ADAPT-SELECT: Historical Correction for Multi-Evidence Regression Test Selection

It typically takes about 3–4 minutes in the reference environment, excluding installation. The uncompressed inputs occupy about 385 MB; allow additional space for the virtual environment and fresh generated results.</p><p dir="ltr">Success means exit status zero and <code>results/full/verification.json</code> containing <code>"passed": true</code> and <code>"mode": "full"</code>. <code>results/full/commands.json</code> lists completed stages and durations; the adjacent logs explain any failure.

The runner refuses to overwrite an existing result directory. For another run, choose <code>--output-dir results/full-2</code>.</p><h2 dir="ltr">Faster checks with narrower scope</h2><pre>.\.venv\Scripts\python.exe -I -B reproduce.py --verify-only<br>.\.venv\Scripts\python.exe -I -B reproduce.py --replay-only --output-dir results/replay<br></pre><p dir="ltr"><code>--verify-only</code> checks package integrity. <code>--replay-only</code> recalculates A/B analyses and compares the paper with retained C evidence, but does <b>not</b> execute the application to regenerate observations.

It therefore does not establish fresh execution reproducibility. Full execution is the default. See <code>REPRODUCIBILITY.md</code> for per-study commands, outputs, and failure interpretation.</p><h2 dir="ltr">Studies and expected results</h2><table><tr><th><p dir="ltr">Study</p></th><th><p dir="ltr">Design</p></th><th><p dir="ltr">Principal reproducibility checks</p></th></tr><tr><td><p dir="ltr">A: foundation</p></td><td><p dir="ltr">510 tests; 8 snapshots; 3 repetitions; separate K=15/30/45 histories</p></td><td><p dir="ltr">Each budget matches 4,080 profiles, 3,570 label records and 420 prediction artifacts.

Independent analysis checks native and equal-count results separately.</p></td></tr><tr><td><p dir="ltr">B: exploratory</p></td><td><p dir="ltr">220 tests; 5 snapshots; 2 repetitions; K=15/30/45 replay</p></td><td><p dir="ltr">Fresh primary records match 1,100 profiles, 880 labels, 168 prediction artifacts and 2,200 individual execution traces. At K30, base/guided find 232/236 proxy impacts; the semantic component finds 238.</p></td></tr><tr><td><p dir="ltr">C: controlled</p></td><td><p dir="ltr">108 requirements; 27 crossed evaluation cells per output panel; 2 panels; K=9</p></td><td><p dir="ltr">Per panel, base=110, route+stage=110, route-only=116, development-selected component=117.

Fresh execution matches 540 checked cell files and 6 summaries.</p></td></tr></table><p dir="ltr">The negative C finding is retained: stage conditioning does not improve the base policy in this design. The panels are not pooled as independent changes. These workloads share one local application backend; they are not independent software projects.

B's outcome is a proxy-impact label, while C uses versioned factual requirements. The older differing B archive is historical provenance and is not the paper's primary evidence.</p><h2 dir="ltr">Package map</h2><ul><li><code>study_a/source/</code>, <code>study_b/source/</code>: complete Python application, evidence, selection, scheduling, test generation and analysis implementation used by the studies, with catalogs and corpus inputs.</li><li><code>study_a/scripts/</code>, <code>study_b/scripts/</code>, <code>study_b/validation_scripts/</code>: regeneration, policy/control/sensitivity analysis, recovered-case evidence and validation code. <code>study_b/tests/</code> contains the executable checks invoked by the runner.</li><li><code>study_a/inputs/</code>, <code>study_b/raw/</code>: archived primary scientific records; <code>study_a/expected/</code>, <code>study_b/expected/</code> and <code>study_b/expected_reported/</code> retain expected analytical results.</li><li><code>study_c/run_controlled_experiment.py</code>, <code>study_c/frozen/</code>: controlled-experiment source and frozen catalogs, schedules, fault allocations, selector masks and source hashes. <code>study_c/expected_execution/</code> contains raw cell-level materializations, selection commits, fitted retrieval state, answers and oracle records.</li><li><code>study_b/historical_case_dossiers/</code>, <code>study_b/historical_archive_diagnostic/</code>, <code>regeneration_investigation/</code>: explicitly historical case/context and environmental diagnostics, separated from refreshed primary results.</li><li><code>paper/</code>: current manuscript source, references, compiled PDF and implementation supplement.

Numerical checks consume the source directly; compiling the PDF is not needed to run experiments.</li><li><code>SHA256.json</code>: current per-file SHA-256 manifest. <code>ANONYMIZATION.md</code> explains retained scientific identifiers and the limits of identity removal; <code>RIGHTS.md</code> preserves third-party attribution and licensing status.</li></ul><p dir="ltr">Missing original producing environments and Study A's original per-replicate associations cannot be reconstructed retroactively.

Current primary A/B/C records regenerate in the reference environment; refreshed B and C include complete current oracle/output associations. Preserved historical fingerprints describe the original records, while <code>SHA256.json</code> describes the bytes of this review package.</p><p dir="ltr">Repository history, direct author/workstation metadata and private revision identifiers are excluded. Necessary original benchmark identifiers and corpus text remain unchanged so that exact computation is preserved; this is not a guarantee that scientific content cannot be linked to prior work.

See <code>ANONYMIZATION.md</code> before public deposition. No upload or DOI registration is performed by this package. Manuscript URL placeholders are intentionally unfilled.</p>