Trace hidden state and restored-result disagreement

Understand the likely causes

• Possible causes include free names in an enclosing environment, formula lookup outside data, incomplete preprocessing metadata, changed factor levels, mutable shared environments, locale or timezone changes, random-state differences, or persistent options and graphics settings. A successful load does not validate these interpretation contracts.

1. Freeze a clean input-to-output reference

Create a small self-contained synthetic script with explicit inputs, units, feature order, category levels, preprocessing, and intended outputs. Record the actual R environment when execution is available, including relevant package and random-generator configuration. Compare the same known input in the old session and a fresh session without relying on an automatically restored workspace. Preserve warnings and names of any externally resolved objects. If no R runtime is available, keep this as a documented investigation plan rather than claiming that a fresh-session comparison was performed.

2. Inspect closures and formula environments

Find free names in reusable functions and terms missing from the supplied model data. A formula can resolve an absent name in its associated environment, while a closure can deliberately retain configuration or accidentally depend on mutable state. The fixture forces a scale of two into a closure and checks an in-memory self-created round trip. The later global scale of 100 is a separate binding. This does not prove that every serialized function is independent: references to packages, environments, external resources, or other helpers need their own explicit contract.

make_multiplier <- function(scale) {
  force(scale)
  function(x) x*scale
}
f <- make_multiplier(2)
scale <- 100
reference <- f(c(1, 3))
restored <- unserialize(serialize(list(fn = f, features = "synthetic value"), NULL, version = 3))
cat("reference=", paste(reference, collapse = ","), "\n", sep = "")
cat("same restored=", identical(reference, restored$fn(c(1, 3))), "\n", sep = "")

reference=2,6
same restored=TRUE

3. Compare representations before final numbers

Check parsed tokens, classes, units, missingness, row identities, factor levels, contrast matrices, and transformed features before comparing final predictions. If those agree, compare model state, relevant options, random call order, and numerical settings. An identical number in one ordinary case can conceal a feature-order mismatch that appears on another input. Keep identity checks exact and numeric comparisons within a justified tolerance. For a decision threshold, test nearby inputs because a tiny numeric difference can change the action even when an average error is small.

4. Repair ownership and verify the full behavior

Pass ordinary data explicitly, capture deliberate configuration at a defined time, and isolate mutable caches by ownership. Preserve preprocessing and category metadata with artifacts. Functions that change options or devices should restore the state they own through a documented cleanup path. After the earliest mismatch is repaired, rerun a known valid case, missing or unsupported input, repeated calls, and the restored behavior. Distinguish a same-environment round trip from portability across versions or platforms. Record any remaining unsupported dependency rather than promising that the object file alone is a complete analysis.

Keep it from recurring

• Keep self-contained scripts and environment records with results, use explicit data and schema checks, document captured state, and separate structural identity from numeric tolerance. Test restoration and cleanup in the environments actually claimed to be supported.