Dalmutladdo pollen prediction under strict LegacyAge age intervals¶
We tested the frozen Norway-excluded direct observed-pollen predictions against a conservative age support rule for Neotoma dataset 20034 (DALMUT). This is a prior-exposed development sensitivity, not a new held-out fit, LRA reconstruction, or vegetation validation. The response remains the parent's closed subcomposition of 16 literal pollen labels. Its denominator is the sum of those labels' native NISP counts, not all pollen or total vegetation.
The independent LegacyAge S5 depth join provides 95% age ranges at exact sample depths. We retained a sample only when its entire reported interval fits inside the original half-open 500-cal-BP bin. The 95% endpoints are treated as a range, with no inferred posterior probability or draw. An upper endpoint exactly on a bin edge fails this rule. No age interpolation was made for the 4.5 cm sample.
| Support decision | Samples |
|---|---|
| Parent Dalmutladdo cohort | 72 |
| Exact LegacyAge S5 depth match | 71 |
| Full interval inside original bin | 2 |
| Refused because interval crosses a bin boundary | 69 |
| Refused because no exact S5 depth exists | 1 |
The two retained samples are 175359 in bin 1 (500–1,000 cal BP) and 175366
in bin 3 (1,500–2,000 cal BP). Their LegacyAge ranges are respectively
604–982 and 1,528–1,965 cal BP. Both stay in their source-selected Neotoma
point-age bins. The parent had 22 scored Dalmutladdo bins; 20 have no sample
under this rule. All 70 sample refusals are recorded in
sample-support.csv. All eight reused predictions, four
fixed methods for two bins, remain in predictions.json.
| Fixed method | Strict two-bin MAE | Parent MAE, same bins | Strict Hellinger | Parent Hellinger, same bins |
|---|---|---|---|---|
| Training mean | 0.038759 | 0.038960 | 0.352103 | 0.348078 |
| Nearest site | 0.001746 | 0.001831 | 0.079432 | 0.066567 |
| Inverse distance squared | 0.005701 | 0.005698 | 0.155724 | 0.148674 |
| CLR/RBF ridge | 0.003165 | 0.003102 | 0.103466 | 0.094985 |
These are one-site, two-bin descriptive means, with no independent site uncertainty interval. The strict scores use only the two retained sample count profiles. The paired nominal score on those same samples is identical because neither sample changes bins. The parent columns use its complete original outcome composition for bins 1 and 3 (six and three samples, respectively). Differences between those columns reflect the changed sample support, not an age-driven reassignment or an improved model. The score ledger keeps each method and bin visible.
The protocol pins the parent prediction bytes, source join,
count axis and interval rule. freeze reads the age claims and existing
predictions but no target pollen observations; score subsequently opens the
source observations and checks all 22 reconstructed nominal bin denominators
against the parent scores. The machine report records the source
identities and result denominators. To replay from the repository root into a
new disposable output directory:
PYTHONDONTWRITEBYTECODE=1 \
PYTHONPATH=packages/bijux-pollenomics/src:packages/bijux-pollenomics-dev/src \
.venv/bin/python docs/report/regions/nordic/dalmutladdo-age-interval-sensitivity/replay.py \
freeze --repository . --output artifacts/jobs/dalmutladdo-age-interval-check
PYTHONDONTWRITEBYTECODE=1 \
PYTHONPATH=packages/bijux-pollenomics/src:packages/bijux-pollenomics-dev/src \
.venv/bin/python docs/report/regions/nordic/dalmutladdo-age-interval-sensitivity/replay.py \
score --repository . --output artifacts/jobs/dalmutladdo-age-interval-check
The source does not identify these LegacyAge ages as the historical LandClim ages. Two supported samples cannot establish Nordic transfer quality or chronology robustness over the other 20 bins.