Computational analysis can create an impression of certainty very quickly. A score appears, a category is assigned, and the result can look more authoritative than it really is. VerseVAD was built to resist that tendency.
The goal is not simply to produce numbers. The goal is to make those numbers understandable, inspectable, and reproducible. When VerseVAD reports a result, it also tries to show where that result came from, what evidence supported it, what was missing, and which methodological choices shaped the outcome.
A Score Is Only the Beginning
A poem-level average can be useful, but it is never the whole story. A valence score, for example, may look surprising until we examine the words contributing to it. A readability score may appear high or low until we look at word frequency, age of acquisition, line length, or syllable complexity. A meter candidate may seem convincing until we inspect pronunciation coverage and the lines that could not be resolved.
For that reason, VerseVAD reports more than final values. It includes denominators, match rates, coverage figures, weighting choices, contributor words, configuration details, and methodological cautions. Where possible, it also preserves line-level, stanza-level, token-level, and type-level evidence in companion data files. That information is not decorative. It is part of the result.
Missing Evidence Stays Missing
One of the project’s core principles is that missing evidence should not quietly become neutral evidence.
If a word is not found in a lexicon, VerseVAD does not automatically assign it the midpoint. If a pronunciation cannot be resolved, the software does not invent one simply to complete the line. If a formula requires evidence that is not available, that limitation should remain visible.
This matters especially in poetry, where proper names, archaic words, unusual spellings, invented language, dialect, and poetic elision are common. A missing word may be central to the poem. Treating it as emotionally neutral, perfectly ordinary, or automatically resolved would create a cleaner-looking result at the cost of honesty.
Coverage figures help readers understand how much of the poem actually contributed to a particular analysis. They do not tell us whether an interpretation is correct, but they do tell us how much evidence the calculation had to work with.
The Method Should Be Reproducible
VerseVAD is designed so that an analysis can be revisited later rather than disappearing into a single screen.
Reports record the text being analyzed, the analysis configuration, resource names, result identifiers, denominators, and relevant methodological notes. The exported CSV files preserve the underlying tabular evidence, including summaries, token audits, structural results, matched terms, and coverage details. This makes it possible to inspect how a result was produced and to compare analyses conducted under the same settings.
Reproducibility also means being clear about which version of a method was used. A formula can change. A lexicon can be updated. A scoring profile can be revised. VerseVAD therefore treats configurations and scoring profiles as versioned analytical choices rather than timeless truths.
That may feel overly careful for casual exploration, but it becomes essential when results are used in scholarship, teaching, corpus comparison, or publication.
Formulas Should Be Explainable
VerseVAD does not rely on secret scoring rules. When the software introduces a composite measure, the components and weights should be visible.
The VerseVAD Poetry Reading Ease score, for example, reports its four components, their weights, their raw values, their coverage, and the anchors used to translate them into ease scores. The score combines frequency, age of acquisition, line accessibility, and word complexity, but it also states clearly that it measures surface-level linguistic accessibility rather than thematic or interpretive depth.
The same principle applies elsewhere. A cumulative lexical load is identified as length-sensitive. A type-weighted result is distinguished from a token-weighted one. A nearest meter candidate is not presented as definitive scansion. A PoetryID label is described as a lexical-affective neighborhood, not a declaration of the poem’s emotion, genre, or quality.
A method does not become weaker when its limits are named. It becomes more useful.
Transparency Makes Revision Possible
VerseVAD is still developing, and that development includes changing our minds. Sometimes a result confirms an expectation. Sometimes it exposes a bug. Sometimes two methods disagree. Sometimes a new comparison shows that a distinction we expected to find is simply not present in the data. Those moments are not failures to hide. They are part of building a better analytical tool.
Transparent formulas and auditable outputs make that process possible. When a score behaves strangely, we can trace the calculation, inspect the contributing words, compare alternate settings, and decide whether the result reflects the poem, the resource, the method, or an implementation problem. Continual improvement depends on being able to see what the software is doing.
A Tool for Better Questions
VerseVAD is not meant to produce unquestionable answers about poetry. It is meant to support closer attention. Showing the work allows readers to disagree intelligently. They can question a lexical match, compare token and type weighting, examine missing words, inspect a line-level result, or decide that a particular measure is not useful for the poem in front of them.
That is a healthier relationship with computational analysis than simply accepting or rejecting a score. VerseVAD’s commitment to transparency is ultimately a commitment to the reader. The software should not ask for trust merely because it produces numbers. It should provide enough evidence for users to understand the result, test it, challenge it, reproduce it, and decide what it contributes to the reading.
The number is not the argument. The evidence is.