Investigate wrong summaries after coercion or recycling

Understand the likely causes

• Possible causes include factor codes used as measurements, character coercion, silent exact recycling, a changed missingness policy, or averaging group means without their counts. Finite output and absence of a warning do not establish that observations, units, or denominators are correct.

1. Freeze tokens, types, and observation identities

Capture the input token, parsed value, type, class, length, names, and relevant attributes at each transformation boundary. Keep a stable id so a reordered vector cannot masquerade as a type-conversion defect. Confirm the measurement unit and whether zero has a genuine meaning. If the failure appears only after import, construct a tiny in-memory CSV with the same token pattern. Avoid copying private rows into a diagnostic artifact. The goal is to locate the first change in meaning, not merely the first place where the final number differs.

2. Distinguish category positions from parsed labels

The fixture intentionally converts a factor in two ways. Direct numeric conversion produces level codes one, two, one, while converting these known numeric labels through character produces ten, twenty, ten. Both totals are finite, so a generic finite-value assertion cannot detect the mistake. This parsing path is justified only because the fixture declares the labels to represent numeric measurements. For arbitrary categories such as low and high, inventing a number is not a repair. Inspect original labels and define a supported numeric format or explicit category mapping before changing conversion code.

labels <- factor(c("10", "20", "10"), levels = c("10", "20"))
codes <- as.numeric(labels)
measurements <- as.numeric(as.character(labels))
cat("codes=", paste(codes, collapse = ","), "\n", sep = "")
cat("measurements=", paste(measurements, collapse = ","), "\n", sep = "")
cat("code total=", sum(codes), "; measurement total=", sum(measurements), "\n", sep = "")

codes=1,2,1
measurements=10,20,10
code total=4; measurement total=40

3. Audit every length and denominator

Print or assert the lengths of both operands before elementwise arithmetic. A shorter vector can recycle silently when its length divides the longer vector, so test a four-versus-two case even if no warning appears. State whether a scalar, paired vector, or repeating pattern is permitted. Then reconstruct the summary numerator and denominator from the retained observations. Count missing and excluded values, and compare group counts if group means were combined. A change from three rows to two observed measurements is an analytical policy that must be visible in the reported result.

4. Verify meaning as well as the resulting number

Choose a repair that predicts a specific corrected representation, length, pairing, and summary. Preserve a failing factor, unequal-length operands, missingness, an all-missing case, and an empty vector as focused fixtures. Compare known id-value pairs after every row-changing step. A corrected total alone can conceal another mistake that happens to cancel numerically. Re-run the calculation in a fresh supported R session when available and inspect warnings rather than suppressing them. Keep the original derived expectation and actual execution evidence in separate records.

Keep it from recurring

• Use typed import schemas, exact factor policies, length assertions, and counts beside missing-value summaries. Test domain meaning with small known records, and review changes to eligible rows or category vocabularies as changes to the analysis contract.