The Sniff schemas.

A bounded question space over a defined population, where the absence of an assertion is itself informative because an answer was expected. The unit of reasoned absence: it names the question, the populations, the expected cardinality, and how the expected answers resolve into answered vs dark cells.

slots: id, question, subject_population, object_population, predicate, expected_cardinality, answered_count, dark_count, dark_by_cause, cell_source, sources_checked, as_of

One subject in a CoverageFrame. Its status is answered (a cited assertion exists), negated (an assertion that the relationship is false exists), or dark (no assertion exists, with a named cause). A dark cell is the payload: it turns a missing edge into a typed, actionable statement about what is not yet known and why.

slots: frame, subject, symbol, object, status, dark_cause, association, recoverability, n, basis, floor_id, as_of

One cause's share of a frame's dark cells. Serialized in the compact dict form, so a tally reads as {cause: count} keyed by DarkCauseEnum.

The three resolutions of an expected answer cell.

• answered · An assertion exists and is cited.
• negated · An assertion exists that the relationship is false. Corresponds to a Biolink association carrying negated = true; biolink:negated is a slot, not a concept, so it is a see_also rather than a meaning.
• dark · No assertion exists; see dark_cause.

The causes of darkness, modeled as they are actually computed. Only absent is expressible in Biolink today (as Association.negated). Prior art verified 2026-08-19 against Biolink v4.4.4, the ECO 2026-07-10 release, GO, and FHIR: the nearest existing vocabularies are single-cause and record-scoped. ECO:0000035 (no evidence data found; GO's ND code is its manual-assertion child ECO:0000307) is the closest cousin of unstudied (a literature or exhaustion search that found nothing) and is bound below as a close mapping. no_disease_anchor is a different claim: no ClinVar P/LP disease-gene anchor to bridge from, not that the gene is unstudied. FHIR data-absent-reason types why one field on one existing record is empty. No vocabulary we could find types the cause of absence per cell over a declared, bounded question space, and below_bar, unreachable, method_limited, and no_disease_anchor have no term in Biolink or any evidence ontology checked. Where no mapping appears below, that absence is verified and load-bearing, not an oversight.

• absent · Studied, and the entity or relationship does not exist. The one cause Biolink can already express, as an association with negated = true. A cell whose cause is absent has status negated, never dark: absent is an evidenced no, not a darkness, and check-conformance refuses the pairing dark + absent.
• unstudied · A literature or exhaustion search addressing this question in this species found nothing. Not for "no Mendelian disease-gene anchor in ClinVar"; that is no_disease_anchor.
• no_disease_anchor · Dog ortholog exists but no human P/LP disease evidence to bridge from (ascertainment darkness). Not a claim that the gene is unstudied in the literature.
• below_bar · Human evidence exists but sits under our review bar (for example under a 3-star ClinVar threshold); distinct from method_limited. Relates to evidence strength (ECO) but no ECO term names "below a downstream bar".
• unreachable · No ortholog or entity exists for the assay to observe.
• method_limited · The method cannot resolve this class of entity even where the underlying data are present.

How a present claim was produced. Companion to DarkCauseEnum, which names absence. Strength is n plus basis plus floor. A score float is forbidden.

• measured · Instrument or lab-backed on this subject.
• estimated · Inferred or owner-sketched, never dressed as a measurement.
• cited_cohort · A published or atlas cohort statistic with a cited n.
• owner_register · An owner-contributed aggregate, cohort-of, never population-of.
• this_dog_lived · Repeated measures on one dog. Never a breed fact.

The reproducible procedure that resolved a frame's cells. The strongest form of a methods section is a pointer to runnable open code at a pinned version. Determinism is the contract; the specific precedence order is a protocol parameter, so two parties running the same protocol version on the same populations must produce the same cells, and a disagreement is diagnosable rather than a shrug.

slots: id, protocol_version, pipeline, match_criteria, completeness, cause_precedence, dial_references, positive_controls, darkness_controls, executed_at, recall, end_to_end_recall, filter_independence, negative_controls

A versioned, counted population a frame enumerates over. Identity is the INTENSION: the id's version (an integer, as in sniff:population/x@1) bumps only when the selection procedure itself changes, never when a new source release changes the membership. The source_version and count are per-release measurements. The count doubles as a drift checksum: same definition version, same source_version, and a differing count means the pipeline drifted (a bug); a newer source_version under the same definition is the world moving, and the delta module enumerates the membership change as boundary events.

slots: id, source, source_version, selection, count

One held-out cell known to be DARK FOR A SPECIFIC CAUSE, independently of the pipeline: a primate-only gene for unreachable, a gene whose ClinVar records all sit under the declared bar for below_bar, a gene with a clean ortholog and no disease anchor for no_disease_anchor. The protocol must observe the expected cause on every darkness control before its dark cells may ship, and it must carry at least one for every cause its tallies claim. Positive controls prove the instrument finds what exists; darkness controls prove it names why it does not, which is the product. Added after both dark cells of a demonstration frame turned out to be mistyped while every positive control passed: an instrument calibrated only on the cells that light has proven nothing about the cells that do not.

slots: subject, expected_cause, observed_cause, observed_status, basis, basis_kind, receipt_source, receipt_as_of, outcome

One held-out cell known to be answerable, independent of the pipeline (for Sniff's reference frame: canine disease genes documented in OMIA before this pipeline existed). The protocol must observe the expected status on every control before its dark cells may ship. Calibration is the difference between measured ignorance and a parsing bug wearing ignorance's clothes.

slots: subject, expected_status, observed_status, basis, basis_kind, receipt_source, receipt_as_of, outcome

A CoverageFrame that carries its warrant: the protocol that resolved it and the versioned populations it enumerated. The core class stays frozen; adopters who cannot yet publish protocols emit plain CoverageFrames and graduate to this class when the procedure is publishable.

slots: search_protocol, subject_population_ref, object_population_ref, release_id, conforms_to, scope, constituents

What this frame measured, and what it deliberately did not. A question that outruns its sources is an overclaim every dark cell inherits. sources_queried must match the frame's sources_checked. sources_not_queried names known authorities left out, with a reason and a named path to close each gap, so an unqueried source is a typed limitation rather than silence.

slots: measures, sources_queried, sources_not_queried

One authority this frame did not query. The reason is why, for this question. close_path is how the gap becomes queried rather than remaining an implicit hole in sources_checked.

slots: source, reason, close_path

A measured recall figure bound to the identity of the control set that produced it. A recall number whose denominator can silently change is the failure this class exists to prevent.

slots: value, control_set, control_set_hash, control_set_size, dial, measured_at

How much two candidate filters fail on the same items. When the observed shared-miss count materially exceeds the independence expectation, stacking further filters of that family will not recover recall.

slots: filters_compared, shared_miss_jaccard, observed_over_expected, marginal_gain

How completely the protocol enumerated the population, defined operationally so the term is checkable rather than a vibe.

• exhaustive · The protocol evaluated every cell in the declared population.
• sampled · The protocol evaluated a declared fraction under a declared sampling method. Dark tallies extrapolate; absent is not assertable.
• heuristic · Enumeration is not guaranteed (for example keyword search). The weakest honest claim; absent is not assertable.

The kind of evidence a control's basis is, ranked by independence from the protocol's own source.

• independent_literature · A published paper. Independent of every database the protocol queries; immutable, so a mismatch is always a pipeline defect.
• independent_source · A database other than the one the protocol queries for this cell (for example OMIA for a ClinVar-anchored protocol). Independent of the source, not of the world.
• same_source_hand_verified · The protocol's own source, read by hand outside the pipeline, with a dated receipt. Catches builder bugs, parsing errors and wrong filters; catches no error in the source. A control of this kind on a mutable database is a world-movement tripwire and routes to world_moved, never to a build failure on its own.

What the protocol run did to a control.

• held · Observed equals expected. The only outcome that counts toward calibration.
• world_moved · Observed differs from expected and the receipt facts no longer hold in the ingested source. The world moved; re-plant a control for the cause.
• failed · Observed differs from expected while the receipt facts still hold. The funnel is wrong. Nothing ships.

How a cell resolves across the constituent frames of an intersection. These tokens are scoped to the named stack. They are not claims about sources outside that stack. A third frame must use these values; it must not invent a parallel convention.

• lit_all · Answered in every constituent frame of this stack.
• lit_any · Answered in at least one constituent frame of this stack, and not in all of them.
• dark_in_all_frames · Dark in every constituent frame of this stack. Not a claim that nobody has looked, and not a claim about sources that are not in the stack.
• unresolved · Not evaluated by this stack. The subject is off the shared axis, or a constituent did not emit a cell.

The typed change between two releases of one frame. Comparable only when kind is world_moved; a frame_moved delta documents that the question itself changed and its transitions are annotations, not knowledge flow.

slots: id, previous_frame, current_frame, kind, transitions, entered, left

One cell's move between statuses or causes across a release, with a typed trigger and a citation for what moved it. The dark-to-answered transition is the product's core loop made schema: the record updates itself and the owner is told, with a citation.

slots: subject, symbol, from_status, from_cause, to_status, to_cause, trigger, reference

One cell entering or leaving the frame's universe across a release. Under world_moved this is knowledge flow (a new source release minted or retired the question cell itself, for example a gene newly carrying a pathogenic classification); under frame_moved it is an annotation of the changed question. Either way it is enumerated, because a silently appearing or vanishing cell is the delta equivalent of an untyped gap.

slots: subject, symbol, status, dark_cause, trigger, reference

What changed between two releases. Never blended. Held two values until a release corrected a source identifier without the world or the method moving: identical definitions over an identical upstream snapshot, and not one cell flipped. That is not the world moving, and world_moved was acting as the residual bucket rather than as a claim about the world.

• world_moved · Identical populations and protocol; cells flipped because knowledge changed. The conservation identity binds.
• frame_moved · Populations, protocol, or dials changed; the question changed. Not comparable to the predecessor as knowledge flow.
• correction · Nothing moved. Identical populations, protocol and dials over an identical upstream snapshot; the release exists to repair our own record. Counts are expected to be unchanged, and any cell that does flip flips because the previous release was wrong about it.

The typed causes of a cell transition.

• source_release · An upstream source shipped a new release.
• new_publication · A new primary publication answered or negated the cell.
• method_change · The protocol or pipeline changed how it resolves cells.
• criteria_change · A named dial changed (the bar moved, visibly).
• population_change · The population definition or its source version changed.
• correction · The previous release was our error. The typed, public, cited repair path; never a silent overwrite.

One owner-contributed fact about a dog, born with a FactVisibility intent (refuse-null). The value is carried under a stable typed key; provenance records whether it was declared, measured, or imported.

slots: id, dog_id, kind, key, value, visibility, provenance, created_at, n, basis, floor_id

The result of composing a fact's visibility intent with the dog's visibility tier and the owner's trust tier. science here is an eligibility gate to count, not a permission to render a cohort rate.

slots: owner, science, showcase

The category of an owner-contributed fact.

• answer · A response to a dog-questions item.
• lifestyle · A structured lifestyle seed such as acquisition window or diet class.
• note · A free-text owner note; born private.
• vet_doc · A clinical document handle; born private, handling deferred.

The owner's per-field visibility intent. The same three-value vocabulary is shared across the record (see INV-116); it is refuse-null and fails closed to private.

• private · The owner's own lens only.
• science · Eligible for an anonymized cohort count when trust clears; never shown publicly.
• public · Eligible for showcase when the dog is public and trust clears.

How a value was obtained. ECO evidence-code bindings are a planned Phase 2 extension; they are omitted here rather than asserted unverified.

• declared · Owner-reported, carried in estimated language.
• measured · Instrument or lab-backed.
• imported · Brought in from an external source with a provenance trail.

One entity mention in one document, with the offsets that locate it and the adjudication that judged it.

slots: id, subject, taxon, document, passage_offset, annotation_offset, annotation_length, extractor, extractor_version, retrieved_date, archive_hash, filter_flags, proposal, adjudication, adjudication_basis, adjudication_note, sentence_text

A machine's opinion about a sighting, recorded before any human judged it. Held as its own class rather than as fields alongside adjudication on purpose: a proposal and a verdict are different kinds of thing, and a flat proposed sitting beside adjudication is one mis-read away from a compiler typing a cell on a machine's say-so. INV-145 forbids exactly that, so the shape is built to make the mistake awkward rather than merely detectable. Nothing here may promote a cell; only a merged human adjudication may. A proposal is what the adjudication ladder measures. Blind review withholds it from the reviewer, and blind agreement between this and the human verdict is the only agreement figure that means anything.

The three states of a sighting.

• pending · Observed, not yet judged. May not support a frame cell.
• confirmed · Judged to support a claim about this entity in this taxon.
• rejected · Judged not to support such a claim.

Why an adjudication went the way it did. Turns the species-assignment problem from a caveat into a measured quantity.

• confirmed_canine · The document studies this entity in this taxon.
• species_ambiguous · Entity and taxon both present, but which species the claim attaches to is not resolvable from the text.
• wrong_species_context · The entity is studied in another species and this taxon is mentioned incidentally, as in a comparative genomics survey.
• dog_as_comparator · The taxon appears as a comparison or control group, not the subject.
• gene_mismatch · The annotation resolves to a different entity, an alias or symbol collision.
• insufficient_text · Undecidable without full text.