Evidence Chain¶
Evidence in Pollenomics is a linked set of claims, not a single confidence label. A public row can involve source identity, record identity, place, time, coordinates, scientific role, and publication eligibility. Each claim keeps its own provenance and precision.
This model serves two kinds of reader at once. A scientific reader can see what a plotted or summarized record means. An auditor can follow the same row back through its governing surface, source locator, transformation, caveat, and admission decision.
Evidence decision records explains how the claim, proposed use, evidence basis, rule, outcome, and recovery condition stay linked when evidence becomes a product member or a reasoned non-member.
Trust Model¶
Four invariants govern the evidence chain:
- Authority is scoped. Project records govern project facts; sample records govern sample facts; aggregates summarize but do not replace them.
- Transformation cannot strengthen evidence. Normalization may make a supported claim comparable, but cannot invent finer place, time, or identity resolution.
- Joins require evidence. Shared labels, proximity, and convenient project context are not sufficient linkage on their own.
- Refusal is data. Unresolved, conflicted, blocked, and deferred outcomes remain visible so published coverage is not mistaken for source coverage.
Evidence Has More Than Two Outcomes¶
The evidence database does not force every question into supported versus unsupported. Those labels hide whether evidence was examined, whether sources disagree, and whether the requested use exceeds the available precision.
| Outcome | Meaning | Safe public treatment |
|---|---|---|
| supported | governing evidence satisfies the declared claim and use | publish at the supported scope and precision |
| qualified | the claim is usable only with a material bound or caveat | publish the qualification with the member |
| conflicted | two or more attributable claims remain incompatible | preserve each claim and withhold an unqualified conclusion |
| unknown | the required evidence is absent or has not been recovered | state the missing dimension; do not convert it to absence |
| refused | evidence was evaluated and fails the named use contract | retain the record and reason outside that product membership |
| out of scope | the claim was not evaluated because the product does not ask it | do not describe the outcome as scientific rejection |
flowchart LR
Claim["claim + intended use"] --> Evidence{"governing evidence available?"}
Evidence -->|no| Unknown["unknown or recovery state"]
Evidence -->|yes| Agreement{"consistent and sufficient?"}
Agreement -->|yes| Supported["supported or qualified"]
Agreement -->|conflict| Conflict["conflicted"]
Agreement -->|insufficient| Refused["refused for this use"]This vocabulary protects negative claims. “No published point” can mean unknown coordinates, a spatial refusal, exclusion from a geographic scope, or a reader filter; none of those proves that the specimen or phenomenon was absent.
Cross-Domain Evidence¶
All source families preserve origin, version, and their materially present lifecycle posture. Normalized and review artifacts are family-specific rather than assumed from publication. Their scientific evidence units also remain different:
- pollen sources govern sites, sequences, samples, and modelled context;
- archaeology sources govern sites and contextual records;
- SVAR governs hydrographic registry records;
- AADR governs release-versioned human ancient-DNA metadata;
- animal aDNA governs project-, paper-, supplement-, sample-, and site-owned evidence; and
- boundaries govern geographic selection, never scientific support.
Temporal semantics explains how time claims from these families can be compared without pretending they have equal resolution.
Animal Sample Evidence¶
Animal aDNA has the deepest explicit chain because a project accession or paper citation is not enough to justify a sample-level map point.
flowchart LR
Source["paper, project, supplement"] --> Sample["stable sample identity"]
Sample --> Site["sample-to-site linkage"]
Site --> Locality["locality class and provenance"]
Sample --> Chronology["date claim and provenance"]
Locality --> Coordinates["coordinate basis and precision"]
Chronology --> Fitness{"scientific fitness"}
Coordinates --> Fitness
Fitness -->|admit or qualify| Point["atlas or country evidence row"]
Fitness -->|block or defer| Ledger["exclusion or recovery surface"]Every arrow represents a claim that can fail independently. A stable sample identifier does not prove a site. A named site does not prove coordinates. A date attached to a project does not automatically belong to every sample.
The final decision is product-specific. A row may be valid evidence for a regional count, qualified contextual layer, or curation inventory while still being ineligible for an exact point or time-aware comparison.
Animal Evidence Has Several Governed Populations¶
Three prominent animal counts describe different database contracts:
| Population | Rows | Governing question |
|---|---|---|
| sample-foundation truth | 894 | which curated source rows are fully grounded, partly grounded, or blocked by metadata, locality, or chronology? |
| recovered project sample master | 868 | which source rows resolve to stable project-owned sample identities? |
| point publication | 234 | which sample-backed or explicitly qualified context members satisfy the spatial product contract? |
The foundation contains 502 fully grounded rows, 256 partially grounded rows, 29 blocked for missing metadata, four blocked for missing location detail, and 103 blocked for weak chronology. Those classes measure evidence preparation; they are not publication decisions. Likewise, the 234-member point surface contains 233 final sample-backed features and one provisional project-context feature, so it is not a subset count that can be divided mechanically by 868 or 894.
flowchart LR
Foundation["894 foundation rows<br/>grounding and blockers"] --> Identity["868 recovered sample identities"]
Identity --> Samples["233 sample-backed points"]
Context["1 qualified project-context feature"] --> Product["234 publication members"]
Samples --> ProductThe arrows express governed relations, not guaranteed one-to-one attrition. Every count or percentage must name its observation unit and population.
Species Views Preserve Taxon Posture, Not Fact Ownership¶
The directories under data/adna/species/ group evidence by accepted taxon
and expose current curation, recovery, spatial, and release posture. They are
projections over source- and project-owned evidence, not independent source
databases. Human AADR currently exposes a capture-only species route, while
the ten non-human views materialize broader normalized and review surfaces;
the common directory shape must not be read as equivalent lifecycle depth.
Species evidence views defines how to interpret those differences, audit one taxon claim, and construct a cross-species comparison without mixing projects, samples, sites, or publication members.
Evidence Dimensions¶
| Dimension | Governing question | Failure if flattened |
|---|---|---|
| Identity | Which physical or analytical sample does this row represent? | duplicate or conflated samples |
| Lineage | Which project, paper, supplement, table, and source row support it? | unverifiable extraction |
| Locality | Is the place sample-specific, site-specific, regional, substituted, or unresolved? | false geographic precision |
| Chronology | Is the date direct, derived, interval-based, textual, or unresolved? | false temporal precision |
| Coordinates | What created the point and with what confidence? | map marker outranks place evidence |
| Fitness | Is the combined record eligible for the declared publication? | presentation silently strengthens evidence |
Minimum Evidence Depends On The Claim¶
There is no universal “complete record.” Completeness is evaluated against the claim being made:
| Proposed use | Minimum governing evidence | A valid record can still be ineligible when… |
|---|---|---|
| source inventory | stable source-native identity and capture lineage | the source row cannot be distinguished or recovered |
| sample inventory | resolved sample identity and project lineage | only a project accession or unlinked paper label is known |
| named-site summary | sample-to-site evidence and locality class | place exists only at project or regional scope |
| exact point map | sample-owned locality plus source-backed or verified site coordinates | a coordinate is inferred, substituted, or broader than the locality claim |
| numeric temporal comparison | comparable numeric interval, dating basis, precision posture, and overlap rule | time is textual, contextual, unresolved, or measured under an incompatible contract |
| cross-domain association | eligible records from both domains plus explicit spatial and temporal bridges | proximity is available but chronology or evidence role is not comparable |
The required dimensions are conjunctive. Exact coordinates do not compensate for unresolved sample identity; a direct date does not compensate for a project-only locality; strong evidence in one domain does not upgrade a contextual source in another.
Completeness Is A Claim Vector¶
For any governed object, completeness is a vector of independently reviewable states rather than a percentage attached to the row:
flowchart LR
Object["governed object"] --> Identity["identity state"]
Object --> Place["locality state"]
Object --> Time["chronology state"]
Object --> Geometry["coordinate state"]
Object --> Lineage["lineage state"]
Identity --> Use{"requested use"}
Place --> Use
Time --> Use
Geometry --> Use
Lineage --> UseThe requested use selects which vector components are mandatory and which may remain qualified. This preserves useful partial evidence without letting a high-confidence component conceal an unresolved one.
Evidence Capability Is A Query¶
Evidence capability is evaluated for a particular object, claim, and use. It is not inherited from the source family or summarized by the number of filled columns.
flowchart LR
Object["governed object"] --> Claim["claim dimension"]
Claim --> Owner["fact owner + evidence locator"]
Owner --> Posture["precision and review posture"]
Use["requested scientific use"] --> Gate{"capability query"}
Posture --> Gate
Gate -->|supported| Result["bounded claim"]
Gate -->|insufficient| Refusal["qualified, contextual, or refused result"]This means “has coordinates” is not a sufficient query. The database asks whether the governed subject owns the locality, how the pair was produced, what precision it supports, and whether that posture satisfies the requested product. Equivalent queries apply to identity, chronology, taxonomy, and cross-domain association.
Claim Envelope¶
A reusable claim must retain enough context to survive outside the page where it was first seen:
| Envelope field | Why it is indispensable |
|---|---|
| governed object ID | identifies the sample, site, source record, or product member without relying on a label |
| fact owner | identifies the record authorized to define the disputed value |
| source family and locator | leads to the captured upstream object and exact supporting location |
| reported value | preserves what the source expressed before repository interpretation |
| normalized value and method | makes comparison possible without hiding transformation |
| precision and evidence class | bounds spatial, temporal, taxonomic, or identity strength |
| role and product scope | explains what the record contributes and where it was admitted |
| qualification or refusal | prevents missing or conflicted evidence from disappearing in reuse |
The envelope is intentionally larger than a popup or CSV cell. Compact views may point to it, but downstream reuse that drops these fields cannot retain the same evidential claim.
Claims Are Addressable Database Objects¶
A claim is identified by its governed subject, fact type, source or decision identity, and revision—not by the current displayed value. This allows two supported values to remain in conflict without overwriting one another and allows a later decision to change posture without erasing the evidence that was reviewed.
| Claim member | Database responsibility |
|---|---|
| subject identity | binds the claim to one typed sample, site, source member, or product |
| fact type | distinguishes identity, locality, chronology, coordinate, taxon, role, and membership claims |
| assertion identity | keeps several source statements or curated interpretations separately addressable |
| evidence locator | recovers the captured statement or deterministic inputs |
| normalized representation | enables comparison while retaining source wording and method |
| decision state | records acceptance, qualification, conflict, refusal, or unresolved posture |
| supersession relation | explains which later claim or decision replaced an earlier interpretation and why |
Supersession is not deletion. Historical values remain attributable to the database revision and evidence that supported them; current products select only the posture accepted by their own contracts.
A Decision Is Scoped To One Claim¶
Evidence posture belongs to the pair of claim and intended use, not to the record as a whole:
| Governed object | Claim under review | Possible decision without changing object identity |
|---|---|---|
| animal sample | labels identify one recovered analytical unit | final identity, ambiguity, merge, split, or refusal |
| animal sample | locality is sample-owned at named-site precision | accepted named site, qualified regional claim, substitution, or unresolved |
| animal sample | chronology supports numeric comparison | comparable interval, text-only time, contextual range, conflict, or unknown |
| locality | geometry represents the supported place precision | exact, approximate, substituted, region-only, withheld, or refused point |
| evidence row | claim is fit for one publication | admitted, qualified, excluded, deferred, or out of scope |
flowchart LR
Object["stable governed object"] --> IdentityClaim["identity claim"]
Object --> PlaceClaim["place claim"]
Object --> TimeClaim["time claim"]
IdentityClaim --> IdentityDecision["identity posture"]
PlaceClaim --> PlaceDecision["spatial posture"]
TimeClaim --> TimeDecision["temporal posture"]
IdentityDecision --> Admission["product-specific admission"]
PlaceDecision --> Admission
TimeDecision --> AdmissionThis model supports partial but honest records. A sample does not become globally “low quality” because one chronology is unresolved, and it does not become universally publishable because its identity and coordinates are strong.
flowchart LR
Claim["claim value"] --> Object["governed object"]
Claim --> Owner["fact owner"]
Claim --> Source["source and locator"]
Claim --> Method["normalization or curation method"]
Claim --> Precision["precision and class"]
Claim --> Scope["role and product scope"]
Claim --> Limit["qualification or refusal"]Evidence Joins Are Claims¶
Joining records is not a neutral formatting operation. Each relationship needs an identity rule and provenance because a wrong join can create a plausible but unsupported public point.
| Relationship | Required support | Unsafe shortcut |
|---|---|---|
| project to paper | registry linkage or source-backed publication association | matching by title fragment alone |
| paper to supplement | captured artifact identity and supporting-material manifest | assuming every attachment contains sample rows |
| project to sample | recoverable source label and stable repository identity | treating a project accession as one sample |
| sample to site | sample-owned row, defined group, or explicitly broader locality class | assigning all project samples to the project title's place |
| sample to chronology | sample-owned claim or visibly contextual fallback | copying the project age range to every sample |
| locality to coordinate | declared coordinate source, method, and precision | geocoding a broad region as an exact site |
| evidence row to product | stable identifier and successful product admission | plotting every normalized row |
Conflict and substitution ledgers preserve cases where more than one join is possible or where a broader relationship is used provisionally. That record is part of the evidence, not an implementation detail.
Evidence Outcomes¶
- direct evidence resolves to a sample-owned source location such as a supplementary table row;
- derived evidence records the transformation and its assumptions;
- qualified evidence is usable only with an explicit precision or source caveat;
- conflicted evidence preserves incompatible claims pending resolution;
- blocked evidence fails a known publication requirement; and
- deferred evidence awaits source recovery or manual curation.
Blocked and deferred states remain part of the database. Their presence makes coverage gaps and recovery work measurable.
An evidence-chain summary should therefore report at least three quantities: the known candidate population, the population for which the required chain was evaluated, and the population that passed the declared use. Omitting the first hides discovery coverage; omitting the second hides curation coverage; omitting the third hides publication selectivity.
Evidence strength is bounded by the weakest claim needed for the proposed use. Strong identity does not repair unresolved locality; exact coordinates do not repair uncertain ownership; complete lineage does not create chronology; and admission to one product does not establish fitness for another.
Inspect A Claim¶
- Begin with the public evidence identifier and publication posture.
- Resolve the normalized record and governing fact owner.
- Inspect sample identity and source lineage.
- Inspect locality and chronology as separate claims.
- Compare coordinate precision with locality evidence.
- Read conflicts, caveats, exclusions, and release-gate outcomes.
flowchart TD
Public["public row or visual mark"] --> Posture["publication posture"]
Posture --> Normalized["normalized evidence record"]
Normalized --> Owner["governing fact owner"]
Owner --> Locator["source artifact and locator"]
Posture --> Decision["admission, qualification, or exclusion"]
Decision --> Review["review and conflict surfaces"]
Locator --> Source["archive, paper, supplement, or governed dataset"]Start from the public artifact when checking a visible claim. Start from the governing fact owner when checking collection completeness or curation state. Those directions meet at the normalized evidence record, but they answer different questions.
The relevant references are sample records, localities, chronology, and coordinates. The object and relation model defines the typed identities used throughout the chain.