Absence-Evidence Calibration Test¶
Calibration test — instantiates Entity Persistence Across Observation Gaps
Rates how informative a non-detection actually is — by asking how likely the channel would have seen the entity if it were there — so a weak-coverage silence can't be read as strong evidence of absence, and only a genuinely informative absence is allowed to trigger retirement.
The dangerous silence is the ambiguous one: you didn't see the entity, but did you look hard enough that not-seeing means anything? Absence-Evidence Calibration Test answers exactly that question. It doesn't track the entity or predict its state; it characterizes the observation channel — estimating the detection probability, the chance the system would have registered the entity had it been present under the current conditions — and uses that number to convert a raw non-observation into a properly weighted piece of evidence about existence. A missed sighting from a channel that covers 5% of the space is nearly meaningless; the same silence from a channel that would almost certainly have caught the entity is strong evidence it is gone. Its second, downstream job follows directly: it owns the retirement trigger — the rule that an entity may be declared absent and retired only once accumulated non-detections are genuinely informative enough to clear a stated bar.
Example¶
A wilderness search-and-rescue team has spent two days looking for a missing hiker in a defined search area and found nothing. The tempting conclusion — "she isn't here, redeploy elsewhere" — is only as good as the search's coverage. The Absence-Evidence Calibration Test is the discipline that stops the team from over-reading their empty-handedness. For each sector they estimate a probability of detection: how thoroughly it was swept, terrain and vegetation density, visibility, searcher spacing, whether the subject might be unresponsive. A steep, brush-choked ravine searched once at a jog gets a low detection probability; an open meadow walked in a tight line gets a high one.
Now the two days of "nothing" resolve into something usable. The meadow's clean negative genuinely lowers the odds the hiker is there; the ravine's negative barely moves them and flags the sector for a re-search rather than abandonment. Only when the cumulative probability of detection across repeated, well-covered searches is high enough[1] does the calibrated absence justify shifting the search's center of gravity — the field's standard reasoning from probability of detection toward probability of success. The team retires a sector from active search on evidence about their own coverage, not on the bare fact that they came back empty.
How it works¶
- Model the detection channel. Estimate the probability the entity would be observed if present — from sensor coverage, search effort, reporting behavior, vegetation, range, or duty cycle — under the current conditions, not idealized ones.
- Weight the non-detection by it. A missed observation lowers the existence/presence belief in proportion to detectability: high detectability turns silence into strong evidence of absence; low detectability makes it nearly uninformative.
- Accumulate across attempts. Combine repeated non-detections and their coverages so that many weak looks or a few strong ones can, together, cross a threshold that no single look could.
- Own the retirement bar. Declare the entity absent and eligible for retirement only when calibrated absence evidence clears an explicit threshold — never on a raw miss.
Tuning parameters¶
- Detection-probability model — how coverage is estimated (fixed table, effort-based curve, physics-based sensor model). Richer models calibrate better but demand data the channel may not have.
- Retirement threshold — how strong calibrated absence must be before retirement fires. High thresholds guard against premature deletion but keep dormant records alive longer; low ones tidy up faster but risk erasing entities that were merely well-hidden.
- Accumulation rule — how repeated non-detections compound (independent multiplication versus discounted, correlation-aware combination). Naïve independence overstates confidence when searches repeat the same blind spots.
- Condition sensitivity — how strongly current conditions (night, weather, occlusion) discount detectability. Ignoring them lets a fair-weather coverage number license an unfair-weather absence claim.
When it helps, and when it misleads¶
Its strength is that it is the direct antidote to premature deletion and to false confidence in absence: it refuses to let a low-coverage channel manufacture a high-confidence disappearance, and it refuses to let a high-coverage negative be ignored forever. It puts a defensible number under the invariant that an absence claim must never exceed demonstrated detection capability — and it gives retirement an auditable evidentiary basis instead of a bare timer.
Its failure mode is a miscalibrated detection model. Overstate coverage and every silence reads as death, reviving premature deletion in a respectable disguise; understate it and nothing is ever retired, and ghosts accumulate. The subtler misuse is treating the absence of a positive as a positive for absence without the coverage step at all — the well-known trap that absence of evidence is not evidence of absence[2] — which is exactly the reasoning this test exists to discipline. The guard is to validate detection probability against known-present cases (does the channel actually catch entities it should?), recheck it whenever conditions change, and keep the retirement threshold high where a false retirement is costly.
How it implements the components¶
absence_evidence_calibration— its core: it estimates detectability and converts each non-observation into evidence about existence weighted by how likely detection would have been.retirement_rule— it owns the trigger, permitting an entity to be declared absent and retired only when calibrated absence evidence crosses an explicit, auditable threshold.
It does not itself carry a current-state estimate through the gap — the propagated observation_gap_state_model and its confidence_and_uncertainty_state are Predictive State Filter — and it does not store the dormant record or field the visibility_existence_separation and persistent_entity_state_record it acts on; those belong to Dormant Entity Registry. This test rates silence and rules on retirement; it neither tracks nor stores.
Related¶
- Instantiates: Entity Persistence Across Observation Gaps — supplies the calibrated meaning of non-detection and the evidence-based retirement trigger.
- Consumes: Dormant Entity Registry — reads the dormant records and their non-detection history, then rules whether they may retire.
- Sibling mechanisms: Predictive State Filter · Dormant Entity Registry · Reappearance Association Protocol · Multi-Observer Sighting Reconciliation · Soft-Delete Quarantine Window
Editorial Notes¶
Form Classification¶
Form family: Assessment, Review & Assurance
Rationale: The mechanism evaluates accumulated non-detections against modeled detectability to determine how much evidentiary weight they deserve and whether the retirement bar is met, so its operative form is a bounded assessment of existing evidence.
Nearest alternative: Analysis, Modeling & Optimization — Detectability modeling supplies the calibration, but the mechanism's defining result is a case-level evidentiary finding and retirement judgment rather than a reusable prediction or optimized solution.
Review outcome: Adjudicated after independent review; medium confidence.
Origin Attribution¶
Primary origin: Operations Research
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Probability-of-detection reasoning and cumulative search coverage were characteristically formalized in wartime search theory and operations research, especially Koopman's search models.
Related originating lineages:
- Biology & Ecology — Occupancy, detectability, and repeated-survey methods developed this reasoning for elusive species and imperfect sensors.
- Disaster Management & Risk Reduction — Search-and-rescue practice made probability of detection, containment, and sector retirement an operational protocol.
- Statistics & Experimental Design — Likelihood-based updating and calibration against known-present cases supply the general inferential treatment of non-detection.
Review resolution: The mechanism's cumulative search coverage and retire-or-search-again decision match operations-research search theory most closely; statistical inference, occupancy detectability, and search-and-rescue operationalize the same logic in other lineages.
Attribution caveat: The detection model has a clear search-theory origin; extending it into a general entity-retirement rule is an Encyclopedia synthesis.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Researched adjudication after independent review; high confidence.
Sources consulted:
- RAND application of Koopman search theory — Describes operations-research search theory as relating detection probability to search time and effort.
- USGS Evidence of Absence Software User Guide — Shows how search-process detection probabilities calibrate what zero detections imply.
References¶
[1] Koopman, B. O. Search and Screening: General Principles with Historical Applications. Pergamon Press (1980). Supports accumulation of detection probability across repeated, covered search opportunities. registry ↩
[2] "Absence of evidence is not evidence of absence" — the principle that failing to observe something counts against its existence only to the degree observation was capable and expected. It was given its familiar statistical statement by Douglas Altman and J. Martin Bland (BMJ, 1995); used here as an established aphorism, not a quoted result. registry ↩