LOVE optimizer solution quality on shared objectives

The three declared objectives are direct numerical calls to the DISQOVER 0.9.14 tf LOVE objective. The R solver uses the actual source-installed RcppDE::DEoptim; the Python replay uses the engine's bounded optimize_box. Both receive identical deposition, regional background, pollen, PPE, box bounds and iteration cap. Each side also evaluates the other side's selected candidate. The retained summary, R fits, Python fits, R session and source identity report preserve the results and runtime.

All three objective values are within the declared 1e-4 gap of independently derived optima, and every cross-engine evaluation differs from the producing engine's objective by less than the declared 1e-10 absolute tolerance. The symmetric, constrained-boundary and composition-ambiguous examples probe different optimizer geometry. The case arrays, seed and iteration cap are authored study choices retained as configuration. Their custody does not reproduce an optimizer run or establish independent scientific approval. The ambiguous objective has multiple equally good cover vectors, so vector equality would be a false solution-quality criterion. Python reported success after 29–40 iterations and 600–1,230 evaluations. R used the complete 400-iteration cap (8,020 or 12,030 evaluations) in each case; its attained objective is accurate on these examples, while the cap alone does not establish an R convergence event.

The same integer seed does not pair R and Python random streams. This study compares achieved objective quality and common-candidate function values, not identical search trajectories or stochastic full-wrapper draws. The original R oracle separately pairs the LOVE objective and wrapper on supplied candidates, and pairs regional REVEALS count/PPE draws. The separate local stochastic LOVE comparison pairs explicit local count/PPE draws. None of these results establishes ecological prediction skill or a global guarantee for arbitrary LOVE landscapes.

First build the source-locked R library with the current R reference producer. Replay with that library and Rscript, using a new absent output directory under artifacts/:

PYTHONDONTWRITEBYTECODE=1 OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 MKL_NUM_THREADS=1 \
  NUMEXPR_NUM_THREADS=1 VECLIB_MAXIMUM_THREADS=1 \
  .venv/bin/python docs/report/method-certification/optimizer-solution-quality/replay.py \
  --repository . \
  --library artifacts/jobs/native-r-reproduction/library \
  --rscript /usr/local/bin/Rscript \
  --output artifacts/jobs/optimizer-solution-quality/reproduction

The generated library path is a local run product, not a portable dependency. A different machine must provision the locked source archive and dependencies before replay. The producer refuses an archive/installed-source marker mismatch and checks numerical gaps before writing its passed report.