Held-out waveform workload supplement for Evidence-preserving substitution in scientific computation

Reproducibility evidence for the held-out waveform workload in "Evidence-preserving substitution in scientific computation". Scientific purpose. The earlier controlled downstream experiments isolate differences between retained information, executable deduction, complete input delivery and current applicability. This additional experiment tests executable checks for complete candidate-set support through a pre-specified sequence of changing questions and inputs on waveform source groups withheld from development and calibration. Source and protocol. The St Petersburg INCART 12-lead Arrhythmia Database v1.0.0 (Evgeny Yakushenko, 2008; PhysioNet DOI 10.13026/C2V88N) supplies 32 documented Holter source groups. Eight groups were used for development/calibration and 24 were held out. All 1,152 selected test windows were retained. Code, protocol, calibration and numerical argument were locally hash-frozen before test-signal acquisition. The 12 stages include threshold expansion, candidate-catalogue expansion, query replacement, repeats and restoration. This is a predefined execution workload, not external preregistration or independent-team replication. Evidence and results. Four routes (always-native execution, exact-distance caching, dependency-aware safe caching, and evidence-bearing objects) share current consumer requirements and a fixed FIFO acquisition policy, at three native-call budgets. The prespecified half-budget arm contains 5,760 registered analysis requests per route across 24 source groups. Evidence-bearing objects and the same-policy safe cache both complete 5,739 requests, compared with 5,520 for exact caching and 2,880 for always-native execution. The unit of analysis is the source group, not each repeated request. Question and input changes expose distinct unmet requirements; native acquisition restores complete support, and restoring the original conditions permits revalidated reuse without new native calls in the tested restoration stage. A separately implemented full-matrix reference calculation, run after all route arms, found no resolved endpoint or complete candidate-set mismatch. Scope. Equality with the same-policy safe cache is an implementation comparator, not a claim of a new pruning algorithm or superior object representation. The exact cache was faster than the object route in the measured one-shot timing. Results do not establish clinical validity, a general runtime advantage, an optimal adaptive policy, or population-wide guarantees. The entire INCART archive was not held out. Contents and reproduction. The ZIP contains the unchanged V70 scientific evidence directory: source code and reproduction entry point, selected raw waveform windows and provenance receipts, 102 files in the pre-execution freeze, input bindings, complete route ledgers, independent reference outputs, analysis tables, numerical argument, audit and figure sources. Start with dtw_independent_workload_v70/README_V70.txt and run "python3 reproduce_v70.py verify" from that directory for read-only integrity verification. Full upstream .dat caches are omitted; selected raw windows and official checksum receipts are included. File identity: V70_Independent_Workload_Evidence.zip; 20,518,365 bytes; SHA-256 6e35f76d21beebbbc9a1a6e1ff29d20f0804586174e13ca4818afe912ac3c664; MD5 54d5497495ef86906c669abc879d04f2. V70 identifies the frozen scientific experiment; subsequent manuscript revisions do not change these bytes. This standalone supplement complements, and does not replace, the historical V58 cross-task scientific snapshot at DOI 10.5281/zenodo.23852233. Licensing and attribution. The record's Creative Commons Attribution 4.0 license applies to author-created documentation and derived result tables. INCART source data and the included raw selected windows retain the Open Data Commons Attribution License v1.0 and the original PhysioNet attribution requirements; the record license does not relicense third-party materials. Source code is supplied for reproducibility without a separately declared software license in this frozen package. Funding. National Natural Science Foundation of China (62221005 and 62222601); National Key Research and Development Program of China (2025YFF0514700).