Frozen build, harness and locked evaluation results for the TAALAF PID trend digitization validation study

Companion deposit to a controlled validation study of a browser-based DCS trend digitization pipeline for separator-level PI control loops. It contains the evaluated software build, the execution harness, the locked evaluation results, the source inputs, and every analysis output. This deposit follows the data and makes no timing claim. It must not be confused with the pre-registered analysis-code deposit at 10.5281/zenodo.21784512, which was published on 4 August 2026, before the evaluation set was generated, and whose entire value is that timing. The registered protocol and development evidence are at OSF 10.17605/OSF.IO/KG8JE, time-stamped 2026-08-02T09:46:07.832845Z. The evaluated build is published as loose files so it can be inspected without unpacking: app.js (SHA-256 15dd0026c06aa836e5c465fc261081ce1112d1788f840d178c9870dc6b26149f), app_v43_patch_v1.6.js (9759ae22df27ee4b4867bcddb6cb5b930e06da6d8aa6b254bce12d32b6380e25) and app_v45_decision_patch_v1.8.js (74e45669d1a7bc243b46352c36d27489beddf22ef4cc44d498acbccc07bbe073). These digests were verified by the runner at execution time and recorded in the run provenance. The build is an archived research artefact fixed at August 2026, corresponding to the evaluated version. It is not a release of the product line, and later versions are outside the scope of the study and of this deposit. The locked results archive has SHA-256 4f1a126de54a181d4bbd4f358ad8e768dadcb915d295086a26ddeb8349900cdb and was hashed before any output was opened; the results.json inside it has digest d0844e9729c1e1bce1931ff699a90e62d20979c109f799601bd20dec83bf88b7. The 800 evaluation images are not deposited because they are exactly reproducible from the registered evaluation master seed 335287836, itself derived deterministically from the SHA-256 of the registered protocol, and can be checked against the 900 input checksums included in the dataset manifests. An execution log is included that records what happened in order, including one run executed against the wrong dataset, detected by reading the run provenance before any metric was opened, and voided. Headline results: complete extraction 760/800 = 95.0%, exactly at the prespecified threshold; mean NMAE upper bounds 0.40, 0.92 and 0.62 percent of full scale for PV, CV and SP, inside the 2.0 percent margin; legacy overall exact agreement 430/800 = 53.8 percent (95% CI 46.0 to 61.4) and unweighted kappa 0.361, both co-primary decision criteria failed; action coverage 66.4 percent, selective agreement 81.0 percent and selective weighted kappa 0.691, all three registered selective thresholds missed; zero unsafe actions across 128 reference-Hold images in 16 scenarios, exact two-sided 95% upper limit 20.6 percent at the scenario level. The evaluated software is developed and owned by TAALAF Engineering Solutions, the study was funded by that company, and the author holds a commercial interest in it. The study's headline conclusion is unfavourable to the evaluated product.