Abborrtjärnen chronology sensitivity for direct pollen prediction¶

We replayed the three-site Swedish direct-pollen development study with a conservative age rule for Abborrtjärnen (ABBO). The age evidence is the published LegacyAge 1.0 S5 depth join to the same 107 Neotoma samples. A sample enters a 500-calibrated-year BP bin only when its entire reported S5 95% interval lies inside the bin. A sample whose interval merely overlaps a bin remains unresolved; we never assign it by its mean or duplicate its counts across bins. Bussjösjön and Bökesjön retain their native Neotoma point chronologies because no equivalent interval source was admitted for them.

The original study has 13 shared bins. ABBO has eight interval-contained samples in four of those bins: one in 1500–2000 BP, three in 2000–2500 BP, three in 2500–3000 BP, and one in 4500–5000 BP. The eight samples contribute 3,543 NISP in the same closed 16-taxon subcomposition. The sample ledger accounts for all 107 ABBO samples: eight contained, 37 overlapping a selected bin without full containment, and 62 outside the four selected bins. The prepared ABBO profile contains the counts that train or score the interval sensitivity.

We held out each entire physical site in turn, preserving the four fixed predictors and their zero handling. The target pollen profile was not opened by its prediction process; the ABBO, Bussjösjön and Bökesjön prediction files were saved and hashed before the separate scoring process opened target counts. All four methods predicted and scored the same 12 site-bin targets. The comparison below uses exactly those four bins for both the original point-age result and this interval-contained result. Values are site-macro scores: four bins average within a site, then three sites have equal weight.

Fixed method	Original MAE	Interval MAE	Original Hellinger	Interval Hellinger
Training mean	0.060594	0.060258	0.427503	0.426070
Nearest site	0.056852	0.056477	0.425366	0.424255
Inverse distance squared	0.060058	0.059762	0.439302	0.438298
CLR RBF ridge	0.059051	0.058751	0.431725	0.430511

The small paired differences are a development sensitivity on a very sparse and selected subset. They do not establish a better chronology or model. The reported S5 intervals are marginal age summaries, not a joint posterior; full containment in a 95% interval is not a guarantee of true bin membership. Selection of the four bins followed inspection of ABBO source support. The original site outcomes were already exposed, and the three lakes are geographically unbalanced. The two southern chronologies remain point ages, so this is not a common uncertainty treatment across sites. No basin geometry, PPE, dispersal profile, vegetation truth, or LRA reconstruction enters this study. The score rows, per-taxon scored predictions, summary, same-support comparison, config, and input receipt retain the exact denominators and source identities.

Reproduce¶

From the repository root, use an absent output directory under artifacts/. The producer checks the byte identities of the Neotoma input, original profiles and LegacyAge sample join. Run prepare, three site-specific predict processes, then score after all prediction files exist:

PYTHONDONTWRITEBYTECODE=1 .venv/bin/python \
  docs/report/pollen-reconstruction/age-interval-sensitivity/replay.py \
  prepare --repository . --output artifacts/abbo-age-interval-sensitivity
PYTHONDONTWRITEBYTECODE=1 .venv/bin/python \
  docs/report/pollen-reconstruction/age-interval-sensitivity/replay.py \
  predict --repository . --output artifacts/abbo-age-interval-sensitivity \
  --target-dataset-id 19913
PYTHONDONTWRITEBYTECODE=1 .venv/bin/python \
  docs/report/pollen-reconstruction/age-interval-sensitivity/replay.py \
  predict --repository . --output artifacts/abbo-age-interval-sensitivity \
  --target-dataset-id 53995
PYTHONDONTWRITEBYTECODE=1 .venv/bin/python \
  docs/report/pollen-reconstruction/age-interval-sensitivity/replay.py \
  predict --repository . --output artifacts/abbo-age-interval-sensitivity \
  --target-dataset-id 54001
PYTHONDONTWRITEBYTECODE=1 .venv/bin/python \
  docs/report/pollen-reconstruction/age-interval-sensitivity/replay.py \
  score --repository . --output artifacts/abbo-age-interval-sensitivity

Compare the generated age-bin-support.csv, ABBO prepared profile, three prediction JSON files, predictions-scored.csv, scores.csv, summary.json, comparison.json and input-manifest.json with the governed files here. Source credit and original investigator attribution are retained in the parent cohort; the LegacyAge S5 source is Li et al. (2021), PANGAEA 933132, CC BY 4.0.