similarity can show that one headline sign survives a specified perturbation. It cannot multiply confidence as though a second independent dataset reproduced the effect.
9. The next source audit should preserve this decomposition
Volume 20 fixed a balanced list of f105r families for source-lineage checking: the two strongest positive families and the two strongest negative families.
Volume 21 adds a second requirement. When source-lineage alternatives change event eligibility, the audit should report which side of the evidence balance is being removed or altered.
The next lineage analysis should therefore retain separate positive and negative contribution ledgers rather than reporting only a page-level net.
10. The stable headline is a cancellation result
The fixed common model remains positive under both representations. That is a genuine bounded observation.
But the path from +5.693 to +5.734 is not “almost nothing changed.” A strong positive component on f77r is reduced, a strong negative component on f105r disappears, and the two changes nearly balance.
Volume 21 therefore replaces a vague robustness claim with a more informative one: the aggregate sign survives, while the evidential composition changes materially and asymmetrically.
Sources and reproducibility record
[1] Immediate predecessor. Vol. 20 — The Counterevidence-Retention Check, EDKSG-VOY-V020.
[2] Parent fixed-model experiment. Vol. 13 — The Frozen-Coefficient Check, EDKSG-VOY-V013. Its split and exclude-touching stratum/event indices and frozen coefficient are used without refitting.
[3] Executed local record. EDKSG-VOY-V021-AGGREGATE-CANCELLATION-01. The package contains exact parent indices, a standard-library reproduction script, page-level accounting, and the changed-stratum ledger.
Collection: Voynich Research Library · Editorial framework: Wintour House.
EDKSG-VOY-V021 · Edition 1.0.0 · 5 September 2026 · Baseline EDKSG-VOY-B000. Aggregate-cancellation sensitivity audit. No model refitting, source correction, independent replication or semantic decipherment is claimed.