@agenticprimitives/ontology 1.0.0-alpha.4 → 1.0.0-alpha.40

package/tbox/evaluation.ttl

@@ -0,0 +1,330 @@
+# T-box — MEASUREMENTS of a harness run (spec 414 §7, A2). Namespace apeval:.
+#
+# ── What this is and is NOT ───────────────────────────────────────────────────────────────────────
+# A measurement is a NUMBER about a run or a step, computed from its trace alone — completion, verdict counts,
+# latency, model calls, tokens, vault calls, replans. It is DESCRIPTIVE: nothing here says whether an answer was
+# right (that is a judgment — spec 414 §3.4, a different record with a named criterion and judge). It is EVIDENCE,
+# never authority: no gate reads an apeval: term (spec 354 §1 / ADR-0041). And it carries no words: a measurement is
+# a metric, a value and a unit (shape M3).
+#
+# Every term specializes a standard one (spec 414 §2 rule 2): a metric is a dqv:Metric, a measurement a
+# dqv:QualityMeasurement (and a prov:Entity, generated by a measuring activity that used the run's bundle).
+# A metric's IRI is DERIVED from its skos:notation — `m-` + camelCase (`latency-ms` → `apeval:m-latencyMs`) — so
+# code names a metric by notation and never by a lookup table (asserted in the ontology suite).
+
+@prefix apeval: <https://agenticprimitives.dev/ns/evaluation#> .
+@prefix apexec: <https://agenticprimitives.dev/ns/execution#> .
+@prefix dqv:    <http://www.w3.org/ns/dqv#> .
+@prefix prov:   <http://www.w3.org/ns/prov#> .
+@prefix skos:   <http://www.w3.org/2004/02/skos/core#> .
+@prefix rdfs:   <http://www.w3.org/2000/01/rdf-schema#> .
+@prefix owl:    <http://www.w3.org/2002/07/owl#> .
+@prefix xsd:    <http://www.w3.org/2001/XMLSchema#> .
+@prefix dcat:   <http://www.w3.org/ns/dcat#> .
+@prefix earl:   <http://www.w3.org/ns/earl#> .
+@prefix dcterms: <http://purl.org/dc/terms/> .
+@prefix ep-plan: <https://w3id.org/ep-plan#> .
+@prefix p-plan: <http://purl.org/net/p-plan#> .
+@prefix dul:    <http://www.ontologydesignpatterns.org/ont/dul/DUL.owl#> .
+
+apeval: a owl:Ontology ;
+  rdfs:label "Agentic Primitives — run measurements" ;
+  owl:versionInfo "0.5.0" ;
+  rdfs:comment "Descriptive measurements of a harness run, on W3C DQV. Spec 414 §7 / 415 A2." .
+
+apeval:Metric a owl:Class ;
+  rdfs:subClassOf dqv:Metric , skos:Concept ;
+  rdfs:label "Metric" ;
+  rdfs:comment "A quantity computed from a run's trace alone. Declared here once; code names it by skos:notation. NOT a criterion (a stated expectation — spec 414 §3.4) and NOT a score of correctness." .
+
+apeval:Metrics a skos:ConceptScheme ; rdfs:label "Run metrics" .
+
+apeval:RunMeasurement a owl:Class ;
+  rdfs:subClassOf dqv:QualityMeasurement , prov:Entity ;
+  rdfs:label "RunMeasurement" ;
+  rdfs:comment "One measurement of one metric on one run or step (dqv:computedOn the run's or the step's activity). Generated by an apeval:Measuring activity that prov:used the run's bundle." .
+
+apeval:Measuring a owl:Class ;
+  rdfs:subClassOf prov:Activity ;
+  rdfs:label "Measuring" ;
+  rdfs:comment "The activity of computing a run's measurements from its trace. Mechanical and reproducible: the same trace yields the same measurements." .
+
+apeval:RunMeasures a owl:Class ;
+  rdfs:subClassOf dqv:QualityMeasurementDataset , prov:Bundle ;
+  rdfs:label "RunMeasures" ;
+  rdfs:comment "One run's measurements as one record (`run.measures:<runRef>` in the acting agent's vault), prov:wasDerivedFrom the run's provenance bundle." .
+
+apeval:unit a owl:DatatypeProperty ;
+  rdfs:domain apeval:RunMeasurement ; rdfs:range xsd:string ;
+  rdfs:label "unit" ;
+  rdfs:comment "UCUM code of the value: `ms`, `{count}`, `1` (a ratio or a 0/1 fact)." .
+
+apeval:hasMeasurement a owl:ObjectProperty ;
+  rdfs:domain apeval:RunMeasures ; rdfs:range apeval:RunMeasurement ;
+  rdfs:label "hasMeasurement" .
+
+# ── The metrics (spec 414 §7, intrinsic — from the trace alone) ──
+apeval:m-completed a apeval:Metric ; skos:inScheme apeval:Metrics ; skos:notation "completed" ; skos:prefLabel "completed" ; skos:definition "1 when the run completed, 0 otherwise." .
+apeval:m-stepsPlanned a apeval:Metric ; skos:inScheme apeval:Metrics ; skos:notation "steps-planned" ; skos:prefLabel "steps planned" .
+apeval:m-stepsRun a apeval:Metric ; skos:inScheme apeval:Metrics ; skos:notation "steps-run" ; skos:prefLabel "steps run" ; skos:definition "Steps that executed (replayed steps excluded)." .
+apeval:m-stepsRefused a apeval:Metric ; skos:inScheme apeval:Metrics ; skos:notation "steps-refused" ; skos:prefLabel "steps refused" .
+apeval:m-verdictsAllow a apeval:Metric ; skos:inScheme apeval:Metrics ; skos:notation "verdicts-allow" ; skos:prefLabel "authority verdicts: allow" .
+apeval:m-verdictsDeny a apeval:Metric ; skos:inScheme apeval:Metrics ; skos:notation "verdicts-deny" ; skos:prefLabel "authority verdicts: deny" .
+apeval:m-approvalsWaited a apeval:Metric ; skos:inScheme apeval:Metrics ; skos:notation "approvals-waited" ; skos:prefLabel "approvals waited on" .
+apeval:m-replans a apeval:Metric ; skos:inScheme apeval:Metrics ; skos:notation "replans" ; skos:prefLabel "replans forced by admission" .
+apeval:m-resumedSteps a apeval:Metric ; skos:inScheme apeval:Metrics ; skos:notation "resumed-steps" ; skos:prefLabel "steps replayed on resume" .
+apeval:m-latencyMs a apeval:Metric ; skos:inScheme apeval:Metrics ; skos:notation "latency-ms" ; skos:prefLabel "latency" ; skos:definition "From the request's arrival (or the first step) to the run's end; on a step, the step's own window." .
+apeval:m-modelCalls a apeval:Metric ; skos:inScheme apeval:Metrics ; skos:notation "model-calls" ; skos:prefLabel "model calls" .
+apeval:m-tokensIn a apeval:Metric ; skos:inScheme apeval:Metrics ; skos:notation "tokens-in" ; skos:prefLabel "tokens in" ; skos:definition "Provider-reported input tokens; absent when no provider reported them (never estimated)." .
+apeval:m-tokensOut a apeval:Metric ; skos:inScheme apeval:Metrics ; skos:notation "tokens-out" ; skos:prefLabel "tokens out" .
+apeval:m-vaultCalls a apeval:Metric ; skos:inScheme apeval:Metrics ; skos:notation "vault-calls" ; skos:prefLabel "vault calls" .
+apeval:m-effectsConfirmed a apeval:Metric ; skos:inScheme apeval:Metrics ; skos:notation "effects-confirmed" ; skos:prefLabel "effects confirmed" ; skos:definition "Steps whose adapter observation reached `confirmed` (spec 410 §2)." .
+apeval:m-engagements a apeval:Metric ; skos:inScheme apeval:Metrics ; skos:notation "engagements" ; skos:prefLabel "harness-capability engagements" .
+
+# ═════════════════════════════════════════════════════════════════════════════════════════════════
+# Spec 415 S0 + SEL-1 — REPLAY SETS, CRITERIA, and the first JUDGMENT: skill selection.
+#
+# A replay set is a versioned dataset of intents (DCAT), each optionally bound to a starting state. A CRITERION is a
+# stated expectation, pre-registered by digest before any run it judges (spec 414 §3.4 J2): for selection, the gold
+# skill and the acceptable set of each intent. A JUDGMENT is an activity by a NAMED judge of a declared KIND that used
+# a subject and a criterion and generated a verdict (EARL) — evidence, never authority (J1). The selection outcome
+# classes follow the AI Skills Commons trigger-accuracy methodology v1.0 (§1), named here once.
+# ═════════════════════════════════════════════════════════════════════════════════════════════════
+
+apeval:ReplaySet a owl:Class ;
+  rdfs:subClassOf dcat:Dataset ;
+  rdfs:label "ReplaySet" ;
+  rdfs:comment "A versioned, checksummed set of intents to replay under variants (spec 415 S0). Split into development and held-out (pre-registered by digest). NOT the criteria: gold lives in a separate criterion record." .
+
+apeval:ReplayIntent a owl:Class ;
+  rdfs:subClassOf prov:Entity ;
+  rdfs:label "ReplayIntent" ;
+  rdfs:comment "One intent of a replay set, cited by id and digest. Its words live in the set's file; a run's graph never carries them." .
+
+apeval:hasIntent a owl:ObjectProperty ; rdfs:domain apeval:ReplaySet ; rdfs:range apeval:ReplayIntent ; rdfs:label "hasIntent" .
+apeval:split a owl:DatatypeProperty ; rdfs:domain apeval:ReplayIntent ; rdfs:range xsd:string ; rdfs:label "split" ; rdfs:comment "development | held-out." .
+apeval:bucket a owl:DatatypeProperty ; rdfs:domain apeval:ReplayIntent ; rdfs:range xsd:string ; rdfs:label "bucket" ; rdfs:comment "The case bucket (explicit · implicit · contextual · negative-near · negative-scope · negative-out · collision-a · collision-b · collision-ambiguous)." .
+apeval:setDigest a owl:DatatypeProperty ; rdfs:domain apeval:ReplaySet ; rdfs:range xsd:string ; rdfs:label "setDigest" ; rdfs:comment "sha256 over the set's canonical JSON (RFC 8785)." .
+
+apeval:SelectionCriterion a owl:Class ;
+  rdfs:subClassOf ep-plan:Constraint ;
+  rdfs:label "SelectionCriterion" ;
+  rdfs:comment "What selection SHOULD do for one intent: the gold skill (or none, for a negative) and the acceptable alternates, pre-registered by digest. A hypothesis until adjudicated." .
+
+apeval:criterionDigest a owl:DatatypeProperty ; rdfs:domain apeval:Judgment ; rdfs:range xsd:string ; rdfs:label "criterionDigest" ;
+  rdfs:comment "sha256 of the criterion record the judgment applied (J2: stated before, cited by digest)." .
+
+apeval:Judgment a owl:Class ;
+  rdfs:subClassOf prov:Activity ;
+  rdfs:label "Judgment" ;
+  rdfs:comment "An activity that judged a subject (a run or a step) against a criterion and generated a verdict (spec 414 §3.4). Evidence, never authority; never rewrites the trace it judged (J1, J6)." .
+
+apeval:Judge a owl:Class ;
+  rdfs:subClassOf prov:Agent ;
+  rdfs:label "Judge" ;
+  rdfs:comment "Who or what judged. Its KIND is declared (J3) and verdicts of different kinds are never pooled (J4)." .
+
+apeval:judgeKind a owl:DatatypeProperty ; rdfs:domain apeval:Judge ; rdfs:range xsd:string ; rdfs:label "judgeKind" ; rdfs:comment "mechanical | model | human." .
+
+apeval:SelectionOutcome a owl:Class ;
+  rdfs:subClassOf skos:Concept ;
+  rdfs:label "SelectionOutcome" ;
+  rdfs:comment "The class of a selection verdict (methodology v1.0 §1)." .
+apeval:SelectionOutcomes a skos:ConceptScheme ; rdfs:label "Skill-selection outcomes" .
+
+apeval:selectionOutcome a owl:ObjectProperty ; rdfs:domain earl:TestResult ; rdfs:range apeval:SelectionOutcome ; rdfs:label "selectionOutcome" .
+apeval:chargedTo a owl:DatatypeProperty ; rdfs:domain earl:TestResult ; rdfs:range xsd:string ; rdfs:label "chargedTo" ; rdfs:comment "The skill a spurious or wrong selection is charged to (a canonical id)." .
+apeval:evidenceKind a owl:DatatypeProperty ; rdfs:domain earl:TestResult ; rdfs:range xsd:string ; rdfs:label "evidenceKind" ; rdfs:comment "stamped (the harness recorded governedBySkill) | inferred (an opaque agent, by detection tier). Never pooled." .
+
+apeval:so-tp a apeval:SelectionOutcome ; skos:inScheme apeval:SelectionOutcomes ; skos:notation "tp" ; skos:prefLabel "correct" ; skos:definition "The gold skill was selected." .
+apeval:so-tpAlt a apeval:SelectionOutcome ; skos:inScheme apeval:SelectionOutcomes ; skos:notation "tp-alt" ; skos:prefLabel "acceptable alternative" ; skos:definition "A declared acceptable alternate was selected (excluded from accuracy, reported beside it)." .
+apeval:so-misSib a apeval:SelectionOutcome ; skos:inScheme apeval:SelectionOutcomes ; skos:notation "mis-sib" ; skos:prefLabel "sibling miss" ; skos:definition "The gold skill's declared collision partner was selected instead." .
+apeval:so-misFar a apeval:SelectionOutcome ; skos:inScheme apeval:SelectionOutcomes ; skos:notation "mis-far" ; skos:prefLabel "far miss" ; skos:definition "Some other skill was selected." .
+apeval:so-miss a apeval:SelectionOutcome ; skos:inScheme apeval:SelectionOutcomes ; skos:notation "miss" ; skos:prefLabel "miss" ; skos:definition "A skill was expected and none was selected." .
+apeval:so-tn a apeval:SelectionOutcome ; skos:inScheme apeval:SelectionOutcomes ; skos:notation "tn" ; skos:prefLabel "correct abstention" ; skos:definition "No skill was expected and none was selected." .
+apeval:so-spur a apeval:SelectionOutcome ; skos:inScheme apeval:SelectionOutcomes ; skos:notation "spur" ; skos:prefLabel "spurious" ; skos:definition "No skill was expected and one was selected; charged to it." .
+apeval:so-undetected a apeval:SelectionOutcome ; skos:inScheme apeval:SelectionOutcomes ; skos:notation "undetected" ; skos:prefLabel "undetected" ; skos:definition "The selection could not be read; excluded from denominators (a verdict of cantTell, never a pass or a fail)." .
+
+# ── Experiments and the capture window (spec 414 §7–8, 415 A4) ─────────────────────────────────────
+# A comparison is PRE-REGISTERED: the plan names the replay set, the criterion, the variants and the thresholds BY
+# DIGEST before the first run, and scoring refuses across them. An experiment is the activity that ran it. A capture
+# window is the activity inside which results may leave a test estate for the eval store — absent, expired or exceeded
+# ⇒ export refused by name. None of this is authority: nothing here is read by a gate.
+
+apeval:ExperimentPlan a owl:Class ;
+  rdfs:subClassOf p-plan:Plan ;
+  rdfs:label "ExperimentPlan" ;
+  rdfs:comment "A comparison pre-registered by digest: the replay set, the criterion, the variants (each a Variant digest), the split, the repeats and the thresholds, fixed BEFORE the first run (spec 415 §3a). A plan whose components differ in more than one thing between two variants is CONFOUNDED and says so." .
+
+apeval:Experiment a owl:Class ;
+  rdfs:subClassOf prov:Activity ;
+  rdfs:label "Experiment" ;
+  rdfs:comment "The activity that ran an ExperimentPlan: every run it dispatched (by run reference, pseudonymised in the eval store), the judgments and the scores. Evidence, never authority." .
+
+apeval:CaptureWindow a owl:Class ;
+  rdfs:subClassOf prov:Activity ;
+  rdfs:label "CaptureWindow" ;
+  rdfs:comment "The window inside which a test estate's runs may be exported to an eval store (spec 414 §8): the store, the capture level (L0 numbers · L1 structure · L2 graph · L3 words), the roster of test personas, the estate, and when it opens and expires. Export outside it is refused and counted, never downgraded silently." .
+
+apeval:replaySetDigest a owl:DatatypeProperty ; rdfs:domain apeval:ExperimentPlan ; rdfs:range xsd:string ; rdfs:label "replaySetDigest" .
+apeval:variantDigest a owl:DatatypeProperty ; rdfs:domain apeval:ExperimentPlan ; rdfs:range xsd:string ; rdfs:label "variantDigest" ; rdfs:comment "The digest of a Variant a plan compares (apexec:Variant)." .
+apeval:thresholdsDigest a owl:DatatypeProperty ; rdfs:domain apeval:ExperimentPlan ; rdfs:range xsd:string ; rdfs:label "thresholdsDigest" .
+apeval:planDigest a owl:DatatypeProperty ; rdfs:domain apeval:Experiment ; rdfs:range xsd:string ; rdfs:label "planDigest" ; rdfs:comment "The digest an Experiment cites for the plan it ran." .
+apeval:repeats a owl:DatatypeProperty ; rdfs:domain apeval:ExperimentPlan ; rdfs:range xsd:integer ; rdfs:label "repeats" .
+apeval:confounded a owl:DatatypeProperty ; rdfs:domain apeval:ExperimentPlan ; rdfs:range xsd:boolean ; rdfs:label "confounded" ; rdfs:comment "true when two compared variants differ in more than one component (spec 414 §7): the result is labelled, never read as a lift." .
+apeval:captureLevel a owl:DatatypeProperty ; rdfs:domain prov:Entity ; rdfs:range xsd:string ; rdfs:label "captureLevel" ; rdfs:comment "L0 | L1 | L2 | L3 — the level a record was exported at. A number computed with words present is never read as one from the vault-only pipeline." .
+apeval:roster a owl:DatatypeProperty ; rdfs:domain apeval:CaptureWindow ; rdfs:range xsd:string ; rdfs:label "roster" ; rdfs:comment "A test persona admitted by the window, by agent address." .
+apeval:estate a owl:DatatypeProperty ; rdfs:domain apeval:CaptureWindow ; rdfs:range xsd:string ; rdfs:label "estate" ; rdfs:comment "The Home origin whose runs the window admits." .
+apeval:hasRun a owl:ObjectProperty ; rdfs:subPropertyOf prov:used ; rdfs:domain apeval:Experiment ; rdfs:range prov:Entity ; rdfs:label "hasRun" ; rdfs:comment "A run the experiment dispatched (its bundle, pseudonymised)." .
+apeval:refusedCount a owl:DatatypeProperty ; rdfs:domain apeval:Experiment ; rdfs:range xsd:string ; rdfs:label "refusedCount" ; rdfs:comment "`capture.refused.<reason>=<n>` — what the firewall refused, counted, never dropped." .
+
+# ═════════════════════════════════════════════════════════════════════════════════════════════════
+# SPEC 418 §2 — THE TEST CASE, AS DATA. A replay intent read as a TEST: what it expects (a skill, a behaviour, a plan,
+# the inputs to ask for), what it exercises in the domain, from which starting state — so a domain's graph says what
+# its tests consist of and coverage is a query. The JSON sets stay the source of truth (pinned by digest); a domain
+# PROJECTS them into its graph. Evidence and hypothesis, never authority.
+# ═════════════════════════════════════════════════════════════════════════════════════════════════
+
+apeval:expectedSkill a owl:ObjectProperty ; rdfs:domain apeval:ReplayIntent ; rdfs:label "expectedSkill" ;
+  rdfs:comment "The skill the case says should be chosen (a canonical skill IRI). Absent on a negative." .
+apeval:acceptableSkill a owl:ObjectProperty ; rdfs:domain apeval:ReplayIntent ; rdfs:label "acceptableSkill" ;
+  rdfs:comment "A skill that is also right for the case (an adjudicated alternate)." .
+apeval:forbiddenSkill a owl:ObjectProperty ; rdfs:domain apeval:ReplayIntent ; rdfs:label "forbiddenSkill" ;
+  rdfs:comment "A skill the case must NOT be routed to — the near neighbour it is built to trip." .
+apeval:collidesWith a owl:ObjectProperty ; rdfs:domain apeval:ReplayIntent ; rdfs:label "collidesWith" ;
+  rdfs:comment "The skill of the collision pair this case probes (the other side of a routing collision)." .
+apeval:expectedBehaviour a owl:ObjectProperty ; rdfs:domain apeval:ReplayIntent ; rdfs:range apeval:SelectionBehaviour ; rdfs:label "expectedBehaviour" ;
+  rdfs:comment "What selection should DO: route to one skill, decline, ask which (clarify), or name several." .
+apeval:SelectionBehaviour a owl:Class ; rdfs:subClassOf skos:Concept ; rdfs:label "SelectionBehaviour" ;
+  rdfs:comment "route · decline · clarify · several · plan (a chain of skills)." .
+apeval:behaviour-route a apeval:SelectionBehaviour ; skos:notation "route" ; skos:prefLabel "route to one skill" .
+apeval:behaviour-decline a apeval:SelectionBehaviour ; skos:notation "decline" ; skos:prefLabel "decline — no skill fits" .
+apeval:behaviour-clarify a apeval:SelectionBehaviour ; skos:notation "clarify" ; skos:prefLabel "ask which of named options" .
+apeval:behaviour-several a apeval:SelectionBehaviour ; skos:notation "several" ; skos:prefLabel "the request needs several skills" .
+apeval:behaviour-plan a apeval:SelectionBehaviour ; skos:notation "plan" ; skos:prefLabel "reach the outcome by a chain of skills" .
+apeval:optionSkill a owl:ObjectProperty ; rdfs:domain apeval:ReplayIntent ; rdfs:label "optionSkill" ;
+  rdfs:comment "A skill among the options a clarify/several case names." .
+apeval:ExpectedPlanStep a owl:Class ; rdfs:subClassOf p-plan:Step ; rdfs:label "ExpectedPlanStep" ;
+  rdfs:comment "One step of the plan an outcome case expects: a skill at a position. A plan in the gold, not a run." .
+apeval:expectedPlanStep a owl:ObjectProperty ; rdfs:domain apeval:ReplayIntent ; rdfs:range apeval:ExpectedPlanStep ; rdfs:label "expectedPlanStep" .
+apeval:planPosition a owl:DatatypeProperty ; rdfs:domain apeval:ExpectedPlanStep ; rdfs:range xsd:integer ; rdfs:label "planPosition" ; rdfs:comment "0-based order in the expected plan." .
+apeval:stepSkill a owl:ObjectProperty ; rdfs:domain apeval:ExpectedPlanStep ; rdfs:label "stepSkill" ; rdfs:comment "The skill (capability) the step runs." .
+apeval:expectedMissing a owl:ObjectProperty ; rdfs:domain apeval:ReplayIntent ; rdfs:label "expectedMissing" ;
+  rdfs:comment "A domain class the right plan must name as missing — to be asked for, not invented." .
+apeval:exercisesClass a owl:ObjectProperty ; rdfs:domain apeval:ReplayIntent ; rdfs:label "exercisesClass" ;
+  rdfs:comment "A domain class the case's words or gold exercise — what the case tests of the ontology." .
+apeval:adjudication a owl:DatatypeProperty ; rdfs:domain apeval:ReplayIntent ; rdfs:range xsd:string ; rdfs:label "adjudication" ;
+  rdfs:comment "hypothesis (authored, pre-registered) | adjudicated (a named annotator agreed)." .
+apeval:message a owl:DatatypeProperty ; rdfs:domain apeval:ReplayIntent ; rdfs:range xsd:string ; rdfs:label "message" ;
+  rdfs:comment "The request's words — carried in a DOMAIN's test projection (authored sentences, never a person's), never in a run's graph." .
+apeval:inReplaySet a owl:ObjectProperty ; rdfs:domain apeval:ReplayIntent ; rdfs:range apeval:ReplaySet ; owl:inverseOf apeval:hasIntent ; rdfs:label "inReplaySet" .
+apeval:underStartingState a owl:ObjectProperty ; rdfs:domain apeval:ReplayIntent ; rdfs:range apexec:StartingState ; rdfs:label "underStartingState" ;
+  rdfs:comment "The seeded state the case is asked from (the asker's history, memory, held records)." .
+apeval:holdsClass a owl:ObjectProperty ; rdfs:domain apexec:StartingState ; rdfs:label "holdsClass" ;
+  rdfs:comment "A domain class the asker already holds a record of in this state." .
+apeval:recentSkill a owl:ObjectProperty ; rdfs:domain apexec:StartingState ; rdfs:label "recentSkill" ;
+  rdfs:comment "A skill the asker used recently in this state." .
+apeval:memoryTag a owl:DatatypeProperty ; rdfs:domain apexec:StartingState ; rdfs:range xsd:string ; rdfs:label "memoryTag" ;
+  rdfs:comment "What the asker's own agent remembers about them in this state (a short tag)." .
+apeval:latestResult a owl:ObjectProperty ; rdfs:domain apeval:ReplayIntent ; rdfs:range earl:Assertion ; rdfs:label "latestResult" ;
+  rdfs:comment "The newest judged run of the case (an EARL assertion with its selection outcome, time and tokens)." .
+# Ranges (spec 414 §2 rule 2 — every term reaches a standard one): a skill is a prov:Entity (its SKILL.md), a class is an rdfs:Class.
+apeval:expectedSkill rdfs:range prov:Entity . apeval:acceptableSkill rdfs:range prov:Entity . apeval:forbiddenSkill rdfs:range prov:Entity .
+apeval:collidesWith rdfs:range prov:Entity . apeval:optionSkill rdfs:range prov:Entity . apeval:stepSkill rdfs:range prov:Entity . apeval:recentSkill rdfs:range prov:Entity .
+apeval:expectedMissing rdfs:range rdfs:Class . apeval:exercisesClass rdfs:range rdfs:Class . apeval:holdsClass rdfs:range rdfs:Class .
+apeval:expectedMissing rdfs:subPropertyOf dcterms:subject . apeval:exercisesClass rdfs:subPropertyOf dcterms:subject . apeval:holdsClass rdfs:subPropertyOf dcterms:subject .
+
+# ═══ Spec 438 W3 — an ASSESSMENT is a separate signed assertion ════════════════════════════════════
+#
+# Spec 414 §3.4 fixed the SHAPE of a judgment and deferred the record to "a later spec". This is it. The thing
+# that makes an assessment safe is what it does NOT do: it never edits the run record it is about. A correction
+# is a NEW assessment that supersedes the old one, so the trace of what happened and the opinions about it stay
+# separable forever (J6). J1 still holds — no verifier, admission stage or mandate reads one.
+apeval:Assessment a owl:Class ;
+  rdfs:subClassOf apeval:Judgment ;
+  rdfs:label "Assessment" ;
+  rdfs:comment "A signed judgment of one run against one criteria revision, bound to the subject by digest. Stored beside the run, NEVER inside it; superseded, never edited. Evidence, never authority (spec 438 §5, 414 J1/J6)." .
+
+apeval:supersedes a owl:ObjectProperty ;
+  rdfs:domain apeval:Assessment ; rdfs:range apeval:Assessment ;
+  rdfs:label "supersedes" ; rdfs:comment "The earlier assessment this one replaces. The earlier one remains readable — a correction is an addition." .
+
+apeval:criteriaRevision a owl:DatatypeProperty ;
+  rdfs:domain apeval:Assessment ; rdfs:range xsd:integer ;
+  rdfs:label "criteriaRevision" ;
+  rdfs:comment "The OutcomeSpecification revision judged against. MUST pre-date the subject (J2): no judging against a post-hoc expectation." .
+
+apeval:evaluatorImplementation a owl:DatatypeProperty ;
+  rdfs:domain apeval:Judge ; rdfs:range xsd:string ;
+  rdfs:label "evaluatorImplementation" ;
+  rdfs:comment "Digest of the evaluator's implementation (mechanical) or of its prompt (model), so J5 reproducibility can be claimed or refused honestly." .
+
+
+# ═════════════════════════════════════════════════════════════════════════════════════════════════
+# 2026-10-14 — EVERY NUMBER INTERPRETABLE, AND THE OUTCOME GROUNDED ("evals article gap" W1, W2, W4, W6).
+#
+# W1. A test result that is NOT a verdict says WHY — the agent, the grader, or the environment — so an estate outage is
+# never read as a miss; an assertion says which TRIAL (repeat) it is, which KIND of grader asserted it, and (EARL's own
+# terms) its mode and who asserted it. W2. A case may carry a REFERENCE ANSWER: a solvability proof the author wrote,
+# held to 1.0 under the graders, never a target to diff against. W4. A replay set says what KIND of suite it is —
+# capability or regression — so the two are never pooled. W6. The outcome gold is a dul:Description; what a run
+# delivered is a situation that satisfies it or does not — the same spine the substrate's own situations stand on.
+# ═════════════════════════════════════════════════════════════════════════════════════════════════
+
+# ── W1: the failure cause ──
+apeval:FailureCause a owl:Class ;
+  rdfs:subClassOf skos:Concept ;
+  rdfs:label "FailureCause" ;
+  rdfs:comment "Why a test result is not a verdict: agent (the subject's own failure, in its denominator) · grader (the judge failed — the instrument, not the subject) · environment (the estate failed — a store, a provider, a timeout, a trace that never landed). Only `agent` stays in the subject's denominator; the others are counted beside it." .
+apeval:FailureCauses a skos:ConceptScheme ; rdfs:label "Failure causes" .
+apeval:fc-agent a apeval:FailureCause ; skos:inScheme apeval:FailureCauses ; skos:notation "agent" ; skos:prefLabel "the agent" ; skos:definition "The subject produced no verdict for a reason the harness owns. Counted against it." .
+apeval:fc-grader a apeval:FailureCause ; skos:inScheme apeval:FailureCauses ; skos:notation "grader" ; skos:prefLabel "the grader" ; skos:definition "The judge failed: timed out, errored, answered unreadably. The instrument, not the subject; never counted against it." .
+apeval:fc-environment a apeval:FailureCause ; skos:inScheme apeval:FailureCauses ; skos:notation "environment" ; skos:prefLabel "the environment" ; skos:definition "The estate failed: a store, a provider, a timeout, a trace that never landed, a repeat refused for isolation. Never counted against the subject." .
+apeval:failureCause a owl:ObjectProperty ; rdfs:domain earl:TestResult ; rdfs:range apeval:FailureCause ; rdfs:label "failureCause" ;
+  rdfs:comment "On a result whose earl:outcome is earl:cantTell (or on a run with no result): why. Absent on a pass or a fail." .
+apeval:trialIndex a owl:DatatypeProperty ; rdfs:domain earl:Assertion ; rdfs:range xsd:integer ; rdfs:label "trialIndex" ;
+  rdfs:comment "Which repeat of the case this assertion judged (0 = the first; an experiment's Stage-1 run). pass@1 reads trial 0; pass^k reads every trial." .
+apeval:graderKind a owl:DatatypeProperty ; rdfs:domain earl:Assertion ; rdfs:range xsd:string ; rdfs:label "graderKind" ;
+  rdfs:comment "The kind of grader that asserted it — mechanical | model | human — beside earl:assertedBy (the grader itself) and earl:mode (automatic for mechanical and model graders; manual for an adjudication by a person). Verdicts of different kinds are never pooled (J4)." .
+
+# ── W2: the reference answer ──
+apeval:referenceAnswer a owl:DatatypeProperty ; rdfs:domain apeval:ReplayIntent ; rdfs:range xsd:string ; rdfs:label "referenceAnswer" ;
+  rdfs:comment "An answer the case's author wrote that SHOULD score 1.0 under the graders — a proof the case is solvable and the graders can see it, checked before any run. Never a target to diff an agent's answer against. Carried in the domain's projection as authored text; a run's graph never carries it." .
+apeval:referenceScore a owl:DatatypeProperty ; rdfs:domain apeval:ReplayIntent ; rdfs:range xsd:decimal ; rdfs:label "referenceScore" ;
+  rdfs:comment "What the graders gave the reference answer when it was last checked (1.0 is the bar; below it the case or the grader is wrong, and the build says which)." .
+
+# ── W4: the suite kind ──
+apeval:SuiteKind a owl:Class ;
+  rdfs:subClassOf skos:Concept ;
+  rdfs:label "SuiteKind" ;
+  rdfs:comment "What a replay set is for. capability: a fresh, hard set written to find out what the harness can do — its number is a capability number. regression: a saturated set kept to catch a step backwards — a pass is expected and never published as a capability number. A capability set that passes three consecutive runs at ≥ 95% graduates to regression." .
+apeval:SuiteKinds a skos:ConceptScheme ; rdfs:label "Suite kinds" .
+apeval:suite-capability a apeval:SuiteKind ; skos:inScheme apeval:SuiteKinds ; skos:notation "capability" ; skos:prefLabel "capability suite" .
+apeval:suite-regression a apeval:SuiteKind ; skos:inScheme apeval:SuiteKinds ; skos:notation "regression" ; skos:prefLabel "regression suite" .
+apeval:suiteKind a owl:ObjectProperty ; rdfs:domain apeval:ReplaySet ; rdfs:range apeval:SuiteKind ; rdfs:label "suiteKind" .
+
+# ── W6: the outcome gold as a description; what was delivered as a situation ──
+apeval:OutcomeDescription a owl:Class ;
+  rdfs:subClassOf dul:Description , prov:Entity ;
+  rdfs:label "OutcomeDescription" ;
+  rdfs:comment "What a case EXPECTS the world to look like after the run, as a dul:Description: the classes whose records should exist (the gold plan's products), the inputs that should be named missing rather than invented, or — for a clarify case — that the person should be asked which option. A run's delivered situation satisfies it or does not. The compiled `behaviour` string (route · decline · clarify · several · plan) is the description's compiled form, kept as is." .
+apeval:DeliveredSituation a owl:Class ;
+  rdfs:subClassOf dul:Situation , prov:Entity ;
+  rdfs:label "DeliveredSituation" ;
+  rdfs:comment "What one run actually left: the records it produced (each a member of the situation), the inputs it named missing, the question it asked. Judged against the case's OutcomeDescription: satisfied (earl:passed) · not satisfied (earl:failed) · could not tell (earl:cantTell). A clarify case is a situation under TWO candidate descriptions; a decline case a situation no description satisfies." .
+apeval:expectedOutcome a owl:ObjectProperty ; rdfs:domain apeval:ReplayIntent ; rdfs:range apeval:OutcomeDescription ; rdfs:label "expectedOutcome" ;
+  rdfs:comment "The case's outcome gold, as a description. Pre-registered by digest with the set." .
+apeval:candidateOutcome a owl:ObjectProperty ; rdfs:domain apeval:ReplayIntent ; rdfs:range apeval:OutcomeDescription ; rdfs:label "candidateOutcome" ;
+  rdfs:comment "A clarify or several case: one of the descriptions the ask could reach; the right first move is to ask which (clarify) or to reach each (several)." .
+apeval:describesClass a owl:ObjectProperty ; rdfs:domain apeval:OutcomeDescription ; rdfs:range rdfs:Class ; rdfs:subPropertyOf dcterms:subject ; rdfs:label "describesClass" ;
+  rdfs:comment "A class whose record the description expects to exist after the run (a gold-plan skill's product, honouring `within` and alternative groups)." .
+apeval:describesMissing a owl:ObjectProperty ; rdfs:domain apeval:OutcomeDescription ; rdfs:range rdfs:Class ; rdfs:subPropertyOf dcterms:subject ; rdfs:label "describesMissing" ;
+  rdfs:comment "A class the description expects to be NAMED AS MISSING — asked for, never invented." .
+apeval:describesAsk a owl:DatatypeProperty ; rdfs:domain apeval:OutcomeDescription ; rdfs:range xsd:boolean ; rdfs:label "describesAsk" ;
+  rdfs:comment "true when the description expects the run to ask the person which option they mean (a clarify case)." .
+apeval:deliveredSituation a owl:ObjectProperty ; rdfs:domain earl:TestResult ; rdfs:range apeval:DeliveredSituation ; rdfs:label "deliveredSituation" ;
+  rdfs:comment "What the judged run left, as the situation the verdict is about." .
+apeval:satisfies a owl:ObjectProperty ; rdfs:domain apeval:DeliveredSituation ; rdfs:range apeval:OutcomeDescription ; rdfs:subPropertyOf dul:satisfies ; rdfs:label "satisfies" ;
+  rdfs:comment "The situation satisfies the description (every expected class exists and validates; every missing input was named; the ask was asked). The state check decides existence and validity mechanically; the judge reads the rest." .
+apeval:reachedUnder a owl:ObjectProperty ; rdfs:domain apeval:DeliveredSituation ; rdfs:range apeval:OutcomeDescription ; rdfs:label "reachedUnder" ;
+  rdfs:comment "The description the run's OWN plan was reaching for (the agent-plan expectation). When it differs from the case's expectedOutcome the run reached a situation under the WRONG description — SPARQL can list those cases." .