Skip to content

Analytic gradient diagnostic for constrained LOVE

The governed English constrained LOVE study refused the BECKHAM held-out fold because its CUTHILL2, HANGINGS and LONGBREACH training LOVE fits did not converge. This source-bound development diagnostic asks whether an exact derivative of the same log-Euclidean LOVE objective resolves those three SLSQP line-search failures. The 30-taxon closed subcomposition, regional handoff, uniform start, nonnegative simplex, two local rings, one repeat, 300-iteration limit, 16,000-objective-evaluation limit and ftol=1e-8 remain fixed. No held-out target pollen is opened or scored.

baseline.json freezes the original request SHA-256 and both ring diagnostics for each training item, plus the governed source experiment and private-source manifest identities. replay.py checks those identities against fresh prepared inputs, rebuilds each training-only request, and requires the original finite-difference SLSQP to reproduce its stop message, iteration count, evaluation count and objective within 1e-10. It checks the analytic derivative against tangent and off-simplex finite differences at positive cover points before running the six bounded fits. The gradient differentiates LOVE's internal cover / cover.sum() normalization, which matters for the off-simplex values SLSQP may evaluate.

BECKHAM-fold training site Original SLSQP, each ring Analytic gradient SLSQP, each ring
CUTHILL2 refused; 147 iterations, 4,898 evaluations, objective 0.823088680 refused; 120 iterations, 402 evaluations, objective 0.823088682
HANGINGS refused; 129 iterations, 4,210 evaluations, objective 1.116182207 converged; 110 iterations, 313 evaluations, objective 1.116182165
LONGBREACH refused; 156 iterations, 5,123 evaluations, objective 0.587411646 refused; 154 iterations, 540 evaluations, objective 0.587410833

All six original fits reproduced their frozen diagnostics exactly. The largest scaled finite-difference discrepancy was 2.79e-8 in tangent directions and 7.42e-9 in ambient coordinates. The full-precision results are in report.json. Exact derivatives reduce objective evaluations and recover HANGINGS, but two training sites still refuse. BECKHAM remains an incomplete fold with no LRA prediction or score. A smaller objective at a refused fit does not waive convergence. This prior-exposed English development result supplies neither measured vegetation accuracy nor Nordic LRA admission.

Recreate the prepared inputs from the private author source capsule and write a new report under artifacts/:

job=artifacts/jobs/author-love-gradient-replay
PYTHONDONTWRITEBYTECODE=1 .venv/bin/python -m \
  bijux_pollenomics_dev.essd_reproduction.author_lra_development \
  --author-source-root /path/to/private/landclim/author-reference \
  --output "$job/source-inputs"
PYTHONDONTWRITEBYTECODE=1 .venv/bin/python \
  docs/report/pollen-reconstruction/author-lra-constrained/prepare.py \
  --source "$job/source-inputs" --destination "$job/prepared"
OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 MKL_NUM_THREADS=1 \
NUMEXPR_NUM_THREADS=1 VECLIB_MAXIMUM_THREADS=1 PYTHONDONTWRITEBYTECODE=1 \
  .venv/bin/python \
  docs/report/pollen-reconstruction/author-lra-constrained/optimizer-gradient/replay.py \
  --prepared "$job/prepared" --output "$job/report.json"
cmp "$job/report.json" \
  docs/report/pollen-reconstruction/author-lra-constrained/optimizer-gradient/report.json

Use an absent job directory. The replay deliberately publishes no LRA product and does not run the complete 15-fold consumer.