Скоринг
линейной
системы.

CONSUMES / PROVIDES

LineStructureCandidate[]              // from 2.19
MetricResult[]                    // owner results 2.02–2.18
PenaltyResult[]?
Candidate.hardViolations[]
Candidate.auditEvidence[]
Candidate.correctionCost?
Candidate.lineage
LineSystemTarget                  // semantic + style
LineScoringPolicy
Compo0.seed

LineSystemScoreResult {
  status,
  rankEligible,
  localScore,
  rawPositiveScore,
  softPenalty,
  confidence,
  evidenceCoverage,
  targetCoverage,
  groupBreakdown[],
  selectedVotes[],
  suppressedVotes[],
  appliedPenalties[],
  suppressedPenalties[],
  rejectedBy[],
  unresolvedInputs[],
  repairHints[],
  dirtyOwners[],
  comparisonKey,
  audit
}

Главное правило

Никакого beautyScore. Локальный scalar нужен только для выбора между line candidates в одинаковом explicit target context. Он не утверждает, что один постер «красивее» другого.

Canonical owner inputs

Composite-first policy: если owner уже экспортирует локальный composite, scorer использует его как основной голос, а дочерние компоненты сохраняет в diagnostics. Это канонический механизм защиты от double counting.

MetricResult {
  key: string,
  owner: "Compo2.xx",
  value: float,                  // 0..1; higher = better fit to owner target
  applicability: float,          // 0..1
  confidence: float,             // 0..1
  status: "ok|warn|fail|na",
  provenance:
    "DERIVED|STANDARD|EMPIRICAL|INITIAL_HEURISTIC|
     IMPLEMENTATION_HEURISTIC|STYLE_PRESET|DEFERRED",

  composite: bool,
  componentOf: string|null,
  dependsOn: string[],
  derivedFrom: string[],
  scoringGroup: string|null,
  independentVoteAllowed: bool,

  votePriority: number?,
  evidence: object,
  repairHints: []
}

PenaltyResult {
  key: string,
  owner: "Compo2.xx",
  severity: float,               // 0..1
  applicability: float,
  confidence: float,
  status: "ok|warn|fail|na",
  provenance: string,

  componentOf: string|null,
  dependsOn: string[],
  derivedFrom: string[],
  scoringGroup: string|null,
  independentVoteAllowed: bool,

  evidence: object,
  repairHints: []
}

Metadata semantics

Validation / legacy adapter

FEASIBILITY GATE

1. N/A / applicability = 0
2. confidence = 0
3. invalid / incomplete metadata
4. componentOf active composite
5. duplicate owner+key+version
6. same scoringGroup → primary vote
   additional sibling only if independentVoteAllowed
7. cross-group derivedFrom / evidence overlap
   → suppress lower-priority vote unless BOTH explicitly allow independence
8. dependsOn alone → NEVER suppress
9. target says group not applicable
10. global / deferred metric → ignored by 2.20

{
  key,
  owner,
  reason:
    "COMPONENT_OF_COMPOSITE" |
    "GROUP_PRIMARY_SELECTED" |
    "CROSS_GROUP_DERIVED_DUPLICATE" |
    "DUPLICATE_EVIDENCE" |
    "ZERO_APPLICABILITY" |
    "ZERO_CONFIDENCE" |
    "INCOMPLETE_METADATA" |
    "TARGET_NOT_APPLICABLE" |
    "GLOBAL_DEFERRED",
  suppressedBy?: string,
  evidenceRef?: string
}

Cross-group rule

Penalty de-duplication

Penalty проходит ту же lineage-проверку, что и positive vote. Одинаковый evidenceFingerprint не может быть списан дважды даже в разных scoringGroups. Если тот же evidence уже представлен active composite vote, penalty подавляется, пока owner явно не выставит independentVoteAllowed=true с обеих сторон.

Missing required group не получает value=0. При requiredEvidencePolicy="require_all" candidate становится insufficient_evidence и не ранжируется. При rank_with_low_confidence он может остаться в ranking, но с пониженным coverage/confidence.

Глобальный solver позже имеет право выбрать локально второй candidate, если он лучше работает с hierarchy, balance, negative space, typography, image interaction или global eye-flow. 2.20 не пытается предвосхитить этот слой.

Edge cases

contributionLoss_i =
  normalizedEffectiveWeight_i × (1 - value_i)

lowestPrimaryVotes =
  sort(selectedVotes by contributionLoss desc)
  take first 3

repairHints = mergeUnique(
  hardViolations.repairHints,
  lowestPrimaryVotes.repairHints,
  appliedPenalties.repairHints
)

dirtyOwners = unique(repairHints.targetOwner)

Candidate generation boundary

Live mini-scorer

{
  "module": "Compo 2.20",
  "name": "line_system_scoring",
  "version": "1.2-corrective",
  "scope": "local scoring and deterministic ranking of Compo 2.19 line-system candidates under one explicit semantic/style target",
  "consumes": [
    "Compo2.19.LineStructureCandidate[]",
    "MetricResult[] from Compo2.02-Compo2.18",
    "PenaltyResult[] from metric owners",
    "Candidate.hardViolations[]",
    "Candidate.auditEvidence[]",
    "Candidate.correctionCost?",
    "Candidate.lineage",
    "LineSystemTarget.semanticTarget",
    "LineSystemTarget.styleTarget",
    "LineScoringPolicy",
    "Compo0.seed"
  ],
  "provides": [
    "LineSystemScoreResult",
    "lineSystem.local_score",
    "lineSystem.local_score_confidence",
    "lineSystem.local_rank_key",
    "lineSystem.suppression_log",
    "repairHints[]",
    "dirtyOwners[]"
  ],
  "owns_metrics": [
    "lineSystem.local_score",
    "lineSystem.local_score_confidence",
    "lineSystem.local_rank_key"
  ],
  "descriptors_only": [
    "suppression_log",
    "group_breakdown",
    "evidenceCoverage",
    "targetCoverage",
    "comparisonKey",
    "audit"
  ],
  "deferred_inputs": [
    "global.balance",
    "global.hierarchy",
    "global.negativeSpace",
    "global.eyeFlow",
    "typography.*",
    "imageCrop.*",
    "posterGlobalScore",
    "Compo2.23 context overrides"
  ],
  "hard_constraints": [
    "hard violation rejects before score",
    "foreign mutation forbidden",
    "raw geometry descriptors cannot vote",
    "normalization belongs to metric owner",
    "N/A excluded from vote and denominator",
    "global metrics cannot influence localScore"
  ],
  "soft_constraints": [
    "owner penalty requires evidence",
    "same evidence may not be penalized twice",
    "group influence is capped",
    "incomplete metadata becomes diagnostic-only"
  ],
  "preferences": [
    "composite-first",
    "explicit semantic/style target",
    "higher-confidence evidence has greater voting influence",
    "lower correction cost wins only after score/coverage/confidence tie rules"
  ],
  "overrides_allowed_by": [
    "Compo2.23 context override layer",
    "future global composition solver",
    "explicit StylePreset / semantic target for weights and applicability only"
  ],
  "initial_heuristics": {
    "groupVoteCap": 1.0,
    "maxSoftPenalty": 0.35,
    "scoreEpsilon": 0.015,
    "confidenceEpsilon": 0.03
  },
  "style_preset_values": {
    "groupVoteCaps": "optional explicit per-group",
    "semanticMultipliers": "optional explicit",
    "styleMultipliers": "optional explicit",
    "thresholds": "must carry STYLE_PRESET provenance"
  },
  "metric_policy": {
    "range": [
      0,
      1
    ],
    "higher_is_better": "owner target fit only, not universal beauty",
    "normalization_owner": "metric_owner",
    "composite_first": true,
    "exclude_na_from_denominator": true,
    "exclude_zero_confidence": true,
    "required_dependency_metadata": [
      "componentOf",
      "dependsOn",
      "derivedFrom",
      "scoringGroup",
      "independentVoteAllowed"
    ],
    "incomplete_metadata": "diagnostic_only",
    "double_counting": "check within scoringGroup and across scoringGroups"
  },
  "dependency_semantics": {
    "componentOf": "suppress component when parent composite votes",
    "derivedFrom": "lineage evidence used for duplicate suppression",
    "dependsOn": "recompute/order dependency only; does not suppress by itself",
    "scoringGroup": "vote-budget family",
    "independentVoteAllowed": "shared evidence may vote independently only when both overlapping metrics explicitly opt in"
  },
  "legacy_metadata_adapter": {
    "version": "line-owner-primary-bridge-v2",
    "scope": "Compo2.02-Compo2.14 only; explicit primary allowlist; unknown legacy owner metrics are diagnostic-only",
    "default_independentVoteAllowed": false,
    "new_owners_2_15_2_18": "must export explicit metadata"
  },
  "group_policy": {
    "groupWeight": "min(sum(scoreWeight), groupVoteCap)",
    "additional_same_group_vote": "requires explicit independentVoteAllowed on both representations; shared lineage is allowed only by explicit bilateral opt-in; group cap still bounds total influence",
    "group_cap_provenance": "INITIAL_HEURISTIC or STYLE_PRESET"
  },
  "aggregation": {
    "policyWeight": "baseWeight * semanticMultiplier * styleMultiplier * applicability",
    "scoreWeight": "policyWeight * confidence",
    "groupValue": "weightedMean(metric.value by scoreWeight)",
    "positiveScore": "weightedMean(groupValue by capped groupWeight)",
    "softPenalty": "clamp(sum(deduplicated owner penalty contributions),0,maxSoftPenalty)",
    "localScore": "clamp(positiveScore-softPenalty,0,1)"
  },
  "confidence": {
    "groupConfidence": "weightedMean(metric.confidence by policyWeight without confidence factor)",
    "sceneConfidence": "weightedMean(groupConfidence by active groupWeight)",
    "evidenceCoverage": "presentExpectedGroups / expectedApplicableGroups",
    "targetCoverage": "presentRequiredGroups / requiredGroups",
    "finalConfidence": "sceneConfidence * evidenceCoverage",
    "missing_required_group": "never inject value=0"
  },
  "penalty_policy": {
    "deduplicate_across_groups": true,
    "deduplicate_against_selected_votes": true,
    "independent_exception": "explicit independentVoteAllowed on both representations"
  },
  "hard_policy": {
    "any_owner_hard_violation": "REJECT",
    "hard_violation_compensable_by_score": false,
    "scorer_may_invent_hard_violation": false
  },
  "ranking": [
    "feasibility",
    "localScore outside scoreEpsilon",
    "targetCoverage",
    "confidence outside confidenceEpsilon",
    "lower correctionCost",
    "stable hash(candidateId|seed|lineageHash)"
  ],
  "correction_policy": {
    "direct_foreign_mutation": "FORBIDDEN",
    "repair_source": "owner-issued repairHints / correction requests",
    "orchestrator": "Compo 2.19",
    "recompute_when": "dirty owner dependency graph indicates upstream evidence changed",
    "recompute": "dirty owners only, then rescore"
  },
  "tests": [
    "T01_NA",
    "T02_HARD_REJECT",
    "T03_COMPOSITE",
    "T04_GROUP_CAP",
    "T05_CROSS_GROUP_DERIVED",
    "T06_INDEPENDENT_VOTE",
    "T07_DEPENDSON_ONLY",
    "T08_TARGET_APPLICABILITY",
    "T09_NO_GLOBAL_LEAKAGE",
    "T10_CONFIDENCE_TIE",
    "T11_DETERMINISTIC_TIE",
    "T12_REJECTED_TIE",
    "T13_PENALTY_DEDUP",
    "T14_PENALTY_VOTE_DEDUP",
    "T15_MISSING_METADATA",
    "T16_LEGACY_ADAPTER",
    "T17_EVIDENCE_COVERAGE",
    "T18_DESCRIPTOR_NOT_SCORE",
    "T19_EXPLICIT_TARGET_GATE"
  ],
  "recompute_when": [
    "candidate geometry/relations changed by owner correction",
    "semantic/style target changed",
    "owner MetricResult/PenaltyResult changed",
    "scoring policy or legacy adapter version changed"
  ],
  "finality": "LOCAL_LINE_SYSTEM_RANKING_ONLY",
  "invariants": [
    "ONE_METRIC_ONE_OWNER",
    "N_A_NOT_ZERO",
    "LOCAL_NOT_GLOBAL",
    "GEOMETRY_NOT_INTERPRETATION",
    "DESCRIPTOR_NOT_SCORE",
    "COMPOSITE_FIRST",
    "NO_BEAUTYSCORE",
    "NO_FOREIGN_MUTATION"
  ]
}