Table of contents
ADR 0038: Owner-Local Class Instance Governance
Status: accepted Date: 2026-05-29
Context
ADR 0020 establishes class-instance occurrences as the generic lane for durable household identities such as People, pets, and owned vehicles. ADR 0037 establishes collaborative/shared class instances as the long-term sharing target, with People collaboration as the first transitional specialization.
The current People implementation predates the generic class-instance model and therefore contains useful behavior and some semantics that should not be copied blindly:
owner_face_recognitionworker results can match faces to existingperson_recordsor create unnamed candidate people.- Legacy
person_assignmentscreated by recognition are usuallyowner_tentative. - The new generic bridge currently mirrors some People worker proposals as
class_instance_occurrence_candidates.state = acceptedandclass_instance_occurrences.state = accepted, even when the legacy assignment remains tentative. - Legacy People has owner-policy automation for unnamed-candidate merge/retarget flows and shared-person auto-contribution.
This creates a target-governance question: which legacy behaviors are architectural decisions worth generalizing, and which are transitional debt?
Decision
Memories will use the following owner-local class-instance governance rules for the generic target model.
1. Accepted Means Accepted
In the generic class-instance model, accepted means Core has accepted the occurrence as authoritative current owner-local truth by one of these paths:
- explicit owner action;
- explicit owner-enabled Core auto-accept policy;
- trusted migration/import that is recorded as such.
Worker confidence alone does not produce accepted class-instance occurrences.
The current People bridge behavior where generic rows may be marked accepted while legacy person_assignments remain owner_tentative is transitional compatibility debt. New generic implementations must not copy that semantic mismatch.
2. Worker Proposals Remain Evidence
Workers may propose:
- new instance candidates;
- existing-instance occurrence candidates;
- supporting class/subclass evidence;
- representative media/occurrence evidence;
- normalized confidence and provenance.
Workers must not decide final household-facing labels, accepted membership, merge/split state, sharing, or collaborative membership.
3. Owner-Local Auto-Accept Is Narrow
Generic class-instance auto-accept may be implemented only as Core-owned policy and must be owner-effective. Super Admin may configure technical defaults and thresholds, but is not a privacy-bypass reviewer for owner-private results.
For owner-local class-instance domains, auto-accept may accept an occurrence only when all of the following are true:
- the target instance already exists and is active;
- the target instance belongs to the same owner, domain, and class;
- the owner or sensitivity policy allows auto-accept for that domain;
- the candidate confidence satisfies the Core-owned threshold for the active manifest/model/profile semantics;
- the occurrence does not conflict with an existing active accepted/tentative assignment or explicit owner rejection;
- provenance records the policy, threshold, confidence, producer/model/profile, manifest version, and acceptance reason.
Auto-accept must not:
- create a new visible/named household identity;
- authoritatively name or alias an instance;
- merge two accepted instances;
- split an accepted instance;
- share an instance or media;
- accept collaborative contributions unless the collaborative policy explicitly allows that separate workflow.
Identity-sensitive domains such as People default to auto-accept off. If enabled, People auto-accept should be limited to high-confidence occurrences for existing active owner-local people.
4. New Instance Candidates Stay Reviewable
New instance candidates are evidence-first review items. Core may store them as pending candidates, but it should not create household-visible authoritative instances unless an owner accepts them or a trusted migration/import creates them with explicit provenance.
For People, the legacy behavior of creating unnamed person_records from dense unknown clusters remains a compatibility/UX stepping stone. The generic target should model this as pending candidate evidence until the owner or policy explicitly promotes it.
5. Merge/Split Are Owner-Governed
Owner-local merge and split are same-owner, same-domain, same-class governance operations.
Manual owner merge may mark a source instance as merged, link it to the target, move or supersede accepted occurrences with provenance, and record a review/governance decision. It must not erase history.
Manual owner split means moving selected occurrences out of one instance into a new or existing same-owner/domain/class instance with provenance. Split is not modeled as perfect time-travel undo; it is a new corrective governance action.
Workers may suggest merge/split evidence, but must not execute merges or splits.
Automatic merge is not a generic accepted-instance operation. The legacy People auto-merge of unnamed tentative candidate people into named people is allowed only as transitional owner-policy behavior over unnamed/tentative evidence. The generic equivalent should be modeled as policy-based retargeting/acceptance of tentative occurrences, not as automatic merging of two accepted identities.
6. Collaborative Class Instances Are Separate
Cross-owner and shared identity behavior is not an owner-local merge/split operation. It belongs to the collaborative class-instance model from ADR 0037:
- shares/audiences;
- instance-scoped roles;
- contributions;
- representative identity metadata;
- fork/revocation provenance;
- authorization separate from media visibility.
People-specific collaborative tables remain transitional. New non-People collaborative domains must not introduce bespoke sharing systems.
Consequences
- The generic class-instance model has clearer state semantics than the current People bridge.
- People migration still needs a corrective slice to stop treating tentative legacy recognition as generic accepted truth unless owner/policy acceptance really occurred.
- Auto-accept can remain useful without letting workers or admins silently rewrite identity truth.
- Merge/split can be implemented incrementally as owner-local operations without prematurely solving collaborative fork/revocation.
- Existing legacy automation is not invalidated, but it is explicitly classified as compatibility behavior to converge away from.
Revisit Triggers
Revisit this decision if:
- real-library People testing shows owner-local existing-person auto-accept is too noisy even at conservative thresholds;
- a non-People class-instance domain needs safe automatic new-instance creation;
- collaborative class-instance implementation requires different owner-local/collaborative merge boundaries;
- export/restore needs a richer event log than review decisions and provenance JSON provide.
Memories Wiki
Using Memories
- Accounts, the Library Hub, and Data Export
- Albums and Automatic Suggestions
- Explore and Discovery Domains
- The Media Viewer
- People and Faces
- Places and Map
- Plants (In Development)
- Datenschutzerklärung (Privacy Policy)
- Fast, Progressive Image Loading
- Search
- Sharing and Households
- Timeline and Photos
- Uploading and Import
Self-Hosting
- Backup and Restore
- CLI Password Reset
- Deployment Architecture
- Known Issues and Operator Diagnostics
- Map Tiles and Place Data Privacy
- Storage Layout
- Upgrades and Versioning
- Workers
Reference