Skip to content

Swedish observed-pollen chronology block

This development comparison asks whether fixed pollen predictors can predict a Swedish target age bin using source pollen from exactly two 500 BP bins older. For every target, the predictor excludes source bins at the target age and the immediately younger and older ages. It uses the already retained four-country training folds, which exclude the target physical site and group. The preparation declares this rule and binds the sealed outcome identity before scoring. The prediction freeze contains all predictions and refusals before the target counts are opened.

The 31-taxon response is the retained closed observed-pollen subcomposition. Each source composition comes from one positive-count bin at the declared older age. Training mean, nearest site and inverse-distance-squared use fixed methods and site-center distances. Every one of the 54 target site bins had at least two eligible training groups: 21 at Abborrtjärnen, 20 at Bussjösjön and 13 at Bökesjön. There were 162 frozen method predictions, 54 scored bins and no refusals. The coverage, scores, paired scores and machine report retain those denominators and each source bin.

The paired comparator is the previously retained same-bin, all-site predictor from the support sensitivity study. It uses the identical target bins and observed target counts. Equal-site MAE first averages bins within each of the three sites; sample-weighted MAE is a separate descriptive view. The scientific interpretation reports both the effect and its limits.

This is a nominal source-bin separation on previously exposed data. It does not establish nonoverlap of true age distributions, propagate age-model uncertainty, provide independent confirmation, reconstruct LRA vegetation or measure vegetation accuracy.

Reproduce

From the repository root, run the three stages into absent artifacts/ paths:

PYTHONDONTWRITEBYTECODE=1 .venv/bin/python -m \
  bijux_pollenomics_dev.source_completion.swedish_chronology_block prepare \
  --repository . --output artifacts/jobs/swedish-chronology-block/replay/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 -m \
  bijux_pollenomics_dev.source_completion.swedish_chronology_block predict \
  --repository . \
  --prepared artifacts/jobs/swedish-chronology-block/replay/prepared \
  --output artifacts/jobs/swedish-chronology-block/replay/frozen
PYTHONDONTWRITEBYTECODE=1 .venv/bin/python -m \
  bijux_pollenomics_dev.source_completion.swedish_chronology_block score \
  --repository . \
  --prepared artifacts/jobs/swedish-chronology-block/replay/prepared \
  --frozen artifacts/jobs/swedish-chronology-block/replay/frozen \
  --freeze-sha256 8592b580600da80a76903c9fa68435619bc53039d4f958e7f5b826a361360c49 \
  --output artifacts/jobs/swedish-chronology-block/replay/scored

The score stage authenticates the frozen predictions before reading target outcomes. Compare the replayed files byte for byte with the retained products.