Reproducing plots from retained inputs and execution records

The separately callable reproducibility service reruns a revision’s retained R script against its retained selected records and compares the new result with stored identities. Replay does not run automatically during confirmation, human review, or export. It returns evidence for a caller to inspect; it does not promote a revision, replace its accepted artifacts, or approve a plot.

View in Full Screen

The diagram shows the service sequence and optional evidence write. Details of data reconstruction and comparison rules follow below.

What replay reads

new_reproducibility_service() accepts an evidence repository, readable artifact store, runner, and verification service. Its reproduce_revision(revision_id, idempotency_key = NULL) method retrieves the persisted revision and its latest revision_context entry. Missing context stops the operation. With the SQLite repository, entries are ordered by created_at, then evidence_id; “latest” means the last matching row in that order.

.revision_context_details() retains the question, confirmed specification, input identity, selected records, generated script, execution request and result, environment, profile warnings, and checks. Replay reads the script bytes from the artifact store using the revision’s code_hash, rather than recompiling the specification or using the script copy in context JSON. The local store verifies an artifact’s SHA-256 hash before returning its bytes through get().

The retained request supplies the original specification hash, script hash, snapshot identity and hash, and harness version. Replay restores its execution_request class directly. It does not call the request constructor again or re-confirm plotting choices.

Reconstructing selected data

.evidence_selected_data() rebuilds a data frame from input_data$selected_data. It flattens list-valued columns, restores selected_data_columns order when recorded, and otherwise uses the retained column names. Row values retain their stored sequence.

The helper converts every value equal to the literal string "NA" to an R missing value. It then restores recorded numeric and integer columns using as.numeric() and as.integer(). Recorded class names may be normalized by stripping their prefix through the last dot; positional names are used when class names are absent and counts match.

This is limited reconstruction, not a general R-object serializer. Factors, dates, custom classes, and attributes are not explicitly restored. A genuine character value "NA" can become missing. Those differences can make the reconstructed input fail the runner’s snapshot-hash check. Retaining a class label alone does not guarantee that every input type can be recreated.

Running and checking the result

The default replay runner is new_execution_runner(), which executes locally in the current R process. The normal governed execution path uses a clean callr subprocess, described in script generation and execution. A caller can inject a different runner into replay; the default does not recreate an isolated or locked historical environment.

The runner receives the retained request, script, reconstructed analysis_data, and recorded classification. If classification was omitted, replay passes "synthetic". This default does not establish synthetic provenance. The local runner accepts synthetic or de-identified classifications and checks script and data hashes against the immutable request before execution.

Replay invokes the verification service and retains its result. It then calculates a separate comparison set:

Check	Comparison
specification	Retained request specification hash equals the revision specification hash.
generated_r_code	SHA-256 of replayed script text equals the revision code hash.
analytical_result	New analytical hash equals the revision analytical hash.
environment	New environment fingerprint equals the recorded context fingerprint.
image_bytes	New image hash equals the revision image hash, only when fingerprints match.

An environment mismatch makes environment false and sets image_bytes to NA. The overall result still fails because the environment comparison failed. The fingerprint covers recorded R/platform/OS details, versions of the application, ggplot2, and patchwork, graphics device, and image dimensions. It is not a complete environment lock or inventory of every font, library, dependency, or operating-system setting.

The returned revision_reproducibility_result has status failed when any comparison is exactly FALSE; otherwise it has status passed. Its embedded verification$status does not participate in that calculation. Callers therefore need to inspect both comparison and verification results. See execution verification and revision evidence for the separate execution checks. Hash agreement does not independently establish statistical correctness.

Recording and using replay evidence

Supplying an idempotency_key records a reproducibility_rerun entry containing the result and its outcome. Omitting the key returns the result without that evidence write. The key controls repository command deduplication; the service still runs the script before attempting the write on every call.

There is no revision-status prerequisite in this method and no normal UI replay gate. Replay neither changes review status nor records a new governed execution attempt. Its practical prerequisites are available context, accessible script bytes, reconstructable records, and a runnable environment.

Coverage and limits

test-revision-evidence.R checks retained context and a passing replay with recorded evidence. Its injected runner sources the script but writes fixed "png bytes"; that test does not demonstrate real PNG reproducibility or clean-subprocess behavior. The tests were inspected for this document and were not executed. Source inspection also identifies the reconstruction and status-calculation limits above; no browser or regulated validation claim follows from this documentation.