Evidence Decision Records

An evidence decision records what one governed claim supports for one declared use. It connects source evidence to an outcome without overwriting the source, discarding an unresolved object, or granting the same fitness to every downstream product.

The domain language defines claim posture, product admission, qualification, exclusion, and recovery condition.

Evidence, Product, And Release Decisions Differ

Decision class	Governing question	Typical outcomes
evidence decision	What does this source-supported claim establish, at what scope and precision?	accepted, qualified, conflicted, unresolved, or refused
product decision	Does that evidence belong in this named publication and role?	admitted, excluded, deferred, outside scope, or qualified member
release decision	Do the complete product and its dependencies support the proposed release language?	allowed, blocked, or refused with dimensions

flowchart LR
    Source["captured evidence"] --> Claim["typed claim"]
    Claim --> Evidence{"evidence decision"}
    Evidence --> Product{"product decision"}
    Product --> Member["member or accounted non-member"]
    Member --> Release{"release decision"}

These decisions can legitimately differ. A location claim can be strong enough for a qualified context marker while the object remains ineligible for a recovered-sample count. A retained publication can remain inspectable while the current database snapshot refuses stronger release language because a contracted lifecycle stage is absent.

Anatomy Of A Decision Record

A reusable decision keeps the following members together:

Member	Responsibility
subject	stable typed object or claim identity under review
proposed use	exact comparison, publication, geography, role, or release claim being evaluated
evidence basis	source identities, locators, normalized facts, and relations considered
rule identity	contract or predicate applied to the evidence
outcome	admission, qualification, conflict, exclusion, deferral, refusal, or scope result
reason	evidence-grounded explanation for the outcome
precision and limits	spatial, temporal, identity, taxonomic, or role ceiling that survives reuse
recovery condition	named material or relation that could change an incomplete outcome
revision	database and product state in which the decision was evaluated
descendants	manifests, tables, maps, reviews, or release language governed by the decision

An outcome without its subject and rule is only a label. A reason without an evidence locator cannot be independently checked. A public member without the decision that admitted it cannot explain why similar candidates are absent.

Decision Identity And Supersession

A decision is keyed by subject, claim or proposed use, scope, rule identity, and evidence revision. Decisions with different keys may coexist; a newer row does not supersede an older row merely because it was written later.

Relationship between decisions	Required treatment
same subject, different claim dimension	retain both; locality, chronology, identity, and role are independent
same evidence, different product or geography	evaluate and retain separate admission decisions
broad project context and narrow sample claim	retain both at natural scope; the narrow claim governs only its sample
same key, stronger evidence, new revision	record explicit supersession, reason, prior outcome, and descendant impact
same key, conflicting evidence, no justified owner	preserve both claims and move the decision to conflicted or unresolved
evidence decision and release decision	keep separate; repository release posture cannot rewrite the underlying claim

Supersession changes which decision governs a named use; it does not delete the prior evidence or outcome. This makes a later admission explainable without making the earlier exclusion look like an undocumented error.

A Decision Can Become Stale Without Becoming False

When an input identity changes, the prior decision remains a valid historical record for its revision but no longer governs the new state. Invalidation and supersession therefore answer different questions:

Event	Prior decision	Required next state
evidence or relation changes	retained, marked outside the current read set	reevaluate the same proposed use
product rule or geography changes	retained for the prior product identity	create a decision under the new rule and scope
stronger evidence confirms the same outcome	retained and superseded with an explained no-membership-change result	propagate the new evidence revision
changed evidence reverses the outcome	retained with affected descendants	superseding decision plus member-level reconciliation

This distinction prevents a stale admission from silently surviving a changed coordinate or rule, while preserving the evidence needed to reproduce an older publication.

One Object Can Carry Several Decisions

The Wadi Halfa dromedary context illustrates why decisions remain scoped:

Claim or use	Governed posture
project identity	archive project SRP073444 is tracked
source context	paper-backed Site 1040 wording supplies a recoverable named-place claim
sample identity	a final sample-master identity is not yet recoverable
coordinate claim	named-place geocoding supports an approximate point, not a supplied excavation coordinate
point product	admitted as a visibly qualified project-context feature
recovered-sample population	not admitted as a final recovered sample
release posture	incomplete recovery remains part of the repository release refusal

flowchart TB
    Project["SRP073444 project"] --> Context["paper-backed place context"]
    Context --> Approximate["approximate named-place coordinate"]
    Approximate --> ContextMember["qualified context member"]
    Project --> Missing["final sample evidence missing"]
    Missing --> Recovery["recovery condition"]
    Recovery --> Release["stronger release language refused"]

The context member and recovery failure are not contradictory. They answer different questions and remain linked so readers do not count the marker as a fully recovered sample.

Decision Surfaces In The Checked-In Database

Surface	Decision represented
data/adna/final/atlas/animal_atlas_candidate_accountability.json	whether each animal atlas candidate has the required sample, site, chronology, coordinate, and lineage companions
data/adna/governance/source_library/sample_locality_conflict_ledger.json	locality disagreement and ownership posture
data/adna/governance/source_library/sample_chronology_conflict_ledger.json	chronology normalization, precision, and conflict posture
data/adna/governance/source_library/source_recovery_release_guard.json	projects whose implausibly low recovery blocks stronger release language
docs/report/animal_atlas_exclusion_report.json	known evidence rows that do not satisfy the animal point product
docs/report/repository_final_release_refusal.json	repository-wide dimensions that refuse final-release language

These surfaces are projections of related decisions rather than one universal quality table. Their denominators differ: atlas candidates, chronology rows, tracked projects, excluded evidence rows, and repository release dimensions are distinct observation units.

Read Candidate Accountability Correctly

The current animal accountability surface contains 234 candidates and 233 rows that pass its complete sample-accountability test. The remaining Wadi Halfa context feature is intentionally represented under a different point class. Its presence does not turn the failed sample-accountability predicate into a pass.

Question	Correct authority
Is the candidate fully sample-accountable?	candidate accountability row
May a qualified project-context marker appear?	point publication contract and traceability row
May it enter a recovered-sample count?	sample-foundation and sample-master population contract
Does the overall repository permit final-release language?	repository release-refusal surface

One summary boolean cannot answer all four questions. Preserve the decision class and observation unit when citing a total.

Audit A Decision

1. Identify the stable subject and the exact claim or use being evaluated.
2. Recover every evidence locator named by the decision.
3. Confirm which record owns each repeated fact.
4. Read the rule as a conjunction of required dimensions; do not let strength in one dimension compensate for missing identity, locality, chronology, or lineage unless the contract explicitly allows it.
5. Compare the outcome with the product member or non-member surface.
6. Verify that the qualification, conflict, or recovery condition remains reachable from the public result.
7. Check that later geographic products preserve the same meaning.

flowchart LR
    Subject["subject + proposed use"] --> Evidence["evidence packet"]
    Evidence --> Rule["named rule"]
    Rule --> Outcome["scoped outcome"]
    Outcome --> Product["member or non-member"]
    Outcome --> Recovery["qualification or recovery"]

The audit fails if the public wording is stronger than the decision, if a qualified member loses its qualification, or if an exclusion cannot be connected to a known candidate and failed predicate.

Decision Changes And Supersession

Stronger evidence can change a decision without erasing the earlier state. The revised record should preserve:

• the prior subject, evidence, rule, and outcome;
• the new evidence locator or corrected fact owner;
• the reason the earlier posture no longer governs;
• affected product memberships and aggregate counts; and
• the revision in which each outcome applied.

This makes decision history interpretable as evidence improved. Replacing a reason with a new label or editing only the visible map would destroy that history and leave dependent products without a causal explanation.

Reuse Contract

When carrying a decision outside its original bundle, retain the stable subject, proposed use, evidence locators, rule identity, outcome, reason, precision, recovery condition, revision, and product scope. If those members cannot travel, describe the result as an orientation aid rather than a reproducible evidence decision.

Continue with record admission, conflicts and recovery, querying governed evidence, and the publication model.