Results and model use

Implementation correspondence and archive differences

The installed Python publication profile matches the independent author-R example on 14,074 supported mean and standard-error values within predeclared tolerances. Maximum residuals are about 1.84e-13 for means and 7.16e-14 for standard errors in fraction units. The raw-to-prepared replay also reproduces the 17 released example files byte for byte.

We separately ran the 2016 author REVEALS function and a pinned later Petr Kuneš LRA source on one author 225 BP window in four site configurations. The source comparison retains 124 paired cover values and 124 paired standard errors that agree at CSV precision, plus identical 1,000-draw full-grid bootstrap indices. Our historical native profile matches the tagged R full-grid result within declared tolerances when given those indices. This is conditional numerical correspondence within a shared code lineage; it does not certify corrected uncertainty or Nordic ecological accuracy.

We separately compared RcppDE and Python LOVE optimization on three declared shared objectives. Both attained objective gaps below 1e-4 from independent optima, and each selected candidate evaluated in the other implementation within 1e-10. The result rows include a case with equally good but different vegetation vectors. This checks bounded objective quality and arithmetic; it does not show identical random searches, global convergence on other inputs or ecological accuracy.

For the nominated 2024 LOVEr/LOVEoptim paper, all 30 printed fall-speed, pollen-productivity and productivity-error values match the later packaged DISQOVER example by row position. Nine of ten taxon labels match exactly; the paper uses Grasses and the package uses Gras. The example does not contain the eight-site landscape, pollen and random draws needed to rerun the paper's Figure 2 RMSE. Its reported preferred ring range is therefore a result of that simulation, not a selected setting for Nordic lakes. The paper-result matrix records the input match and the unavailable output separately.

The paper/archive comparison finds a genuine difference on the one exact shared 100–350 BP window: archived means differ from the author replay by up to 9.20 percentage points in grid 512 and 10.55 in grid 562. Only one of 25 intervals matches exactly. The PFT mean-arithmetic replay closes the archived ISTS/TBS means but does not identify their biological definition or standard-error aggregation. The source revision and archive transformation remain open; the differences are not ecological accuracy estimates.

Across all 10,500 archived grid/window mean rows, ISTS equals the sum of Corylus.avellana, Fraxinus and Quercus.deciduous within numerical tolerance. The paper assigns deciduous Quercus type to TBS, while this archived output places its mean in ISTS. All checked aggregate standard errors satisfy necessary component-error bounds, but the archive does not expose the covariance needed to reproduce its aggregation. We keep the paper/archive classification difference and the unresolved ISTS meaning visible; numerical agreement does not authorize a taxon relabeling.

HYDE as modeled context

The retained HYDE country table has 84 source-reported cropland areas for Denmark, Finland, Norway and Sweden at seven native CE labels and three source scenarios. Selected native ASCII grids reproduce those country totals within the declared area tolerance. This checks internal modeled-source accounting. HYDE cropland is neither measured vegetation nor cereal pollen, and the scenarios are not confidence bounds. Release-specific rights, formal CRS, calendar support and comparable class/footprint rules are still unresolved, so we have not run or scored a no-HYDE/HYDE pollen-model comparison. The source review records those exact limits.

Direct pollen and LRA forward pollen

The English constrained LOVE study uses 15 author sites in 1250–1750 BP and 30 source-bound pollen targets. It publishes three forward-pollen predictions for each of 14 held-out sites; one fold refuses after three training optimizations fail. The 42 predictions were frozen before target pollen was opened. On those same 14 sites and 30 taxa, fixed direct-pollen methods have lower mean error:

Fixed method	Direct pollen MAE	LRA forward-pollen MAE	Direct Hellinger	LRA Hellinger
Training mean	0.020496	0.026273	0.282082	0.360671
Nearest site	0.019552	0.021132	0.255767	0.286001
Inverse distance squared	0.018579	0.021088	0.243673	0.285415

MAE averages fraction errors over the named taxa; Hellinger compares closed compositions. Both are first averaged within each scored site. The paired MAE rows and Hellinger result retain the precise result identity. The direct route uses all 14 training pollen sites; the LRA route uses one regional anchor and 13 local reconstruction predictors. The sites and outcomes were previously explored. This result does not compare vegetation accuracy or establish the best method for new Nordic cores.

The refused BECKHAM fold remains refused after a source-bound analytic-gradient check. The check reproduced all six original ring-fit failures and recovered both HANGINGS fits with an exact derivative of the same LOVE objective. The four CUTHILL2 and LONGBREACH ring fits still failed their convergence test, so no BECKHAM LRA prediction or held-out score was added. A lower objective at a failed fit does not certify a reconstruction.

The site-influence result finds the lower direct mean Hellinger distance persists after any one scored site is omitted. Individual sites still disagree, especially for nearest site, where only seven of 14 favor direct. This is a descriptive sensitivity check on previously explored examples, not a confidence interval.

We also varied the declared English author lake radii by 20% in the actual LRA forward-pollen consumer, keeping fitted vegetation and regional cover fixed. The governed prediction freeze and scores cover 126 projections on the same 14 scored sites. For inverse-distance-squared vegetation, site-macro pollen MAE is 0.021032, 0.021088 and 0.021457 at 80%, 100% and 120% of the assumed radius. This diagnoses forward transport sensitivity on prior-exposed English examples; it neither calibrates a lake radius nor admits Swedish basins to LRA. On the same matched sites, our paired direct-pollen comparison finds lower aggregate direct-pollen MAE and Hellinger at all three radius factors for each fixed method. Individual sites disagree, and direct and LRA used different training information.

Additional Swedish lake transfer

Three further Swedish EPD lakes were selected from site and age metadata before their pollen counts were read. The three-lake transfer withheld Gilltjärnen, Lake Avegöl and Flinkasjön from the 81-site Denmark/Finland/Norway training set. All 63 site by 500 BP bins received three fixed 31-target pollen predictions, giving 189 matched score rows. Inverse distance squared had the lowest three-site mean absolute fraction error (0.017334), compared with nearest site (0.017932) and training mean (0.020682); nearest site performed better at Lake Avegöl. These are previously exposed development outcomes for observed pollen. They do not admit these lakes to LRA or measure vegetation accuracy.

Swedish training scope on matched pollen targets

The whole-site comparison holds out each of three Swedish sites. It compares the other two Swedish sites with those same two plus 81 Denmark, Finland and Norway sites. On the 39 bins shared by both scopes, inverse-distance-squared equal-site observed-pollen MAE is 0.032846 with Sweden-only training and 0.023226 with four-country training. Sweden-only training refuses 15 of 54 candidate bins; four-country training predicts all 54. The paired rows retain the exact common support. These three Swedish sites and their outcomes were previously explored, and most of the mean improvement comes from one site. The comparison does not establish independent prediction skill or LRA vegetation accuracy.

Swedish age and distance sensitivity

We used the same three whole-site Swedish targets to ask how observed-pollen predictions change when training pollen comes from an adjacent nominal 500-year bin, only the nearest five source sites, or source sites at least 100 km from the target center. The source-bound sensitivity retains 486 attempted site-bin/policy queries, 473 supported predictions and 13 refusals. On 52 bins shared with the unshifted all-site reference, excluding sites within 100 km raises inverse-distance-squared equal-site pollen MAE from 0.023151 to 0.029171. The paired score rows show the site and bin effects. Three previously explored sites and nominal source-bin shifts cannot establish independent spatial skill, chronology uncertainty, LRA accuracy or vegetation accuracy.

Swedish chronology-block development result

We also withheld each Swedish target site and predicted its 500 BP pollen bin using training pollen exactly 1,000 nominal BP older, excluding the target and adjacent source bins. All 54 target bins received three frozen predictions and matched same-bin scores, with no refusals. Inverse-distance-squared equal-site pollen MAE changed from 0.023647 to 0.024023; nearest-site and training-mean changes were small and in the opposite direction. The chronology-block report keeps the 162 paired rows and count denominators. These source-assigned bins do not establish nonoverlap of the cores' true age distributions. This is a prior-exposed observed-pollen development comparison, not a test of LRA or measured vegetation.

Abborrtjärnen with an alternative chronology

The LegacyAge source join matches published age estimates to all 107 Abborrtjärnen pollen sample depths and 71 of 72 Dalmutladdo depths by exact S5 row. The remaining 4.5 cm Dalmutladdo sample has a source-bound derived age-summary allocation from the adjacent 4 and 5 cm rows: minimum −94, maximum 541.5, median 257 and mean 215 calibrated years BP. This is not an exact S5 row or joint age posterior. These are alternative ages, not the historical LandClim age vector.

In a separate three-site development sensitivity, we admitted a sample to a 500-year bin only when its entire reported LegacyAge 95% interval fit inside that bin. Only eight Abborrtjärnen samples qualified for four shared bins. Four fixed pollen predictors scored the same 12 site-bin targets each; inverse-distance-squared site-macro MAE changed from 0.060058 to 0.059762 on that matched subset.

The other two sites retain point ages, and the intervals do not provide a joint age posterior. This sparse, previously exposed 16-label pollen study does not establish a better chronology, independent accuracy or LRA inputs.

A conditional membership account also bounds the fraction of all native pollen counts represented by the selected 16 taxa, using the same 107 samples. Its four per-bin lower and upper bounds are separate limits. A joint-membership check finds that none of the 16 combinations of those endpoints can occur together when each sample has one possible age within its reported interval. The individual bounds remain valid, but combining their endpoints would imply incompatible sample assignments. These conditional count-fraction checks supply neither age probabilities nor reconstructed vegetation.

For the Norwegian Dalmutladdo core, a separate age-interval support check keeps only two of 72 pollen samples: their full reported LegacyAge 95% intervals stay inside their original 500-year bins. Four previously frozen Norway-excluded methods provide eight predictions across those two bins; 20 previously scored bins have no eligible sample. The two samples do not change bins, so the paired same-sample scores are identical under the nominal and strict rules. Differences from the original full-bin scores reflect fewer counted samples, not a new model result or LRA validation. The separate [31-author-target analysis] (../../../report/regions/nordic/dalmutladdo-author-target-age-interval/README.md) reuses six frozen predictions on those two bins. Its inverse-distance-squared two-bin MAE is 0.006445, compared with 0.006341 for the parent's full compositions on the same bins. Those scores have a different taxon axis and denominator from the 16-label result; the two studies cannot be pooled.

At Danish Solsø, a [primary LegacyAge depth join] (../../../report/countries/denmark/solsoe-source-lineage/legacyage-chronology/README.md) matches all 92 Neotoma samples across two cores. The 1981 and 1983 core chronologies have different agreement with these alternative age intervals. The join does not recover LandClim II's published selection of 79 samples or its historical basin inputs. A separate [Denmark-held-out pollen sensitivity] (../../../report/countries/denmark/solsoe-source-lineage/legacyage-interval-sensitivity/README.md) retains 15 of 91 comparable samples whose full age intervals fit their original bins. Eight of 24 scored bins and 32 frozen four-method predictions remain supported. This is a previously exposed direct-pollen result on one physical lake, with no new LRA or vegetation validation claim. The distinct [31-author-target Solsø sensitivity] (../../../report/countries/denmark/solsoe-source-lineage/author-target-age-interval/README.md) uses author-assigned pollen counts and 24 frozen three-method predictions on the same 15 supported samples and eight bins. Its nearest-site eight-bin MAE is 0.025673, compared with 0.025513 on the parent's complete counts in those bins. We do not pool that score with the 16-label result or infer independent temporal skill from this one previously exposed lake.

A newly scored Swedish lake

We applied three fixed observed-pollen predictors to Lake Sämbosjön, a physical site absent from the earlier target studies. All 22 candidate 500-year bins were predicted and scored. Inverse-distance-squared equal-bin MAE was 0.022991 on a closed 16-label pollen response containing 224,291 of 393,822 native counts (56.95%). The governed study keeps the source identity, frozen predictions, scores and count denominators. Aggregate target count support was inspected before prediction, and aggregate scores were inspected before a later protocol annotation correction. The prediction bytes stayed fixed; this one-site result is descriptive development evidence, not independent confirmation or vegetation accuracy. A separate source age-bound screen finds that 58 of 111 sample intervals cross a nominal 500-year bin boundary, with at least one crossing in each scored bin. The published MAE remains a point-age-bin result, not an age-uncertainty-adjusted score. The contained-age sensitivity keeps 53 samples in 21 bins with the same frozen predictions. Primary MAE is 0.023054 versus 0.022881 for original counts on those same bins; bin 1 has no contained sample and is refused. Marginal source-bound containment is a development sensitivity, with no joint age posterior or confidence interval.

A second newly scored Swedish lake

At Lilla Gloppsjön, three fixed methods predicted every one of 23 nominal 500-year pollen bins before the target counts were opened for scoring. All 69 method/bin scores completed. On the same 16-label closed pollen response, the prespecified inverse-distance-squared method has equal-bin MAE 0.017295 and Hellinger distance 0.237576. Training mean has lower MAE (0.016272), while nearest site has lower Hellinger distance (0.228724); the method order depends on the metric at this lake. The research report and score table give the denominators and per-bin results.

Only 55,638 of 127,027 native pollen NISP counts lie on the exact 16-label axis (43.80%). Two of 86 source samples are excluded from age-bin scores due to negative source age bounds, with their counts retained separately. The separate contained-age sensitivity keeps 42 of 84 age-selected samples in 20 of 23 bins. On those same 20 bins, the frozen primary method's MAE is 0.019316 on contained samples versus 0.018367 on the original full-bin observations. The remaining three bins refuse; this conditional check is not a joint age uncertainty model. The training source-validation screen had previously parsed this target's source partition, so this is descriptive new-site development evidence, not an untouched independent confirmation. It does not evaluate vegetation or LRA.

What two lake pollen series can and cannot say

We compared observed Alnus and Plantago lanceolata pollen fractions at Bussjösjön and Bökesjön in 13 common 500-cal-BP bins. The lake trajectory report retains each site's source counts, age support and 48 taxon/threshold/interval comparisons. Some site changes agree in direction and others contradict. A review of the source counts found that the 48 classifications are 24 distinct taxon transitions evaluated at two thresholds: one Alnus contradiction persists under both thresholds and denominators. A different Bussjösjön Alnus transition reverses direction when the 16-taxon subtotal replaces all native pollen counts. Neither comparison is a vegetation-cover change. The 1000-year binning and alternate nearby-site circles expose time and location sensitivity. Giving both lakes equal weight versus pooling their pollen counts changes 10 of 48 all-native direction classes and 9 of 48 closed-16 classes; four comparisons in each denominator reverse between a qualified increase and decrease. The matching count denominator is used for each composition.

Independent dated SEAD records describe nearby archaeological context, but their feature counts are not vegetation or land-use measurements. The LandClim model grid includes these pollen sites and cannot independently validate them. The supported answer is a bounded description of two pollen series; whether local land use caused their changes remains unresolved.

Parameters and next experiments

Measured vegetation evidence

Our EMPD2–NFI source screen admits zero matched vegetation pairs for repository prediction accuracy. A Färskesjön plot provides measured tree volume but is a 7 m radius plot whose source point is 427 m from the lake pollen point. Historical Småland maps give map coverage and shared pollen samples for three bogs, without the full numerical vegetation fractions. A published southern Swedish figure supports a bounded nine-label comparison of its authors' estimates, not a score of this repository's predictions. These source results answer different questions; none supplies our missing independent vegetation outcome. An authenticated Namibian source join adds ten modern soil-pollen sites paired with satellite-image-derived vegetation areas at 15 reported radii. It is a separate, previously inspected development source with no model score; its savanna soil samples do not validate Nordic lake reconstructions. A descriptive radius comparison finds mean five-taxon pollen-to-area Hellinger distances of 0.3664 at 100 m, 0.3427 at 400 m and 0.3777 at 1,500 m across the same ten sites. The source tables do not settle ring versus cumulative area semantics or a suitable pollen-source radius.

Model settings

For the English LRA run, published PPE and fall speeds were fixed after a declared one-taxon transport exclusion. A simplex-constrained LOVE fit used a uniform-start SLSQP search, 300 iterations, 16,000 objective evaluations and 1e-8 tolerance. Three of 195 training fits failed; a bounded 600-iteration check still refused those cases. The original author-R optimizer differs. These settings are a reproducible development profile, not universal ideal parameters. The two-case synthetic PPE and transport sweep completed 12 development trials and ranked one candidate first within that set. Its independent pollen generator and installed REVEALS use different transport laws, so the winning candidate may compensate for model mismatch. We do not recommend its PPE or transport settings for a real core. The repository's model-use guide sets out the target-specific choices, source gates, numerical settings, uncertainty scope and refusal conditions in one place.

One bounded author direct-pollen CLR-RBF search has lower common-support MAE than inverse distance squared but less coverage; the development account specifies its candidate set, denominator and prior exposure. Country-excluded Nordic tuning did not establish a general advantage over the fixed IDW baseline. A new independent cohort must be qualified and predictions frozen before outcome inspection. Measured vegetation validation needs an admitted reference with matching site, time, taxa and spatial support.