Skip to content

Verification Evidence

Every published claim sits on several independently reviewable proof layers. Verification establishes whether encoded contracts hold across those layers; it does not enlarge the underlying evidence or erase uncertainty recorded by a source.

flowchart LR
    Source["source identity and retrieval context"] --> Curated["curated evidence records"]
    Curated --> Relation["explicit evidence relations"]
    Relation --> Membership["governed publication membership"]
    Membership --> Product["map, report, or export"]
    Source -. "lineage proof" .-> Claim["proportionate public claim"]
    Curated -. "semantic proof" .-> Claim
    Relation -. "support proof" .-> Claim
    Membership -. "scope proof" .-> Claim
    Product -. "delivery proof" .-> Claim

Proof Layers

Layer What is verified What remains a review judgment
source integrity known source identity, retrieval context, expected assets, and governed hashes where available whether the source is suitable for a new scientific question
curation semantics required fields, preserved nulls, normalized identifiers, chronology meaning, and precision rules whether an ambiguous record should support a stronger interpretation
evidence relations typed links resolve to governed records and retain their declared role whether contextual evidence can be promoted to direct support
publication scope deterministic membership, exclusions, traceability, and declared geographic or thematic boundaries whether the publication scope is sufficient for a reader's intended inference
runtime delivery installed commands, artifacts, and failure behavior conform to their public contracts whether a technically valid output communicates uncertainty well
scientific review source-specific assumptions and visible caveats receive human review whether later evidence warrants a revised interpretation

No single layer substitutes for another. A valid artifact with missing lineage is not fully traceable; a traceable record outside a publication contract is not silently included; and a passing runtime check does not certify scientific completeness.

Reading A Verification Result

Result Supported conclusion Unsupported conclusion
source checks pass governed inputs have the expected identity and structure the source has complete spatial, temporal, or taxonomic coverage
curation checks pass encoded transformations preserve declared invariants every ambiguity has one scientifically correct resolution
relation checks pass evidence links are structurally valid and typed every linked record provides direct support
publication checks pass the product contains the membership its contract declares omitted records are irrelevant to every possible analysis
delivery checks pass the supported interface produces conforming outputs the output is appropriate for an undeclared use case

The decisive question is therefore not simply whether verification passed. It is whether the proof that passed governs the claim being made.

Match Proof To Failure Mode

Different checks catch different classes of error. A credible review names the failure mode and selects evidence capable of detecting it:

Possible failure Decisive proof Insufficient proof by itself
wrong upstream release source identity, retrieval metadata, and governed hashes unchanged published count
record meaning changed during normalization field-level semantic comparison and focused normalization behavior schema validity
evidence attached to the wrong sample or site stable identifiers and relation resolution across the chain valid individual records
eligible member missing from a product admission inventory, exclusion ledger, and manifest membership comparison successful rendering
product scope changed declared geography and product contract diff unchanged file names
caveat disappeared from presentation structured warning plus rendered-output review structured artifact validity

This is why a broad green test run is not a complete evidence statement. Its value depends on whether the relevant assertion was present and whether the reviewed artifact is the one the assertion governed.

Worked Verification Paths

A mapped SEAD site

A reader can begin with one feature in the Nordic bundle and verify, in order:

  1. the feature is named by the regional manifest;
  2. its stable site identity resolves to the normalized SEAD record;
  3. its country assignment explains regional admission;
  4. the source page preserves the SEAD identity and current source limitations;
  5. the layer role remains environmental-archaeology context.

That chain verifies membership, geography, identity, and role. It does not establish a dated event because the current SEAD capture lacks dating, period, and bibliography rows.

An animal atlas point

An animal point has a longer path: product membership, point traceability, admission decision, sample master, sample-to-site relation, locality evidence, chronology evidence, coordinate provenance, paper and supplement identity, and archive project. Verification must follow the links used by the visible claim. A valid project link cannot substitute for a missing sample locality; a valid site cannot substitute for sample chronology.

A source-family count

The collection summary can verify that a family was captured and record its count. A publication manifest can verify how many members entered one product. Those counts use different populations. Comparing them is meaningful only after the record unit, geography, and admission rule are named.

Verification Evidence Hierarchy

Public statement Minimum evidence that should be recoverable
“1,231 AADR samples are in the Nordic bundle” bundle manifest, sample identities, four-country selection, and v66 release lineage
“200 Neotoma sites provide pollen context” normalized site identities, source linkage, temporal review, and primary-context role
“23 SEAD rows are not mapped” reviewed inventory denominator and explicit missing-country state
“two animal localities are published in the Nordic product” two admitted feature identities plus sample, locality, chronology, coordinate, and source links
“one fieldwork location is shown” observation record, site identity, coordinates, date, contributor, and sampling-context role

The hierarchy is claim-specific. It does not award one repository-wide confidence score or treat a longer provenance path as inherently stronger.

Failure And Refusal

Verification failures preserve information. Depending on the boundary, the system can reject malformed input, preserve a qualified record, exclude a record with an explicit reason, or refuse publication. These outcomes are preferable to manufacturing certainty or silently weakening a contract.

An exclusion report is consequently part of the evidence, not a secondary debug artifact. It identifies the boundary between what a publication can support and what remains outside its declared claim.

Absence also needs a type. “Not captured,” “captured but unresolved,” “resolved but ineligible,” and “eligible but outside this product scope” describe different populations and different recovery actions. Collapsing all four into a missing row destroys the information needed to audit completeness.

Review Surfaces