Our approach

What we are trying to answer

When you look up a place on our map, the useful question is not simply whether a species exists somewhere nearby. It is: what has actually been documented here, when, and by whom? ISITUSA brings scattered public evidence together so you can ask that question without having to search every agency database or collection yourself.

Our mission is to make invasive species research accessible, understandable, and useful to everyone. That requires preserving uncertainty as carefully as we preserve a finding. A county with many records may have been studied more intensively than its neighbor. A blank county may have little accessible evidence. Neither pattern, on its own, measures ecological health.

This paper follows evidence from its original source to the public site. It describes the implemented research system and its limits, rather than promising complete national coverage or continuous monitoring. The companion data-sources paper explains the kinds of research we use.

Start with a bounded question

Research begins with a defined source, species, and geographic scope. Our public geography uses U.S. counties and county equivalents. A species catalog gives records a shared identity, while separate state-applicability decisions describe whether a species belongs in a particular jurisdiction's research scope. An unknown applicability decision remains unknown; it is not silently excluded to make coverage look better.

The catalog begins with the US-RIIS lower-48 register. It is a starting framework, not a complete inventory of every introduced organism throughout the United States. Research coverage outside that original scope needs additional care. A legal listing, a biological occurrence, and an ecological impact are also different claims: a regulated species is not automatically documented in every county of the state that regulates it.

We record the requested scope before acquisition. This makes a completed search interpretable: it shows which source and pairs of species and counties were examined, rather than implying that every possible source has been exhausted. Jurisdiction configuration and the research protocols define that structure.

Find the source, then define its use

We look for public agency datasets, scientific collections, structured observations, field surveys, and explicit authority statements. Discovery is only the first step. Before a source becomes part of the evidence workflow, we identify its publisher, access method, geographic reach, supported claims, and known caveats in a source registry.

A source-specific importer, called an adapter, defines how its fields should be read. The adapter has a version, allowed parameters, rules for matching species and places, and a review gate. This matters because an identical-looking zero can mean very different things in two datasets. One may record a surveyed sample with no detection; another may simply leave an unreported county blank.

Where a provider offers a national archive, we can retain a versioned copy and examine bounded portions of it locally. Other sources are queried through an API, downloaded as a table, or reviewed as a document. Access failures, incomplete responses, and provider limits are recorded as limitations, not converted into biological conclusions.

Keep a trail back to what we read

A research run records the source, adapter version, parameters, requested scope, and acquisition context. Its receipt brings together accepted evidence, rejected candidates, review outcomes, warnings, and retained artifacts or their hashes. A hash is a fingerprint of a file: it helps establish which exact bytes a review used.

Retention is source-specific. The registry distinguishes versioned artifacts, hash-only retention, and material that is not retained. We do not assume every upstream page can be reproduced indefinitely. A citation and a retrieval record remain different from permission to republish the original material.

An accepted assertion connects a source record to a species, geographic scope, and supported claim. Where available, it carries the original record identifier and URL, observation date, source dates, acquisition details, and normalized payload fingerprint. Missing dates remain missing. A new download date must not make an old observation appear recent. The evidence model and adapter contract describe these fields.

Which species does the record describe?

Common names can refer to several organisms, and scientific names can change. Matching therefore follows each adapter's approved policy, using catalog identities, retained taxonomic mappings, or specifically reviewed interpretations. A similar-looking name is not enough. Ambiguous identities, unsupported ranks, and unresolved synonyms are held or rejected rather than guessed.

Which place does the evidence support?

A county code must correspond to the geographic registry. Some adapters require an explicit county identifier or unambiguous county text; others are specifically designed to resolve suitable coordinates against boundaries. Coordinate inference is not a universal fallback. For example, the retained EDDMapS county-aggregate and USGS NAS archive adapters do not use coordinates to repair missing county information.

A statewide mention does not become a record in every county. Survey areas, point locations, regulatory areas, and county claims retain distinct scopes. Retired county names, uncertain coordinates, and ambiguous boundary matches require an explicit supported resolution. Where a reviewed official determination is inherited from a broader jurisdiction, that derivation must remain visible.

Is this a separate observation?

The same specimen may appear in a collection portal, GBIF, and iDigBio. Stable source identities, lineage keys, and payload hashes help identify repeated evidence within the supported workflows. Cross-provider overlap can still be imperfectly resolved. Multiple links to the same specimen should not be read as multiple independent confirmations or a population count.

Keep different findings separate

The research model keeps historical occurrence, current determination, survey outcome, research progress, freshness, and review status on separate axes. That lets a later finding add context without erasing what happened before.

Recorded presence

Accepted evidence documents an occurrence in a supported place and time. It does not, by itself, establish abundance, an established population, current persistence, or presence throughout a whole county.

Survey non-detection

A source explicitly reports that a particular survey or sample did not detect the target. The conclusion stays tied to its method, area, and time. It does not erase an earlier occurrence.

Official absence or eradication

An explicit authority statement passes the applicable scope, date, corroboration, review, and conflict checks. The presentation distinguishes the authority's meaning, such as a pest-status or disease-distribution statement, from universal biological absence.

Researched but unresolved

Work has been recorded, but it has not produced a qualifying determination. A completed source screen is useful progress, not proof that the species is absent.

Not researched

The system has no qualifying research outcome for that pair under the applicable projection rules. Sparse data files can represent this default without storing an empty row for every possibility.

No matching rows, an empty map, and a rejected record never establish absence. Nor should a regulatory boundary or a report of no known detections be silently recast as a county field survey.

Automation helps search, transform, compare, and check large amounts of public information. AI-assisted research can also help interpret sources and draft explanations. Neither a successful import nor fluent writing makes a claim true. Acceptance depends on the registered source rules and supporting evidence.

The system distinguishes machine-validated, agent-reviewed, and human-approved evidence. These labels are not interchangeable. Machine validation checks a defined contract; agent review records an AI-assisted assessment; human approval records a person's approval. We do not claim that every accepted occurrence has been individually examined by a human expert, or that independent agent review is external scientific peer review.

For official jurisdiction determinations, implemented gates include human approval and specific independently reviewed agent methods. The latter require distinct reviewer identities, a retained independent review, an explicit declaration, separate corroboration, and date and conflict checks. Under the versioned agent policy, project eligibility expires 365 days after the qualifying authority date. That is a review rule, not the biological date of eradication or a guarantee lasting a year. Read the implemented review and expiry rules.

Corrections are recorded as new review, retraction, or superseding events. The evidence ledger is append-only: a corrected conclusion should remain traceable to what changed and why. The compiler blocks unresolved present-versus-absence conflicts instead of quietly selecting a convenient answer.

Turn reviewed evidence into a public view

A compiler reads versioned inputs offline and resolves them for an explicit assessment date. It produces the summaries and per-state or per-county files used by the site. Stable ordering and repeatability checks help ensure that the same inputs and date produce the same output. The local query database is a disposable research aid, not the underlying evidence ledger.

Research acquisition does not run when you open a page. The website reads prepared data. In the hosted delivery path, versioned research releases are served from our data origin; build-time checks compare release metadata and file fingerprints with the application's expectations. This helps prevent a new interface from silently combining incompatible data versions. See the compiler and release-coherence check.

Migration is not the same as completion. At this edition's code baseline, Alabama is configured to publish compiler-owned findings into the compatibility map, while the other configured jurisdictions expose research projections separately. Legacy combined records still have attribution limits. Always read the particular view's status and references; the national map and the research ledger should not be assumed to have identical review coverage.

The registry assigns source-specific refresh targets, ranging from days to months, and some manual or legacy sources have no fixed cadence. These are planning policies, not promises that every source is fetched on schedule. Provider access, source publication cycles, review work, and operational capacity affect when a new release can be published.

Observation freshness answers a different question. In the current freshness implementation, the most recent dated accepted observation is classified relative to the assessment date: up to 365 days is current, 366 through 730 days is aging, and older evidence is stale. Without a usable observation date it is undated. Partial dates use conservative lower bounds. These labels describe age, not certainty that an organism is present today.

Source age matters too. An archive downloaded this morning may contain years-old data. The retained Alien Forest Pest Explorer snapshot, for example, carries an April 2023 underlying-data horizon. The iDigBio registry entry records a frozen historical index. Rechecking either source must not silently renew the biological age of its records.

Operational receipts and validation results help track failed access, rejected records, unresolved questions, and publication readiness. This is a versioned research workflow, not a real-time warning service. Consult the relevant local authority when a decision depends on current conditions.

Three examples of careful interpretation

An old specimen is newly digitized

A herbarium publishes a specimen collected decades ago. If its identity and county are supported, it can add historical evidence. Its recent digitization does not prove a new arrival, an expanding population, or current presence.

A water sample has no detection

A monitoring program explicitly reports no target eDNA in a sample. If the source contract supports that interpretation, it can contribute a scoped non-detection. It does not establish that every water body in the county is free of the species.

An agency declares eradication after earlier detections

Historical occurrences remain in the record. A qualifying later declaration can support a separate current determination within its reviewed geography and dates. An unresolved contradictory record can block that determination; removing the old history is not the solution.

What remains uncertain

Public evidence is uneven. Accessible roads, active collecting institutions, popular observation sites, and well-funded monitoring programs can all concentrate records. County shading is therefore a view of documented evidence, not a census of organisms or a measure of invasion severity.

Source errors, outdated names, imprecise geography, duplicated specimens, delayed updates, and incomplete provenance remain possible. Some legacy combined county records cannot be traced back to a single original source-to-county connection. The EDDMapS compatibility merge is additive and cannot safely subtract an older row without rebuilding the relevant source family.

Species-profile prose is an explanatory layer, separate from the occurrence ledger. A useful identification or management reference does not prove a county occurrence, and a profile correction does not automatically change geographic evidence. Our review workspace supports revision; publication should preserve the distinction between a draft, an accepted explanation, and a supported finding.

To inspect the work, start with a species profile, follow its references, and use Research for evidence and status. If a record seems wrong, send the page, species, place, source identifier, and your supporting reference through the available correction route or contact us. A well-documented correction helps improve the record without replacing one unsupported claim with another.