Evidence Curation

Curation is the accountable work between acquiring a source and publishing a claim. It determines which object a source row describes, which facts the row supports, how precise those facts are, and whether they are fit for one declared product. Normalization makes records structurally consistent; curation makes their scientific meaning and limits explicit.

The domain language defines the claim, fact owner, posture, admission, qualification, and recovery terms used here.

The Curation Boundary

flowchart LR
    Capture["captured source material"] --> Object["resolved object identity"]
    Object --> Claims["identity, place, time, and role claims"]
    Claims --> Evidence["source locators and provenance"]
    Evidence --> Review{"claim-specific review"}
    Review -->|supported| Admit["admit at declared strength"]
    Review -->|limited| Qualify["retain with qualification"]
    Review -->|conflicted| Conflict["preserve competing claims"]
    Review -->|insufficient| Refuse["exclude or defer with reason"]

The review is claim-specific. One sample can have a final identity, a directly supported site, an approximate coordinate, and unresolved numeric chronology. A single record-level quality flag would hide those independent outcomes.

What Is Curated

Curation object	Governing question	Durable result
source	Which release, archive, paper, supplement, or service supplied the material?	source identity, retrieval context, content identity, and use boundary
object identity	Which project, paper, sample, site, sequence, lake, or registry record is this?	stable identity, native key, aliases, and ambiguity state
claim	What does the captured material say about taxonomy, locality, chronology, coordinates, or role?	source wording, subject, value, scope, precision, and locator
relation	How does one governed object relate to another?	typed sample-to-project, sample-to-site, paper-to-project, or product-to-member link
decision	Is the claim usable for the proposed comparison or publication?	admitted, qualified, conflicted, deferred, or refused posture with reason
recovery item	What evidence would change an unresolved decision?	named missing artifact, field, relationship, or review action

These objects live at their natural scope. Project registries govern project identity; project sample masters govern samples; per-sample evidence packets govern locality and chronology claims; species views combine reviewed records; publication manifests govern product membership.

Anatomy Of A Curated Claim

A curated claim is durable when another reader can reconstruct both the reported fact and the decision made about it:

Field	Purpose
subject	typed governed object the claim concerns
dimension and predicate	identity, taxonomy, locality, chronology, coordinate, role, or another declared assertion
reported value	source wording or source-native value before interpretation
normalized value	typed representation when a defensible transformation exists
scope and precision	object level, geography, time basis, uncertainty, and supported granularity
evidence locator	source identity, artifact, table, row, field, excerpt, or deterministic derivation
fact owner	authority that governs repeated copies of this dimension
posture	accepted, qualified, conflicted, unresolved, deferred, or refused
decision reason	evidence and rule that justify the posture
recovery condition	specific evidence capable of changing an incomplete decision
revision identity	database state in which the claim and decision were evaluated

The normalized value may legitimately be absent while the reported value and locator remain governed. Preserving that distinction is stronger than filling a typed field with an assumption that cannot be traced to the source.

Curation Is Database Construction

Curation does more than annotate imported rows. It constructs the joins and negative states that make the evidence database queryable without erasing the source model.

Construction decision	What must remain recoverable
resolve two labels to one sample	both source labels, their locators, the governed sample key, and the resolution basis
split one source row into several claims	the shared source locator and the owner, scope, and precision of each claim
connect a sample to a site	the typed relation, supporting evidence, and whether the relation is direct or curated
decline to create a coordinate	the locality evidence examined and the reason precision would be fabricated
admit a record to a product	the exact evidence state and rule set under which membership was granted
retain a known non-member	candidate identity, failed or inapplicable rule, and product scope

A database that retains only successful joins cannot explain its own denominator. The unresolved, refused, deferred, and outside-scope populations are part of the curated state, not editorial debris.

Curation Preserves Five Conservation Rules

Rule	Conserved information
source conservation	captured wording, native identifiers, locators, and nulls survive interpretation
identity conservation	merges, splits, aliases, and unresolved candidates remain traceable to prior identities
scope conservation	project, sample, site, context, and product claims stay at the level supported by evidence
population conservation	admitted members and every class of known non-member reconcile to the governed denominator
decision conservation	prior outcomes, qualifications, conflicts, and recovery conditions remain explainable after revision

Curation may add a typed interpretation or select a governing claim, but it cannot create source evidence, raise precision, erase a candidate, or transfer a fact to a narrower object without a recorded evidentiary operation. These rules make preparation scientifically reviewable rather than merely tidy.

Curation Does Not Overwrite The Source Row

The captured value, repository interpretation, and product decision remain separate records. That separation is visible in two animal cases:

Case	Captured evidence	Curated database state	Product consequence
Direkli Cave goat	supplementary workbook Table S2, row 2 names archive sample SAMEA4453841, locality, coordinates, and chronology	final project-owned sample identity with separate locality, coordinate, and chronology claims	admitted as a sample-backed point at the supported precision
Wadi Halfa dromedary	paper-backed Site 1040 place context for project SRP073444; no recoverable sample-master row	provisional project-context identity and approximate named-place coordinate, with missing sample evidence retained	admitted only as a qualified context feature, not counted as a recovered sample

flowchart LR
    Capture["captured wording and locator"] --> Claim["typed evidence claim"]
    Claim --> Decision["curation decision"]
    Decision --> Projection["product-specific projection"]
    Capture -. "preserved, never rewritten by" .-> Projection

Both cases produce useful public evidence, but they do not produce the same object or claim class. Flattening them into a common “curated point” row would erase the distinction that makes the 234-member animal surface defensible.

Fact Ownership

Facts often appear in several convenient exports. Repetition does not transfer authority. data/source_fact_ownership_registry.json names the governing surface for repeated facts and identifies downstream supporting surfaces.

flowchart TB
    Owner["governing fact owner"] --> Species["species-normalized view"]
    Owner --> Review["cross-project review"]
    Owner --> Candidate["publication candidate"]
    Candidate --> Product["world, regional, or country member"]
    Product -. "audit back to owner" .-> Owner

For example, a project-owned sample_master.json governs animal sample identity. A species record, atlas candidate, CSV export, and popup may repeat the label, but corrections begin in the project authority and propagate through every descendant.

Curation Outcomes

Outcome	Meaning	Public effect
admitted	required evidence and relations support the product claim	publish with its declared role and precision
qualified	the claim is useful but materially limited	publish only when the qualification remains visible
contextual	evidence informs interpretation without directly supporting the target object	retain as a separately labelled context layer
conflicted	two captured claims disagree and neither may be silently erased	preserve both, select an owner where justified, and expose the conflict
deferred	a named recovery action could resolve the decision	keep outside the stronger product until recovery and review complete
refused	the proposed claim exceeds the available evidence or product contract	publish the refusal or exclusion reason, not the unsupported claim

Refusal is preserved evidence about the boundary of the database. It does not mean that the source object is false or irrelevant; it means that the proposed use is not currently justified.

Decision Granularity

Decisions are attached to a claim and use, not to a source family or file. The following questions therefore receive independent answers:

• Is the sample identity resolved?
• Is its relation to a project or paper recoverable?
• Does locality evidence support a named site, a broader region, or neither?
• Is a coordinate reported, resolved from a name, representative, or derived?
• Is chronology numeric, textual, contextual, conflicted, or absent?
• May the object support the target claim, appear only as context, or remain a known non-member?

Keeping those answers separate prevents a strong source-level reputation from overriding a weak sample-level relation. It also prevents one unresolved dimension from hiding evidence that remains useful at a narrower strength.

Dimension Decisions Remain Independent

Dimension	Typical owner	Decision that must not leak into it
identity	project sample authority or source-native record	a precise coordinate cannot prove which sample it belongs to
locality relation	sample-to-site evidence	project membership cannot make a project place sample-owned
coordinate	coordinate-provenance claim tied to the governed locality	map geometry cannot upgrade an approximate or representative point
chronology	sample, site, or context chronology claim at its natural scope	locality confidence cannot create a numeric interval
taxonomy	governed taxonomic claim and its source locator	species-folder membership cannot resolve conflicting source taxonomy
evidence role	source-family and product contract	co-location cannot turn context or framing into direct evidence
product membership	product-specific admission decision	admission to one product cannot authorize another role or geography

A curation review may assemble these dimensions, but it does not trade strength between them unless a named product contract explicitly permits a qualified alternative. The resulting record is a vector of scoped decisions rather than one confidence score.

Audit One Curated Claim

Start from either the source or the public member and recover the same chain:

1. identify the governed object and its source-native key;
2. locate the captured artifact, field, table row, or excerpt;
3. identify which surface owns the disputed fact;
4. inspect normalization, precision, null, and conflict posture;
5. inspect the product-specific admission decision and scope;
6. confirm that warnings, exclusions, and recovery items remain reachable.

For the final check, compare the public wording with the decision class. A sample-backed member may name the governed sample and source-supported point; a project-context member must retain its provisional identity and approximate spatial posture everywhere it appears. If presentation uses the stronger language for both, the defect is in the projection even when the underlying curation records are correct.

Continue with evidence decision records for the durable connection between claim, rule, outcome, and product; record admission for the product gate; and conflicts and recovery for unresolved evidence. Use source families for acquisition context, evidence dimensions for claim semantics, and publications for manifested outputs.