Entity Persistence Across Observation Gaps

Keep a temporarily unseen entity represented as an uncertain continuing entity, then re-associate its return to the retained identity before declaring disappearance or creating a replacement.

Overview

A system often stops seeing an entity before the entity stops existing. A person walks behind an obstruction, a device goes offline, an asset misses a scan, a field subject is not detected, or a case falls out of contact. The naïve response is to erase the entity from the current model and create a new one if something similar later appears. That response destroys continuity. The opposite naïve response—keeping every unseen entity forever—creates ghost records, false matches, storage burden, and surveillance.

Entity Persistence Across Observation Gaps is the intervention pattern for bridging that interval responsibly. It makes four distinctions explicit: visibility is not existence; an estimate is not an observation; a persistent identifier is not proof of identity; and a missing signal becomes evidence of absence only in proportion to the observation channel’s ability to detect the entity. The system retains a bounded, uncertain continuity hypothesis, then applies a reappearance association gate before restoring the prior identity.

Structural problem

Observation channels are intermittent while entities and obligations persist. Without an explicit continuity layer, systems oscillate between two failures:

1. premature disappearance, where a missed observation deletes the entity and severs its history; and
2. indefinite persistence, where stale hypotheses remain active without uncertainty growth, review, or retirement.

Both failures become more severe when identity matters. A wrong split creates duplicate records and loses lineage. A wrong merge transfers state, permissions, liability, care history, or suspicion to the wrong entity. The core problem is therefore not merely hidden-state inference. It is the governance of identity, evidence, uncertainty, absence, and return across time.

Intervention sequence

1. Define the entity and identity test. State what counts as the same entity across time and what evidence can establish or defeat continuity.
2. Bind stable identity separately from a sighting. A sensor token, filename, location, or local database key must not silently become the entity itself.
3. Separate visibility, existence belief, and lifecycle state. Use states such as observed, unobserved, dormant, absent, retired, and ambiguous rather than one present/absent flag.
4. Anchor the model in confirmed evidence. Preserve the last observation, source, time, measurement quality, and provenance.
5. Propagate only what is justified. During a gap, estimate plausible change and widen uncertainty. A static dormant record is preferable to a precise but fictional prediction.
6. Calibrate non-detection. Ask how likely the observation system would have detected the entity if it were present under current conditions.
7. Bound persistence. Apply confidence decay, review dates, resource tiers, legal limits, and a retirement rule.
8. Gate reappearance. Compare identity features, trajectory, context, provenance, and conflicts before reconnecting the new presentation.
9. Permit abstention. Hold competing hypotheses or seek discriminating evidence when a forced match would be unsafe.
10. Preserve lineage. Rebinding restores the same history; rejection creates a new identity with predecessor or ambiguity links rather than silently overwriting the old one.
11. Reconcile observers. Keep source-specific evidence and resolve delayed, duplicate, or contradictory reports explicitly.
12. Measure errors and recalibrate. Track false merges, false splits, ghost duration, retirement reversals, unmatched returns, and privacy-retention violations.

Key components

Component	Description
Identity persistence criterion and binding ↗	The identity criterion explains why two temporally separated presentations are the same entity rather than merely similar. The persistent identity binding gives that entity a stable reference independent of location, current state, or a source-local identifier. Neither component proves a match by itself; both make the match test and its consequences explicit.
Visibility–existence separation ↗	This component prevents the central category error. An entity can be unobserved while still believed to exist, and it can be retired while a minimal tombstone remains. Visibility, existence belief, lifecycle state, and record retention should be separate fields with separate transition rules.
Persistent entity state record ↗	The record carries the identity, last confirmed observation, state belief, uncertainty, source provenance, lifecycle status, and review information. It must label which values were observed, inferred, propagated, or administratively assigned. A return should update this record, not silently replace it.
Observation-gap state model ↗	The gap model specifies what may happen while the entity is unseen. In some domains it is a stochastic motion model; in others it is only a statement that the entity may persist while detailed state is unknown. The model should become less certain as evidence ages unless new indirect evidence justifies otherwise.
Absence-evidence calibration ↗	A missed observation is informative only relative to search effort, sensor coverage, reporting behavior, and expected detectability. This component prevents a low-coverage channel from producing high-confidence disappearance claims and prevents a high-coverage negative observation from being ignored indefinitely.
Reappearance association gate ↗	The gate determines whether a new presentation is rebound to a retained identity, held as ambiguous, split from an existing identity, merged after review, or assigned a new identity. The gate must expose asymmetric harms: false merges and false splits rarely cost the same.
Retirement, reconciliation, and privacy controls ↗	A retirement rule prevents ghost accumulation; a reconciliation record preserves conflicting evidence; and privacy/data-minimization rules limit retention and re-identification. Together they ensure that object permanence remains bounded rather than becoming an excuse for permanent tracking.

Common mechanisms

A predictive state filter propagates a probability distribution or bounded envelope through a gap. A dormant entity registry retains identity and sparse evidence when detailed prediction is unwarranted. A reappearance association protocol structures matching and abstention. An absence-evidence calibration test measures detection coverage. A persistent identifier resolver links source-local references. Existing mechanisms such as identity-resolution workflows, grace periods, soft-delete quarantine windows, and tombstones can support the pattern, but none alone constitutes it.

Parameter dimensions

Dimension	Lower setting	Higher setting	Design consequence
Expected observation-gap duration	brief occlusion	long dormancy	changes propagation model, storage tier, and review cadence
Detection probability	weak coverage	near-complete coverage	determines how strongly non-detection supports absence
State dynamics	slow or bounded	fast or volatile	controls uncertainty growth and usefulness of prediction
Association threshold	permissive	conservative	trades false merges against false splits
Identity evidence richness	sparse	multi-modal	improves discrimination but can increase privacy and bias risk
Dormant-state budget	small	large	changes compression, prioritization, and retirement pressure
Retirement threshold	early	late	trades continuity against ghost-state and retention costs
Reversibility	provisional link	irreversible merge	determines evidence and review requirements
Observer topology	one source	federated sources	adds provenance, ordering, and reconciliation requirements
Human impact	low	high	increases contestability, explanation, and independent review duties

Invariants to preserve

• Not currently observed never means nonexistent by definition.
• Predicted state is always distinguishable from observed state.
• The strength of an absence claim does not exceed the demonstrated detection capability.
• Every rebind, retirement, split, merge, and new identity remains auditable.
• A continuity decision preserves history; a discontinuity decision preserves predecessor relations.
• Dormant state is bounded by review, retention, and resource rules.
• Human re-identification remains purpose-limited, lawful, contestable, and subject to deletion or contextual separation where required.

Variants

Predictive occlusion continuity

This variant actively propagates a state distribution while the entity is hidden. It is common in robotics, navigation, and computer vision. Its central risks are overconfident dynamics, track switching, and stale tracks that compete with real observations.

Dormant-record continuity

This variant retains identity and history without claiming a detailed current state. It fits offline devices, asset registries, longitudinal studies, and case systems. Its central risks are stale factual claims, indefinite retention, and privacy overreach.

Federated sighting reassociation

This variant reconciles intermittent reports from multiple observers or replicas. It requires provenance, ordering, source authority, and conflict handling. Its central risks are impossible continuity claims, suppressed local uncertainty, and mistaken global merges.

Neighbor distinctions

• State Estimation asks what hidden state is most likely now. This archetype also asks whether the entity still exists, whether a later presentation is the same entity, when the hypothesis should retire, and how lineage should be preserved.
• Explicit State Modeling defines allowable states and transitions. This archetype governs evidence through a period when the entity is outside observation.
• Observability Instrumentation improves the signal. This archetype remains necessary when the signal is still absent or intermittent.
• Equivalence Normalization and Equivalence Class Consolidation resolve multiple presentations into a common class or record. They supply matching machinery but do not own temporal persistence, uncertainty growth, or retirement.
• Continuity Preservation smooths known transitions. This archetype bridges missing evidence.
• Reconciliation After Drift repairs divergent representations. This archetype may invoke reconciliation, but its primary purpose is to carry an identity hypothesis through the gap.
• Identity-Bounded Change decides sameness through modification. This archetype decides sameness through non-observation and reappearance.
• Change-Scoped Revalidation decides what must be rechecked after change. This archetype can apply even when no change event was observed.

Failure modes and safeguards

Ghost entities

Keeping every entity indefinitely turns caution into clutter and risk. Use uncertainty decay, review dates, tiered storage, and explicit retirement.

Premature deletion

Treating a single missed observation as disappearance fragments identity. Calibrate absence against detection coverage and expected gap length.

False rebind

A returning presentation can be attached to the wrong identity, especially when features are common. Require multiple evidence dimensions, conflict checks, and an abstention path.

False split

A legitimate return can be treated as new, producing duplicate records and broken history. Preserve candidate predecessor links and allow reversible provisional associations.

Invented certainty

A state model can quietly turn predictions into facts. Label observations and estimates separately, timestamp them, and widen uncertainty as evidence ages.

Privacy and surveillance creep

Persistent identity features can outlive purpose or consent. Apply data minimization, access boundaries, retention limits, deletion, and unlinkability requirements independently of matching accuracy.

Bias and contestability failure

Re-identification errors can differ across groups or contexts. Measure both false-merge and false-split rates, provide explanations and correction paths, and avoid high-stakes automation where evidence is not adequate.

Extended example

A hospital transport robot observes two medication carts. One moves behind a construction partition. The robot does not delete it. It marks the cart unobserved, retains the persistent cart identity and last confirmed contents, predicts a widening location envelope, and decreases confidence according to time and possible exits. When a cart emerges, the association gate checks its tag, appearance, route, load signature, and whether any incompatible sighting exists elsewhere. A strong match restores the same cart history. An ambiguous match remains unresolved and causes a safe slowdown rather than transferring the wrong medication context. If the cart stays missing beyond validated search coverage, the record becomes a tombstone and a custody exception. Operators can inspect which values were observed, inferred, and retired.

Non-examples

• A state-machine diagram for a continuously stored ticket is Explicit State Modeling, not observation-gap continuity.
• A one-time customer deduplication job is Equivalence Normalization, not this archetype.
• Adding telemetry to expose a hidden temperature is Observability Instrumentation unless entity identity must persist through disappearance and return.
• Keeping deleted rows forever without uncertainty, review, privacy, or reassociation rules is uncontrolled retention, not object permanence.

Implementation checklist

• Is the entity and identity criterion explicit?
• Are visibility, existence belief, lifecycle state, and storage state separate?
• Is the last confirmed observation preserved with time, source, and uncertainty?
• Does the gap model avoid claiming more than the evidence supports?
• Has detection probability been validated under relevant conditions?
• Are false-merge and false-split harms stated separately?
• Can the association process abstain and preserve ambiguity?
• Are retirement, tombstone, and identifier-reuse rules explicit?
• Are multi-observer conflicts and ordering problems auditable?
• Are privacy, deletion, purpose limitation, and appeal built into the lifecycle?
• Are error rates and stale-entity duration monitored and recalibrated?

Disposition rationale

The target is drafted as a full archetype rather than converted to a component or mechanism. Object permanence names a reusable cross-domain intervention with a distinct problem signature, decision cycle, components, parameters, failure modes, and safety boundaries. State estimation, identity resolution, persistent identifiers, grace periods, filters, and tombstones each implement only part of the pattern. No accepted archetype or previous queue output owns the complete conjunction of visibility–existence separation, uncertain state retention, calibrated absence evidence, evidence-gated reappearance association, bounded retirement, and preserved identity lineage.

Common Mechanisms

• absence_evidence_calibration_test
• dormant_entity_registry
• Grace Period
• Identity Resolution Workflow
• multi_observer_sighting_reconciliation
• Persistent Identifier Resolver
• predictive_state_filter
• reappearance_association_protocol
• Soft-Delete Quarantine Window — Makes deletion reversible by first marking a layer deleted and holding it, recoverable, for a grace period sized to how much its loss would hurt — before anything is destroyed for real.
• Tombstone or Deletion Marker — Leaves a durable marker where a removed layer used to be — recording that it existed, that it's gone, and where its references should now resolve — so deletion can't be mistaken for 'never there.'

Compression statement

Define the entity and identity test; separate visibility, existence, and lifecycle state; retain a timestamped identity-bearing record after observation stops; propagate only what the dynamics and evidence justify while confidence decays; estimate how informative non-detection actually is; retire the entity only under an explicit absence or expiry rule; and on reappearance compare candidate evidence against the dormant record, preserving history when the association passes and recording ambiguity, split, merge, or new identity when it does not.

Canonical formula: For entity e with persistent identity p, last observation o_t, state belief B_t, and visibility v_t: if no observation arrives at t + Δ: visibility := unobserved B_(t+Δ) := propagate(B_t, dynamics, Δ) confidence := decay(confidence_t, Δ, detection_coverage, context) do not set existence := false unless absence_evidence ≥ retirement_threshold for a candidate reappearance o_j: score(e, o_j) := identity_match × state_compatibility × provenance_consistency if score ≥ association_threshold and conflicts are absent: rebind o_j to p and update B elif evidence is ambiguous: hold competing hypotheses else: create a distinct identity and preserve predecessor/uncertainty records

Related Abstractions

Abstractions this archetype builds on — directly (a source ingredient) or as a related pattern. Links follow the typed catalog namespace.

Built directly on (6)

• Identity Test: The rule that decides, in a given system, when two presentations refer to the same entity.
• Object Permanence: A system maintains persistent internal state for entities it cannot currently observe, modeling them as continuing to exist and re-binding them on re-emergence.
• Observability: Infer internal state externally.
• Persistent Identifier: A designed token committed to keep resolving to its entity across changes in the entity's location, representation, custodian, or version, via a separately maintained resolver.
• State and State Transition: Captures system condition and evolution.
• Uncertainty: Incomplete knowledge.

Also references 19 related abstractions

• Absence Of Evidence Vs Evidence Of Absence: A null finding becomes evidence against a claim only in proportion to how likely the search was to detect the claim had it been true.
• Associative Memory: Content-addressable storage where a cue retrieves linked content.
• Confidence Intervals: Range of plausible values.
• Continuity vs. Rupture: Gradual vs abrupt change.
• Criteria of Individuation: The rules a system fixes for what makes something one entity — when parts compose a single whole, when two presentations are the same entity, and which kind supplies a thing's persistence — together constituting its inventory of countable individuals.
• Data Integrity: Accuracy and consistency preserved.
• Develops-From Relation: A later entity came to be by stage-wise transformation of a continuing predecessor, the same continuant passing through directed, qualitatively distinct stages under a generative rule.
• Eventual Consistency: Distributed copies of shared state are allowed to diverge under local updates, with a deterministic merge guaranteeing they reconverge once updates stop.
• Frame Problem: After a change, the challenge of bounding which facts now need re-derivation and which persist by default, without re-checking everything.
• Identifiability: Whether an internal unknown can in principle be uniquely recovered from the observable signal a system makes available.

▸ Show 9 more

Variants

Narrower or domain-specific specializations that share this archetype's core structure. Recognized variants are established; candidate variants are provisional.

Predictive Occlusion Continuity · mechanism or method variant · recognized

Maintain a predicted state distribution for a temporarily occluded or unsensed entity and reconnect a later observation to the same identity when the evidence supports it.

• Distinct from parent: It emphasizes motion or state prediction, covariance growth, and gating of candidate reappearances by a dynamic envelope.
• Use when: The entity is expected to move or change during the observation gap; A dynamics model and observation-quality estimate can constrain plausible reappearance states; False track termination or duplicate identity creation has material cost.
• Typical domains: robotics, computer vision, navigation, wildlife tracking
• Common mechanisms: predictive state filter, reappearance association protocol, identity resolution workflow

Dormant-Record Continuity · data or evidence variant · recognized

Retain an identity-bearing record through a period of non-contact or non-reporting without pretending to know its current detailed state.

• Distinct from parent: It uses a dormant lifecycle state, sparse updates, and retention governance instead of continuous predictive tracking.
• Use when: The entity may persist while detailed state prediction would be speculative; Reappearance must recover history, obligations, or provenance rather than create a duplicate record; Retention and deletion duties can be stated explicitly.
• Typical domains: case management, asset registry, offline devices, longitudinal research
• Common mechanisms: dormant entity registry, grace period, soft delete quarantine window, tombstone or deletion marker, identity resolution workflow

Federated Sighting Reassociation · coordination variant · recognized

Maintain continuity when observations arrive intermittently from multiple observers or replicas and must be reconciled to a shared identity.

• Distinct from parent: It makes source authority, clock/order uncertainty, duplicate sightings, and eventual reconciliation central.
• Use when: No single observer has continuous coverage; Reports can arrive late, out of order, or with inconsistent identifiers; Duplicate and conflicting sightings must be reconciled without erasing provenance.
• Typical domains: distributed systems, ecological monitoring, supply chains, multi sensor tracking
• Common mechanisms: multi observer sighting reconciliation, persistent identifier resolver, identity resolution workflow, reconciliation workflow

Near names: Object Permanence, Occlusion-Tolerant Entity Continuity, Unobserved Entity Persistence, Persistent Entity Tracking, Dormant Entity Continuity.