Sämbosjön pollen prediction on contained age support¶
How much does the fixed observed-pollen score change when samples whose reported age interval crosses a nominal 500-cal-BP boundary are excluded? This sensitivity retains 53 of 111 selected source samples in 21 of 22 candidate bins. The primary inverse-distance-squared prediction has equal-bin MAE 0.023054 and Hellinger distance 0.185741. The original observations on these same 21 bins have MAE 0.022881 and Hellinger distance 0.183698. This small descriptive difference does not estimate an age posterior or establish ecological accuracy.
Source support and count denominator¶
The physical site, SAMBOTOT core, pollen dataset and selected MADCAP chronology are those in the parent study. The source age-bound screen provides younger and older marginal bounds, with no stated confidence level. We retain a sample only when its interval has positive-width overlap with just its nominal bin. An older endpoint exactly touching the next bin contributes no width there, matching that source screen. No crossing sample is reassigned. Point ages, training sites, labels and the 66 frozen predictions stay fixed; no method is fitted or selected after scoring.
The 111-row sample ledger partitions 393,822 native NISP pollen counts into 185,150 retained and 208,672 crossing-sample counts. Retained samples contribute 108,163 counts on the fixed 16-label axis and 76,987 outside it; excluded samples contribute 116,128 and 92,544 respectively. These four masses close to the parent's 224,291 closed-axis and 169,531 outside-axis counts. The fraction scores concern the retained closed pollen subcomposition, rather than all pollen or vegetation.
The coverage ledger keeps all 22 attempted bins. Bin 1 (500–1000 cal BP) has no wholly contained sample and is explicitly refused; its three method rows have null scores. Bins 2–22 supply 63 scored method/bin pairs. Refused bins are neither zero-error observations nor omitted successes.
Compare the same bins and predictions¶
| Fixed method | Paired bins | Contained MAE | Original MAE, same bins | Contained Hellinger | Original Hellinger, same bins |
|---|---|---|---|---|---|
| Inverse distance squared, primary | 21 | 0.023054 | 0.022881 | 0.185741 | 0.183698 |
| Nearest site | 21 | 0.026731 | 0.026545 | 0.238368 | 0.235900 |
| Training mean | 21 | 0.037673 | 0.037562 | 0.285562 | 0.284419 |
All 66 score/refusal rows and the machine report retain the full precision, original paired scores and exact input identities. Every nominal-bin count denominator and score was reproduced exactly before calculating the restricted response. Equal-bin means weight each supported bin equally; there is no between-site uncertainty estimate from this single core. The original all-22-bin primary MAE of 0.022991 answers a different support question and is not the matched comparator above.
These counts and parent outcomes had already been exposed. This is a one-site development sensitivity. Marginal containment cannot represent correlated age uncertainty, a confidence interval, an untouched confirmation, Nordic LRA or an independent vegetation comparison. The nearby LandClim SAM point remains a refused core-identity join. Source attribution and reuse conditions remain those of the parent Neotoma/EPD study.
Reproduce this sensitivity¶
The bijux_pollenomics_dev.source_completion.pollen_age_support.sambosjon producer verifies ten pinned parent/age-screen inputs, the preparation/freeze relations and original source-part hashes before reading target counts. It refuses changed inputs, duplicate predictions, altered physical-site exclusion, unusable contained count support and parent-score disagreement. It writes the four numeric files into a previously absent directory:
OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 .venv/bin/python -m \
bijux_pollenomics_dev.source_completion.pollen_age_support.sambosjon \
--repository . --output artifacts/jobs/sambosjon-interval-support/reproduction
A second independent invocation reproduced all four retained files byte for byte. Focused checks exercise matched-bin scores, count-mass closure, refused/null outcomes, duplicate prediction and sample-support rejection, input drift before count reads, and the source-screen boundary convention. Qualified scientific review and a collaborator reading exercise remain outstanding.