Sentinel-Wildlife: putting Florida's contaminant and pathogen surveillance of wild birds and amphibians on one map

Framing (non-negotiable). The unit of analysis is a species × Florida county pair, over the whole period the federation's wildlife record covers (1974-06-10 to 2024-05-13). Everything below is hypothesis generation for sampling design, not exposure assessment and not causal or clinical inference. Two distinct kinds of statement appear and are kept visibly apart throughout: measured — a contaminant concentration actually quantified in that species' tissue — and inferred — a species that is only phylogenetically close to one that was measured, which is a hypothesis about who might accumulate, supported by no sample. Host–pathogen links are literature-derived co-occurrence assertions, not experimental host-competence results. Observation counts are opportunistic community-science records (iNaturalist-derived) and index observer effort as much as animal abundance. Keep these caveats attached to every downstream claim.

Abbreviations. AI = avian influenza · Bd = Batrachochytrium dendrobatidis (amphibian chytrid fungus) · CUI = UMLS Concept Unique Identifier · DOID = Human Disease Ontology identifier · EEE = eastern equine encephalitis · EHR = electronic health record · FIPS = Federal Information Processing Standards (US geographic code) · HP = Human Phenotype Ontology · HPAI = highly pathogenic avian influenza · IAV = influenza A virus · KG = knowledge graph · MONDO = Mondo Disease Ontology · NCBITaxon = NCBI Taxonomy ontology · OKN = Open Knowledge Network · PFAS = per- and polyfluoroalkyl substances · PFBA/PFDA/PFDS/PFHpS/PFHxS/PFNS/PFOS = individual perfluoroalkyl acids · RCM = regional climate model · S2 = Google S2 geospatial cell hierarchy · SDoH = social determinants of health · SPARQL = SPARQL Protocol and RDF Query Language · UMLS = Unified Medical Language System · WNV = West Nile virus · WQP = US Water Quality Portal

1. Executive summary

Wild animals are read as sentinels for two things at once — the chemicals accumulating around them and the pathogens circulating through them — and the OKN federation contains both kinds of evidence. Joining them over Florida shows that, for this study area, the two systems do not overlap at all. The federation holds 5,205 wild bird and amphibian observation records (303 bird and 97 amphibian taxa, 11,654 individuals, 657 named places, 1974-06-10–2024-05-13) and zero contaminant measurements of any kind — biotic or environmental — anywhere in Florida. The intersection of the wildlife record and the contaminant record is not small; it is empty.

Across the whole federation, contaminant body burden has been measured in 67 animal taxa from 12 states, and exactly 2 of them are birds — mallard (Anas platyrhynchos) and Canada goose (Branta canadensis), both sampled only in Washington County, Minnesota, in 2022, with PFOS reaching 1,990 ng/g and 137 ng/g of tissue respectively. 0 amphibians have been measured. Both measured bird species do occur in the Florida record, so the entire measured evidence base for Florida's avifauna is two species sampled 2,000 km away. Everything else is inference: 27 Florida-recorded waterfowl and gamebird taxa sit within a genus, subfamily, family or superorder of a measured species and are therefore plausible sentinels for which no sample exists, and every one of the 97 amphibian taxa (76 of which resolve to NCBITaxon) has no measured relative anywhere in Amphibia.

The pathogen side is populated but disconnected. Of the 339 Florida taxa that resolve to an NCBITaxon term, 17 appear in the NIAID Data Ecosystem and 11 carry an explicit host–pathogen assertion in the biohealth graph, giving 22 taxon–disease pairs across 9 infectious diseases. Every one of those diseases has substantial human evidence in biohealth (9/9), but only 1 has any EHR phenotype evidence in oard-kg (salmonellosis, 128 phenotypes), only 1 has a biomarker (influenza, one variant), and 0 are present in spoke-okn's disease layer. The federation carries West Nile virus in 43 datasets and Bd in 2, and links 0 of them to any avian host.

Ranked by how much a single new sampling effort would tell us, the top counties are Orange, Miami-Dade and Palm Beach (16 of 64 counties in tier A), and the top species are Meleagris gallopavo (wild turkey — 8 infectious diseases, 3,086 human-evidence edges, never sampled), Cairina moschata and Anas platyrhynchos. 8 of 34 ranked species are the specific gap the study set out to find: hosts for a human pathogen that are absent from the contaminant record entirely. What this adds is a defensible, auditable target list — and the finding that a Florida sentinel programme would be starting from zero, not adding to a thin baseline.

2. Sources used

Twelve KGs were queried; all appear above and every claim traces to a logged query in the reproducibility record. Suppliers that the capability index named but that this study did not use are declared in §6.4.

3. Design & rules

The study area. Florida was chosen because it is where the federation's wildlife observations live. The wildlifekn graph stores places as free-text labels — a mixture of Florida city names and county names, with no state, no coordinates and no identifiers — so the only reproducible route to geography is the published label bridge to spatialkg county names. Running that bridge live returns 62 Florida counties, not the 63 in the crosswalk catalogue; the discrepancy is a defect in the published count rather than drift (see §6.5). Normalising Saint → St. recovers two more (St. Johns 12109, St. Lucie 12111), giving the 64 counties used here out of Florida's 67. City-type locations (some 757 records at 474 amphibian and 283 bird city labels) cannot be resolved to a county at all and are excluded from every county-level statement, though they are included in the state-wide inventory and the temporal series.

Two evidence classes, never merged. A species is measured only if sawgraph holds a contaminant concentration in a biota sample of that exact NCBITaxon id. Everything else that shares a clade with a measured species is inferred and tiered by how close that clade is (§4). The exact-id overlap between the Florida wildlife record and the contaminant record is 2 taxa; a naive clade-expanded join returns 339, because sawgraph also asserts phylum- and class-level ancestors — that number is an artefact of the hierarchy, not evidence, and is not used (§6.5).

Host–pathogen links. Two independent routes were run. nde links a host taxon to a dataset's MONDO health condition and infectious agent; the MONDO terms were filtered to genuine infectious disease by rdfs:subClassOf* closure under MONDO:0005550 in ubergraph. biohealth asserts PROCESS_OF between a disease concept and an organism concept, and its organisms are matched to the wildlife record by NCBITaxon label. Both are observational co-occurrence, and the biohealth layer is literature text-mined — two of its edges are flagged as probable extraction artefacts (§6.3).

Human evidence. For each mapped disease, four independent human-side layers were interrogated — biohealth (literature/clinical associations, via the MONDO↔UMLS hasDbXref bridge), oard-kg (EHR phenotype co-occurrence, UNIONing both reified roles), biomarkerkg (biomarkers), prokn and spoke-okn (presence in a curated disease layer).

Scoring. Counties are scored on six min–max-normalised axes (sentinel-capable species richness, best proximity tier present, pathogen-host species count, log EPA PFAS-facility count, total observed species richness, adult asthma prevalence); species are scored on five (number of infectious diseases hosted, proximity tier, human-evidence weight of those diseases, Florida observation footprint, county spread) with a multiplier applied when the species is a pathogen host that has never been sampled. The exact weights, formulas and thresholds are in the reproducibility file; §3 states only what is in the score and why.

Inventory (rebuilt live).

The asymmetry in panel A is the study's central result in one image: the wildlife layer and the human-disease layer are both well populated for this study area, and the contaminant layer — the thing wild animals are most often proposed as sentinels for — contains nothing at all.

4. Confidence tiers

Two ranked products are reported, each tiered A/B/C by score quartile (A = top quartile, B = 40th–75th percentile, C = below the 40th percentile). Because every Florida county has zero contaminant samples, the sampling-deficit term is uniform across the study area and does not discriminate between counties: the county ranking is driven by biological content and environmental pressure, not by differential prior sampling.

Distribution — counties: 16 A, 22 B, 26 C of 64. Species: 9 A, 11 B, 14 C of 34 ranked species, with 2 higher taxa (Anura, Amphibia) held out of the ranking because they are not species-level sampling targets.

5. Findings by axis

5.1 Which species are observed, where, and how the record changed

The record is two datasets in one graph. The amphibian series runs from a single 1974 record to a hard stop at the end of 2018, with 1,709 of its 2,723 records (1709, 63%) dated in that final year; the bird series runs to May 2024 and grows monotonically to a peak of 575 records in 2023. The two clades are also geographically different: bird county-labels are all Florida, whereas the amphibian layer spans the northern Gulf coastal plain, with roughly 72 county labels belonging to Georgia, Alabama or Mississippi. This is why the amphibian county numbers must be read with the homonym caveat in §10.

At the taxon level, the most-recorded amphibians are Osteopilus septentrionalis (277 records, 1,712 individuals), Anaxyrus terrestris (276 / 1,343) and Hyla cinerea (223 / 860); the most-recorded birds are Ardea alba and Ardea herodias (56 records each), Pandion haliaetus (47) and Anhinga anhinga (46). Full per-species and per-county tables are in the workbook.

The growth in both panels tracks community-science participation, not wildlife abundance — the provenance links on every statement are iNaturalist observation URLs. The operational consequence is that the amphibian layer is a historical baseline that stops six years before the bird layer ends, so any joint bird–amphibian sampling design cannot treat the two as contemporaneous.

5.2 Who has a measured contaminant body burden — and who is only inferred

Federation-wide, sawgraph holds PFAS measurements in biota for 67 taxa and 2,284 samples across 12 states — overwhelmingly fish (62 of the 67 taxa, 2,682 taxon–sample assignments), plus white-tailed deer (91), two marine bivalves (40) and two birds (10). There are no amphibians. Mallard was sampled four times and Canada goose six times, all in Washington County, Minnesota (the 3M "East Metro" PFAS study area), in June and August 2022. Mallard PFOS ranged 875–1,990 ng/g and Canada goose 13.3–137 ng/g, with a long tail of shorter-chain analytes (PFHpS, PFHxS, PFDA, PFBA, PFNS, PFDS) detected in the same tissues.

Panel B is the sampling gap in its bluntest form: eleven states with anywhere from 6 to 1,714 biota samples, and the state that is simultaneously an avian-influenza flyway, a PFAS hotspot and an amphibian-disease hotspot has none.

5.3 Phylogenetic inference: which species are plausible sentinels with no sample

Anchoring on the two measured species, ubergraph's NCBITaxon closure places their genera (Anas, Branta), subfamilies (Anatinae, Anserinae), family (Anatidae) and superorder (Galloanserae) over the Florida record. 27 recorded taxa fall into one of those tiers: 5 congeners of the mallard or goose (Anas acuta, A. castanea, A. crecca, A. fulvigula, Branta leucopsis), 12 in the same subfamily (Aix sponsa, Bucephala albeola, Cairina moschata, Lophodytes cucullatus, Mareca americana, Mergus serrator, Nomonyx dominicus, Oxyura jamaicensis, Spatula clypeata, S. discors, Anser anser, A. cygnoides), 6 more in Anatidae (Alopochen aegyptiaca, three Aythya, two Dendrocygna) and 4 gamebirds in Galloanserae (Colinus virginianus, Gallus gallus, Meleagris gallopavo, Pavo cristatus).

The distribution matters as much as the tiers: only two taxa are above the line, 27 are within a plausible extrapolation distance, and 310 of the 339 NCBITaxon-resolved taxa (234 birds and all 76 resolvable amphibians) have no measured relative closer than class or phylum. Any claim that Florida's amphibians are contaminant sentinels currently rests on nothing in this federation.

5.4 Which species are pathogen hosts, and what human disease each pathogen maps to

Of the 339 taxa that resolve to NCBITaxon, 17 intersect nde by exact taxon id — 12 birds and 5 amphibian taxa — but the intersection is dominated by laboratory use rather than wild-host surveillance: Gallus gallus alone accounts for 1,158 datasets and pulls in a long list of unrelated human diseases, which is a study-organism signal, not a host signal. Restricting nde's MONDO conditions to true infectious disease leaves a small, interpretable set, in which mallard is the only wild taxon linked to avian influenza (MONDO:0018695), influenza, arbovirus infection and infectious disease, and house finch (Haemorhous mexicanus) is linked to conjunctivitis — the well-known Mycoplasma gallisepticum system.

The biohealth route is richer and reaches 11 Florida taxa and 22 taxon–disease pairs. Wild turkey is the hub, asserted as host for eight diseases including West Nile fever, avian influenza, salmonellosis, Newcastle disease and coccidiosis; rock pigeon carries three, European starling two (EEE and conjunctivitis), and mallard, Muscovy duck, lesser scaup, great horned owl, common grackle and eastern newt one each.

Panel B is the second disconnect. Every mapped zoonosis has a substantial human literature footprint in biohealth — from 71 edges for EEE to 1,368 for influenza — but the clinical layers are almost empty: salmonellosis is the only disease with EHR phenotype associations (128 in oard-kg), influenza the only one with a biomarker (a single ROBO2 variant), and none of the nine appears in spoke-okn's DOID disease layer. So the chain wild host → pathogen → human disease → human clinical evidence is completable for exactly one of the nine diseases.

5.5 The place-based picture: health, environment, social, climate

For the 64 observation counties, spoke-okn supplies 140–151 chemicals recorded as found in the county and 853–1,140 SDoH measures per county; none of those chemicals is a PFAS, so the absence of Florida PFAS data in sawgraph is not compensated elsewhere in the federation. The available health and social measures are adult asthma prevalence (7.5% in Monroe to 10.4% in Gadsden, 2019) and per-capita income ($15,532 in Hamilton to $47,382 in Monroe, 2020). fiokg gives 4,118 EPA-PFAS-Facility records across the state, from 1 in Liberty County to 279 in Hillsborough — the study's proxy for where PFAS contamination is plausible. This is the PFAS-relevant subset of EPA's Facility Registry Service; the full registry is far larger but counts every site EPA or a state programme has ever tracked, so it indexes economic activity rather than PFAS risk and is not used here (the check that established this is in the reproducibility record). Uninsured rate and children-in-poverty are present as SDoH concepts for all 67 counties but their values did not resolve on the reified-value path used here and are not reported.

The climate record reaches 35 of 64 counties: climatemodelskg holds 246 Florida GeoNames places in 36 Florida counties (35 of them inside the study area), covered by three CORDEX regional climate models (CRCM5, HIRHAM5, REMO2015) and mentioned by 97 distinct climate publications across 125 place–paper mentions, concentrated in Miami-Dade (50 mentions), Brevard (21), Orange (14) and Palm Beach (9). The remaining 29 counties — including several that rank highly on wildlife content, such as Wakulla — have no place in the climate graph at all, so the climate context is systematically absent from exactly the rural, low-facility counties where the wildlife record is richest.

Geography is reported in one place only, as an interactive OpenStreetMap-tiled map of the 64 counties (see the companion file Sentinel-Wildlife_county_map.html, also embedded in the HTML report). Each marker is clickable and carries that county's rank, score, tier, sentinel-capable and pathogen-host species counts, PFAS-facility count, asthma prevalence and contributing sources. County anchor points are the mean position of three S2 Level-13 cells (minimum, maximum and a sample of the cell ids spatialkg records as contained by that county) — a point inside the county, not its centroid.

6. Domain analyses

6.1 The county gap: sentinel-capable richness where nothing has been sampled

Because the sampling deficit is total and uniform across Florida, the county ranking measures biological and environmental content. Orange County leads decisively: 11 sentinel-capable taxa (the most of any county, including the mallard itself, so it is the one county where a measured species and a first-ever Florida sample coincide), four pathogen-host taxa, 110 observed taxa and 210 PFAS facilities. Miami-Dade and Palm Beach follow on a similar profile of high richness and high PFAS-facility pressure. Wakulla at rank 7 is the informative outlier — 73 observed taxa, five sentinel-capable taxa and two pathogen hosts on only 4 PFAS facilities, i.e. high biological value with almost no PFAS-facility pressure, which makes it the natural low-exposure reference site rather than a hotspot candidate. Four counties (Brevard, Alachua, Leon, Citrus) carry three or more sentinel-capable taxa and no pathogen-host taxon, which is a contaminant-only opportunity.

6.2 The species gap: pathogen hosts absent from the contaminant record

8 of the 34 ranked species are hosts for a human pathogen and have never been sampled for any contaminant — the precise gap the study set out to find. Meleagris gallopavo (wild turkey) is first by a wide margin: eight asserted infectious diseases whose human concepts carry 3,086 biohealth evidence edges between them, 12 Florida records across 9 counties, and a proximity tier of only I4 (same superorder as the mallard), meaning even its inferred contaminant relevance is weak — a species where the pathogen case is strong and the contaminant case is entirely unbuilt. Cairina moschata (Muscovy duck) is second: an Anatinae subfamily member, so a much better contaminant extrapolation, an influenza host assertion, and the largest Florida footprint of any waterfowl in the record (38 records in 6 counties). Notophthalmus viridescens (eastern newt) is the only amphibian in tier A — 23 records across 17 county labels and a coccidiosis host assertion, with tier Z proximity, i.e. no measured relative anywhere in its class.

6.3 Data-quality flags raised by the analysis

Three assertions were flagged rather than scored. The biohealth layer asserts Anura → influenza and Amphibia → salmonellosis; the first is almost certainly a text-mining artefact (frogs are not influenza hosts) and both are higher taxa rather than species, so both were held out of the species ranking and are reported separately in the workbook. Separately, biohealth names Colinus virginiuanus as a West Nile fever host — a misspelling of Colinus virginianus — which means the bobwhite quail, a species present in the Florida record, is invisible to the label bridge. That is one demonstrated false negative in a bridge whose failure mode is silent, and it is the reason §10 treats the 11-taxon host set as a lower bound.

6.4 Analysis families: what was run and what was deliberately skipped

find_context_sources was queried for the four context types this study needs, and every supplier it returned is accounted for here rather than silently dropped.

Two analysis families that a biomedical OKN study would normally run were not run, deliberately: functional enrichment (GO and Reactome) and drug/target linkage, because this study has no gene or protein foreground — its entities are whole organisms, places and diseases. Stating that explicitly matters more than the omission: a silently missing enrichment section would read as coverage.

6.5 Two federation-level defects surfaced by this analysis

First, the published verified_count for crosswalk L8 (wildlifekn × spatialkg on county FIPS) is 63; the true number of Florida counties is 62. Re-running the skeleton without the 12-prefix filter reproduces 63, of which one distinct value is the literal string https — a spatialkg region IRI that does not match the administrativeRegion.USA. pattern, so the SUBSTR/REPLACE in the published skeleton yields a non-FIPS token that the COUNT(DISTINCT) then counts as a county.

Second, the taxon clade-expansion trap is unusually stark here. wildlifekn × sawgraph overlaps by 2 taxa on exact NCBITaxon id but by 339 after subClassOf* expansion — i.e. all 339 resolvable wildlife taxa nest under some sawgraph clade, because sawgraph's WQP taxa carry materialised ancestors up to Chordata. Reporting 339 as "species with contaminant data" would inflate the measured evidence base by a factor of 170. The clade number is used nowhere in this report except as this cautionary note.

7. Discussion

Three findings compose into one picture. The wildlife layer is real but effort-driven and temporally split; the pathogen layer is real but wired to laboratory and literature evidence rather than to wild hosts; and the contaminant layer, for this study area, does not exist. A Florida wildlife-sentinel programme designed from the federation would therefore not be extending a baseline — it would be creating the first data point, and the federation's role is to say where and on which animal that first data point buys the most.

For contaminant surveillance, the strongest targets are the counties that combine sentinel-capable waterfowl richness with high facility pressure — Orange, Miami-Dade, Palm Beach — paired with Wakulla as a low-pressure reference. On the species side, the extrapolation is only defensible within Anatidae, which makes Muscovy duck the best contaminant-sentinel candidate: subfamily-level proximity to the measured mallard and the largest Florida waterfowl footprint in the record. For zoonotic surveillance the ranking points elsewhere: wild turkey carries by far the densest host–pathogen evidence but is phylogenetically remote from anything measured, so it is a pathogen target, not a contaminant one. Sampling both in Orange County would test the two axes in one effort.

Three testable predictions follow. (1) If Muscovy ducks are sampled in Orange, Miami-Dade or Hillsborough County, PFOS will be detectable at concentrations within the range measured in mallard and Canada goose in Minnesota — the subfamily-proximity hypothesis's first real test, and the one that would either license or kill the whole inference ladder in Figure 4. (2) Amphibian PFAS burdens in Florida will be non-zero and measurable, because the absence in Figure 3A is a graph-coverage fact and not a biological one (§8, Claim 8). (3) Bd, ranavirus and PFAS co-occur in the same Florida amphibian assemblages, which the federation cannot currently express at all — it holds Bd with no host and amphibians with no contaminant.

The decisions this supports are narrow and concrete: which counties to fund a first biota-sampling round in, which species to collect, and which single reference county to hold as a low-exposure control. What it cannot support is any statement about exposure levels, risk, or trends in Florida, because there are no Florida measurements to trend.

8. Comparison with prior work

Claims were checked against the primary literature retrieved with the PubMed connector, with full text read through the Paperclip corpus where the article was available there; the full per-claim record, with citations, is in Sentinel-Wildlife_literature_comparison.md.

All twelve claims were checked against PubMed abstracts and metadata. Claims 6 and 7 alone were verified against full article text (Ellis et al. and Fiacre et al., read through Paperclip): the starling paper's results section and its published peer-review exchange, and the statement in the WNV paper's body that birds are the principal amplifying hosts. Claims 4 and 5 rest on structured abstracts only — Avian Pathology and the Journal of Wildlife Diseases are paywalled and absent from both PMC and the Paperclip corpus — so the turkey/WNV contradiction in Claim 4 stands on the authors' own abstract conclusion rather than on their results tables, and should be read with that limit attached.

Where the KG evidence diverges from the literature. Two divergences are outright errors in the graphs: the Anura → influenza edge (Claim 10) and the Colinus virginiuanus misspelling that hides a real host species from the label bridge (§6.3). Two are differences of scope rather than error: the wild-turkey West Nile edge (Claim 4) is a true literature co-occurrence that biohealth faithfully records but that host-competence experiments do not support as a transmission role, and the starling edge (Claim 6) has the right species with a pathogen the corroborating literature does not address. The remaining divergences are coverage gaps — WNV without avian hosts (Claim 7), Bd without hosts (Claim 9), amphibians without contaminant measurements (Claim 8) — where the literature is settled and the graphs are simply empty. That distinction matters for how each is fixed: errors need correction upstream, scope differences need better edge semantics, and coverage gaps need data.

9. Full ranked results

The complete rankings are in Sentinel-Wildlife_results.xlsx — Ranked Results (counties), Species Ranking, plus one sheet per supporting extract (observation inventory, temporal series, contaminant record, host–pathogen links, human evidence, place context) and a Methods & Rules sheet. The intermediate TSV/CSV extracts are in data/.

Tip: click a column header to sort (the sources (n) column sorts by how many federation KGs support the row), use the search box for a county name, and use the pull-downs to restrict to a confidence tier or a best-proximity tier. Sources contribute as follows — wildlifekn the observations, spatialkg the county geography, sawgraph the (empty) contaminant record, fiokg the PFAS-facility pressure, spoke-okn the health and social measures, climatemodelskg the climate record where it reaches the county.

The ranking's shape is as informative as its order: the top four counties are separated by a wide margin from the rest, and the tail is not a set of low-value places so much as a set of poorly observed ones — Lafayette, Calhoun and Gilchrist rank last on 1–3 observed taxa each, which is a statement about survey effort, not about wildlife. Read operationally, the table says fund Orange first, hold Wakulla as the reference, and treat the bottom quartile as counties needing an observation survey before a sampling decision can be made at all.

A representative slice of the species ranking:

10. Summary of findings & limitations

Findings recap. For Florida, the OKN federation holds 5,205 wild bird and amphibian observation records and no contaminant measurements whatsoever — not in animal tissue, not in water, soil or air. Federation-wide, contaminant body burden exists for 67 taxa in 12 states, of which 2 are birds (mallard and Canada goose, Minnesota only, PFOS to 1,990 and 137 ng/g) and 0 are amphibians. 27 Florida taxa are close enough to a measured species for extrapolation to be arguable and 310 of the 339 resolvable taxa are not. On the pathogen side, 11 Florida taxa carry 22 host–disease assertions over 9 infectious diseases, all 9 of which have substantial human literature evidence but only 1 of which has EHR phenotype evidence and 1 a biomarker.

The gap, stated plainly: 8 species are hosts for a human pathogen and have never been sampled for any contaminant, led by wild turkey, Muscovy duck and eastern newt; and all 64 counties have zero contaminant samples, so the county ranking is a ranking of what a first sample would buy, with Orange, Miami-Dade and Palm Beach highest and Wakulla the best low-pressure reference. Two federation defects were surfaced along the way: the L8 crosswalk's county count is 63 where the truth is 62, and the wildlifekn × sawgraph clade-expanded overlap of 339 must not be read as 339 species with data.

Limitations.

1. The geographic bridge is label-based and demonstrably lossy. wildlifekn places are free-text names with no state and no coordinates. County attribution rests on exact name agreement with spatialkg, which fails on Saint vs St. (two counties recovered by declared repair) and cannot distinguish Florida's Jackson, Washington, Jefferson, Franklin, Liberty, Calhoun, Holmes, Walton, Madison, Taylor, Baker, Bay and Escambia counties from same-named counties in Georgia, Alabama and Mississippi. Because the amphibian layer demonstrably includes those states, amphibian county counts for homonym counties are upper bounds.
2. 757 records at city-type places are excluded from every county statement. They are counted in the state-wide inventory and the temporal series, so county and state totals do not reconcile by design.
3. The taxon bridge is also label-based, and one false negative is demonstrated (Colinus virginiuanus). The 339-taxon resolution and the 11-taxon host set are therefore lower bounds; subspecies collapse to species, and 61 of the 400 recorded taxa do not resolve to NCBITaxon at all.
4. Observation counts index observer effort. The record is iNaturalist-derived and grows with participation; the amphibian series ends in 2018 and the bird series in May 2024, so the two clades are not contemporaneous and neither supports an abundance trend.
5. The measured contaminant base is n = 10 samples of 2 species in 1 county of 1 state, on a single analyte panel in a single year, at a known industrial PFAS site. Extrapolating from it to Florida is a hypothesis about chemistry and physiology, not a measurement, and the phylogenetic tiers in Figure 4 are a proximity heuristic, not a toxicokinetic model.
6. Host–pathogen links are observational co-occurrence. The biohealth layer is literature text-mined; one edge is a demonstrated artefact, one contradicts host-competence experiments (§8, Claims 4 and 10), and absence of an edge is not evidence of non-host status.
7. nde overlap is dominated by laboratory use. Gallus gallus's 1,158 datasets make it a study organism, not a wild sentinel; conflating the two would badly distort any ranking that used raw dataset counts.
8. The human-evidence layers are shallow for zoonoses. oard-kg is rare-disease-oriented, biomarkerkg covers one of the nine diseases, and none appears in spoke-okn — so "does the disease show up in human clinical and biomarker evidence?" is answered no for eight of nine, which is a statement about these graphs and not about clinical medicine.
9. The place-based layer is coarse and partly unresolved. SDoH values are county-level annual figures; uninsured rate and children-in-poverty did not resolve on the reified-value path and are omitted. The PFAS-facility count is a presence proxy, not an emissions estimate, and does not weight a fluorochemical plant against a small user; the federation carries no PFAS release quantities for Florida, so no graded exposure term is possible.
10. The climate record covers 35 of 64 counties and is a gazetteer-plus-model-coverage layer, not climate observations; counties with no climatemodelskg place have no climate context at all, and that absence is correlated with rurality.
11. County anchor points are not centroids. They are the mean of three S2 Level-13 cell positions per county and are used only to place a marker inside the right county on the map.
12. The scores are a decision aid, not an estimate. Weights were set by the analyst to encode the question asked ("what would one new sample tell us?"); they are stated in the reproducibility file and a reader who weights differently will reorder the middle of both tables, though the top four counties and top three species are robust to reasonable reweighting.

11. Reproducibility

Everything needed to replicate this analysis — the originating prompt, the replicator specification (selection rules, thresholds, join recipes, scoring formulas, verified quantities and limitations), and every supporting SPARQL query verbatim with its row count and the pinned KG versions and timing — is in Sentinel-Wildlife_reproducibility.md, with the scripts in scripts/ and the intermediate extracts in data/.

12. References

Retrieved via the PubMed MCP connector. Full text of the two entries marked full-text-verified was read via the Paperclip MCP connector; every other entry was assessed from PubMed abstracts and metadata.

1. Dulsat-Masvidal M, et al. Assessing Contamination Profiles in Livers from Road-Killed Owls. Environ Toxicol Chem. 2024. PMID:38146916 · doi:10.1002/etc.5816
2. Sun J, et al. Perfluoroalkyl acids and sulfonamides and dietary, biological and ecological associations in peregrine falcons from the Laurentian Great Lakes Basin, Canada. Environ Res. 2020. PMID:32882236 · doi:10.1016/j.envres.2020.110151
3. Prosser DJ, et al. Maintenance and dissemination of avian-origin influenza A virus within the northern Atlantic Flyway of North America. PLoS Pathog. 2022. PMID:35666770 · doi:10.1371/journal.ppat.1010605
4. Wang C, et al. Infection of chicken H9N2 influenza viruses in different species of domestic ducks. Vet Microbiol. 2019. PMID:31176393 · doi:10.1016/j.vetmic.2019.04.018
5. Kunkel MR, et al. Susceptibility of wild turkeys (Meleagris gallopavo) to experimental West Nile virus infection. Avian Pathol. 2022. PMID:36102057 · doi:10.1080/03079457.2022.2123732
6. Malmberg JL, et al. Mortality in Wild Turkeys (Meleagris gallopavo) Associated with Natural Infection with H5N1 Highly Pathogenic Avian Influenza Virus (HPAIV) Subclade 2.3.4.4. J Wildl Dis. 2023. PMID:37486883 · doi:10.7589/JWD-D-22-00161
7. Ellis JW, et al. Avian influenza A virus susceptibility, infection, transmission, and antibody kinetics in European starlings. PLoS Pathog. 2021. PMID:34460868 · doi:10.1371/journal.ppat.1009879 — full-text-verified (PMC8432794)
8. Fiacre L, et al. Evaluation of NS4A, NS4B, NS5 and 3'UTR Genetic Determinants of WNV Lineage 1 Virulence in Birds and Mammals. Viruses. 2023. PMID:37243180 · doi:10.3390/v15051094 — full-text-verified (PMC10222181)
9. Shu Y, et al. Legacy and Emerging Per- and Polyfluoroalkyl Substances Surveillance infrom Inlet Watersheds of Chaohu Lake, China: Tissue Distribution and Bioaccumulation Potential. Environ Sci Technol. 2023. PMID:37565447 · doi:10.1021/acs.est.3c02660
10. Zhu CH, et al. Effects of hibernation on the bioaccumulation and tissue distribution of per- and polyfluoroalkyl substances in frogs (Rana tigrina cantor) via skin exposure. Environ Pollut. 2025. PMID:40681077 · doi:10.1016/j.envpol.2025.126842
11. Rohonczy J, et al. Effects of perfluoroalkyl sulfonic acids on developmental, physiological, and immunological measures in northern leopard frog tadpoles. Chemosphere. 2024. PMID:39271078 · doi:10.1016/j.chemosphere.2024.143333
12. Rohonczy J, et al. The effects of two short-chain perfluoroalkyl carboxylic acids (PFCAs) on northern leopard frog (Rana pipiens) tadpole development. Ecotoxicology. 2024. PMID:38315267 · doi:10.1007/s10646-024-02737-z
13. Wiley DLF, et al. Leveraging machine learning to uncover multi-pathogen infection dynamics across co-distributed frog families. PeerJ. 2025. PMID:39897487 · doi:10.7717/peerj.18901
14. OKN federated SPARQL endpoint (FRINK), queried 2026-07-29 via the mcp-okn MCP server; KG versions and release dates as pinned in §2 and in the reproducibility record.