English author basin-radius forward sensitivity¶
We varied each declared circular lake radius by factors 0.8 and 1.2 around the source profile's radius, while holding the 42 published LRA vegetation fractions, regional cover, PPE, fall speeds, and training fits fixed. The original 14 scored sites have assumed radii from 10 to 200 m. We changed the first forward and inversion annulus inner edges and areas consistently; the outer edges and all later annuli stayed fixed. This is a target-basin forward transport diagnostic, not a new LOVE reconstruction or a claim that the author-example radii represent age-specific surveyed basins.
We froze 126 pollen predictions before opening the
target counts. The unmodified radius reproduces every published forward
pollen fraction exactly. We then scored all 126 rows on
the existing 30-taxon count denominator. The unmodified-radius MAE reproduces
the published 42 held-out scores within 1e-15.
| Fixed vegetation method | Radius factor | Site-macro pollen MAE | Site-macro Hellinger |
|---|---|---|---|
| Training mean | 0.8 | 0.026396 | 0.360975 |
| Training mean | 1.0 | 0.026273 | 0.360671 |
| Training mean | 1.2 | 0.026246 | 0.361927 |
| Nearest site | 0.8 | 0.020960 | 0.283415 |
| Nearest site | 1.0 | 0.021132 | 0.286001 |
| Nearest site | 1.2 | 0.021491 | 0.291210 |
| Inverse distance squared | 0.8 | 0.021032 | 0.283976 |
| Inverse distance squared | 1.0 | 0.021088 | 0.285415 |
| Inverse distance squared | 1.2 | 0.021457 | 0.289825 |
We retain the machine report and per-site scores because the aggregate moves differently by method and metric. The original study attempted 15 folds; BECKHAM still has no LRA prediction because three training fits refused, so this diagnostic scores 14 paired sites. All sites and target outcomes were already exposed in development. These shifts do not select a physical radius, quantify chronology or PPE uncertainty, establish vegetation accuracy, or admit a Swedish core to LRA.
Paired direct-pollen comparison¶
We joined the radius scores to the published direct-pollen MAE and Hellinger rows by target site, method and exact observed raw-row SHA-256. We checked that the unvaried LRA score and prediction identity reproduce the original comparison. The paired rows and summary cover 126 matched method/radius/site rows.
| Method | Radius factor | Direct minus LRA MAE | Direct minus LRA Hellinger | Sites favoring direct by MAE |
|---|---|---|---|---|
| Training mean | 0.8 / 1.0 / 1.2 | −0.005900 / −0.005778 / −0.005750 | −0.078892 / −0.078588 / −0.079845 | 13 / 13 / 13 of 14 |
| Nearest site | 0.8 / 1.0 / 1.2 | −0.001408 / −0.001580 / −0.001939 | −0.027647 / −0.030234 / −0.035443 | 6 / 7 / 7 of 14 |
| Inverse distance squared | 0.8 / 1.0 / 1.2 | −0.002454 / −0.002509 / −0.002878 | −0.040303 / −0.041743 / −0.046152 | 10 / 10 / 9 of 14 |
Negative aggregate differences favor direct pollen. We found that varying the assumed radius by 20% does not reverse the aggregate direct-pollen advantage on this matched English development support. Individual sites disagree, and the direct route used all 14 training pollen sites while LRA used a regional anchor plus 13 local reconstruction predictors. This is not an independent model ranking or vegetation-accuracy comparison.
We regenerated the source inputs and 15-site predictions from the qualified author snapshot. The source experiment hash matched exactly. The fresh prediction bundle has different run metadata, but all 1,260 retained vegetation and forward-pollen values match the governed prediction table exactly. Its 126 radius scores are byte-identical to the governed scores; only the source and freeze identities in the new run report differ. The producer verifies the candidate bundle members and every numerical prediction cell against that retained table before it freezes the radius projections.
To reproduce from the private author snapshot, run these commands from the repository root with an absent job directory:
job=artifacts/jobs/author-lra-radius-replay
source_root=/path/to/private/landclim/author-reference
PYTHONDONTWRITEBYTECODE=1 .venv/bin/python -m \
bijux_pollenomics_dev.essd_reproduction.author_lra_development \
--author-source-root "$source_root" --output "$job/source-inputs"
PYTHONDONTWRITEBYTECODE=1 .venv/bin/python \
docs/report/pollen-reconstruction/author-lra-constrained/prepare.py \
--source "$job/source-inputs" --destination "$job/constrained-inputs"
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.essd_reproduction.author_lra_run \
--inputs "$job/constrained-inputs" --cache-root "$job/cache" \
--output "$job/predictions"
PYTHONDONTWRITEBYTECODE=1 .venv/bin/python \
docs/report/pollen-reconstruction/author-lra-constrained/radius_sensitivity.py \
freeze --inputs "$job/constrained-inputs" --predictions "$job/predictions" \
--output "$job/frozen"
PYTHONDONTWRITEBYTECODE=1 .venv/bin/python \
docs/report/pollen-reconstruction/author-lra-constrained/radius_sensitivity.py \
score --inputs "$job/constrained-inputs" --predictions "$job/predictions" \
--frozen "$job/frozen/predictions.json" --output "$job/scored"
The producer authenticates the source experiment, prediction bundle members, every target geometry and each raw count row against the source declarations. It invokes the installed LRA forward consumer for every projection. Private author inputs stay outside the repository; the governed outputs preserve the original prediction and score identities.
To reproduce the downstream comparison from the retained source-locked scores without private raw files: