Voynich | edkSG Research Volumes · EDKSG-VOY-V020 · Edition 1.0.0 · 5 September 2026, Singapore
Research status: executed fixed-model counterevidence-retention audit of f105r using the three representation policies already declared in Volume 13. No model refitting, new transcription, image verdict, population inference or decipherment is claimed.
A stricter filter makes f105r look more positive.
But it does not do so by increasing any positive stratum.
Under the split policy, the frozen common model’s positive strata contribute +4.933, while its negative strata contribute -2.284, for a net page score of +2.649.
Under the stricter policy that excludes pairs touching doubtful seams, the positive contribution remains exactly +4.933. The negative contribution shrinks to -1.044. The page score consequently rises to +3.889.
The entire increase—approximately +1.240—comes from one negative okaiin / qokaiin family ceasing to provide a scoreable comparison.
The stricter view is not simply “cleaner support”. In this page it selectively removes counterevidence. That removal must remain visible whenever the stricter total is reported.
1. The page result is a balance, not a uniform effect
Volume 13 reported f105r as a positive fixed-model transfer: approximately +2.649 on the primary split representation. That page-level number is correct under its declared input.
The page is not composed entirely of positive same-form families. Twelve strata have both current outcomes under the split policy. Two contribute exactly zero to the model-versus-zero difference, eight are positive, and two are strongly negative.
The positive strata sum to approximately +4.933. The negative strata sum to approximately −2.284. The published page score is their net.
This distinction matters because a later representation policy can change which strata remain scoreable. If negative evidence disappears more readily than positive evidence, the net can improve without the supportive observations becoming any stronger.
2. The largest positive and negative families
| Current-form family | n | Predecessor-y | q-prefixed | Overlap | Conditional contribution |
|---|---|---|---|---|---|
| okeedy / qokeedy | 5 | 3 | 3 | 3 | +1.420 |
| okaiin / qokaiin | 3 | 1 | 2 | 0 | −1.240 |
| okal / qokal | 2 | 1 | 1 | 0 | −1.044 |
| okedy / qokedy | 5 | 4 | 4 | 4 | +0.992 |
| four two-case families | 2 each | 1 | 1 | 1 | +0.500 each |
The two largest contributions point in opposite directions. The strongest positive family aligns all three predecessor-y cases with q-prefixed outcomes and both non-y cases with unprefixed outcomes. The strongest negative family does the reverse: its two q-prefixed outcomes follow non-y predecessors, while its single y-ending predecessor is followed by an unprefixed form.
A page-level positive total therefore hides meaningful internal conflict. That conflict is part of the evidence, not noise to be removed after the result is known.
3. The adverse okaiin / qokaiin family
The three split-policy observations are:
| Source locus | Predecessor | Current form | Feature tuple | Strict policy |
|---|---|---|---|---|
| f105r.8 | air | qokaiin | non-y → qo | Excluded: predecessor span begins after doubtful comma |
| f105r.21 | chedy | okaiin | y → o | Retained |
| f105r.33 | aiiin | qokaiin | non-y → qo | Excluded: predecessor span begins after doubtful comma |
Notice what the stricter rule is doing. The separator between each predecessor and current form is a confident dot. The reason the first and third pairs are removed is that the predecessor span itself touches a doubtful seam on its other side.
After those two q-prefixed observations are excluded, only chedy → okaiin remains. A one-outcome remainder cannot supply within-family current-outcome variation. The entire −1.240 stratum therefore vanishes from the conditional score.
The strict policy has not converted a negative family into a positive one. It has converted an adverse comparison into NO EVALUATION INFORMATION.
4. The page-score increase is exactly the lost counterevidence
| Policy | Positive contribution | Negative contribution | Net f105r score |
|---|---|---|---|
| Split doubtful seams | +4.933 | -2.284 | +2.649 |
| Join doubtful seams | +4.933 | -2.284 | +2.649 |
| Exclude pairs touching doubtful seams | +4.933 | -1.044 | +3.889 |
The split and joined page scores are identical here. The strict score is higher by +1.240011.
The removed negative family’s split contribution is -1.240011. The magnitudes are equal to floating-point precision.
In other words:
strict net − split net
= 3.8887976035 − 2.6487863461
= +1.2400112574
removed adverse stratum
= −1.2400112574No positive stratum gained support. No coefficient changed. The apparent improvement is entirely an accounting consequence of removing that counterexample family from the scoreable population.
5. A conservative filter can still be favourable
“Exclude uncertain material” sounds automatically conservative. It is conservative about admitting observations, but not necessarily conservative about the direction of a downstream result.
If uncertainty is distributed asymmetrically across supportive and adverse observations, exclusion can move an estimate in either direction. Here it moves the frozen score upward because the doubtful-seam contact is concentrated in two adverse q-prefixed observations.
This does not mean the strict rule is wrong. A boundary policy should be justified by representation principles, not chosen according to whether it raises or lowers the model score. The result means that its directional consequence must be disclosed.
A useful research receipt therefore records not only how many events were excluded, but which previously informative claims they supported or contradicted.
6. Counterevidence does not disappear completely
The second strong negative family, okal / qokal, survives the strict policy unchanged. It contains one y-ending predecessor followed by unprefixed okal and one non-y predecessor followed by qokal. At the frozen positive coefficient it contributes approximately −1.044.
That surviving adverse family prevents the strict page from becoming a picture of uniform alignment. Even the higher +3.889 total remains a net over contradictory families.
The positive okeedy / qokeedy and okedy / qokedy families also survive unchanged under the strict view. The stricter result is therefore not fabricated from nothing; it retains genuine positive and negative structure. Its increase, however, is attributable specifically to lost adverse information.
7. What this changes in the interpretation of f105r
The strongest defensible statement is now more precise than “f105r is robust across gap policies”.
We can say that the page remains positive under the three declared policies, while also saying that the strictest policy increases the net score by eliminating one strongly adverse family’s outcome variation.
Those statements are compatible. Robustness of the sign does not imply neutrality of the policy’s effect on the evidence balance.
The f105r page therefore becomes a useful warning against reading sensitivity tables only by their final totals. The route from one total to another matters.
8. Preserve route failures and adverse evidence as first-class records
A research system should distinguish three states: an observation that supports the current model, an observation that contradicts it, and an observation that can no longer answer the question under a stricter representation.
The third state must not be silently converted into the first. Here the strict policy does not make the two excluded qokaiin observations supportive. It makes them unavailable to this particular comparison.
This distinction generalises beyond Voynich. A narrower evidence route can improve reliability while also reducing the number of opportunities for a model to be contradicted. Both consequences should be recorded.
9. The source-lineage audit now has a balanced target list
Volume 19 proposed a source-lineage audit of f105r. The present influence map improves that plan by fixing the high-value families before inspecting historical alternatives.
The next lineage audit should include both sides:
- okeedy / qokeedy — largest positive family, +1.420;
- okaiin / qokaiin — largest negative family, −1.240 and boundary-sensitive;
- okal / qokal — second negative family, −1.044 and retained by the strict policy;
- okedy / qokedy — second major positive family, +0.992.
Selecting only the positive families would turn a source audit into confirmation seeking. Selecting only the negative families would create the opposite bias.
The source-history comparison should ask the same claim-scoped questions used in Volumes 18 and 19: does the current-form family exist under the other row, is the relevant word boundary represented, and does the predecessor retain the y/non-y classification required by the model?
Volume 20 does not strengthen the model by deleting uncertainty. It strengthens the research record by showing exactly when uncertainty deletes counterevidence.
Sources and reproducibility record
[1] Immediate predecessor. Vol. 19 — The Boundary-Lineage Check, EDKSG-VOY-V019.
[2] Parent experiment. Vol. 13 — The Frozen-Coefficient Check, EDKSG-VOY-V013. The frozen coefficient and split/join/exclude-touching event and stratum indices are used without refitting.
[3] Current source representation. ZL3b-derived f105r diplomatic record retained by the Volume 13 source manifest. The strings quoted here identify only the directly audited event relations and are not interpreted semantically.
[4] Executed local record. EDKSG-VOY-V020-COUNTEREVIDENCE-RETENTION-01. The package binds the exact parent indices, ranks the f105r strata, records the three affected qokaiin-family observations and verifies the accounting identity between the lost negative stratum and the strict-policy score increase.
Collection: Voynich Research Library · Editorial framework: Editorial Publishing House.
EDKSG-VOY-V020 · Edition 1.0.0 · 5 September 2026 · Baseline EDKSG-VOY-B000. Counterevidence-retention audit. No source correction, model refitting, independent review, population p-value or semantic decipherment is claimed.