Compo 4.03 — Mass ratios, ordering и relations

Сравниваем массу, не композицию

00 / module api

4.03 отвечает на вопросы «во сколько раз», «кто тяжелее», «насколько близки» и «какая доля массы в наборе». Он не решает, хорошо ли распределены массы.

SCOPE

Pair ratios, log-ratios, relative gaps, ordering, equivalence bands и mass shares.

CONSUMES

Уже готовые comparable mass magnitudes, предпочтительно из 4.02.

FINALITY

VISUAL MASS RELATIONS ONLY.

RELATION ≠ BALANCE. `A mass = 2×B` — локальный relational fact. Сбалансирована ли сцена с таким отношением — вопрос будущего global balance owner.

ComparableMassRef

01 / comparable mass

Сравнение допустимо только внутри одной mass calibration/stage/domain системы.

ComparableMassRef

{
  objectRef,
  massRef,
  massValue,

  stage:
    "INTRINSIC" |
    "CROSS_OBJECT_CONTEXTUALIZED",

  calibrationDomainRef,
  measurementDomainRef,

  confidence,
  provenance
}

Comparable iff

same calibrationDomainRef
same calibrationProfileRef
same measurementDomainRef
same stage
massValue finite >= 0

unless:
  an explicit external
  normalization adapter
  has already reconciled
  the domains.

PREFERRED STAGE. Для cross-object composition comparisons предпочтителен `CROSS_OBJECT_CONTEXTUALIZED` из 4.02. Intrinsic comparison разрешается только как явно выбранный режим.

PairMassRelation

02 / pair relations

Pair relation хранит несколько разных представлений одной связи; они не должны независимо голосовать в 4.04 как четыре «разных сигнала».

PairMassRelation {
  aRef,
  bRef,

  ratio?,
  logRatio,
  relativeGap,
  pairOrder,
  equivalenceClass,

  calibrationDomainRef,
  measurementDomainRef,
  stage,

  confidence,
  status,
  provenance
}

DERIVED FAMILY. `ratio`, `logRatio`, `relativeGap` и `pairOrder` происходят из одной пары mass values. Это relational descriptors одного evidence family, а не четыре независимых scoring votes.

Ratio, log-ratio, relative gap

03 / pair math

Log-ratio удобен как симметричный signed descriptor: swapping A/B просто меняет знак.

Ratio

r = A / B swap: r(B,A)=1/r B≈0: finite JSON ratio=N/A order carries direction

Log ratio

ℓ = ln(A/B) when A,B > ε zero involved: logRatio = N/A swap: ℓ(B,A) = -ℓ(A,B)

Relative gap

g = (A-B) / max(A,B,ε) g ∈ [-1,1] swap: sign flips

Order и equivalence band

04 / ordering

Численные различия меньше tolerance можно считать одной perceptual mass band, но tolerance остаётся эвристикой/калибровкой.

equivalent(A,B) iff |logRatio(A,B)| ≤ equivalenceLogEpsilon Then: pairOrder = EQUIVALENT otherwise: A_GT_B or A_LT_B Special: A≈0 and B≈0 → ZERO_ZERO Important: EQUIVALENT ≠ exactly equal EQUIVALENT ≠ interchangeable in design

TOLERANCE PROVENANCE. Reference `0.08` — `INITIAL_HEURISTIC`, не универсальный psychophysics threshold.

MassOrderingResult

06 / set ordering

Set ordering использует mass magnitude; equivalence groups сохраняют near-ties, а lexical tie-break нужен только для deterministic serialization.

MassOrderingResult {
  comparisonSetRef,

  orderedObjectRefs[],
  tieGroups[],

  massShares[],

  stage,
  calibrationDomainRef,
  measurementDomainRef,

  confidence,
  status,
  provenance
}

Sort: higher massValue first Within same equivalence band: stable objectRef lexical order Lexical order is: deterministic implementation detail NOT perceptual superiority

Global scale invariance

07 / invariance

Если все массы в comparison set умножить на одинаковый положительный коэффициент, relations должны остаться теми же.

For k > 0: A' = kA B' = kB A'/B' = A/B log(A'/B') = log(A/B) relativeGap(A',B') = relativeGap(A,B) share_i(kM) = share_i(M) order(kM) = order(M)

CALIBRATION-SAFE PROPERTY. Простое изменение общей единицы/scale mass model не должно менять relational structure.

Zero mass cases

08 / zero and N/A

Нулевая mass может быть валидной — например explicit alpha=0 в 4.01. Но ratio с нулевым denominator нельзя сериализовать как Infinity в canonical JSON.

A	B	Order	Canonical ratio
0	0	ZERO_ZERO	N/A
>0	0	A_GT_B	N/A; logRatio finite via ε descriptor
0	>0	A_LT_B	0
N/A	any	N_A	N/A

N/A ≠ ZERO. Missing/unreliable mass evidence не участвует в ordering как объект с нулевой массой.

Confidence

09 / relation confidence

Relation confidence зависит от evidence quality сравниваемых masses, а не от того, насколько ratio далёк от 1.

Reference policy: MIN: relationConfidence = min(conf_A, conf_B) GEOMETRIC_MEAN: relationConfidence = sqrt(conf_A × conf_B) For set: aggregate from members by declared confidence policy NOT: confidence = |logRatio| confidence = dominance gap

Mass dominance ≠ hierarchy

10 / semantic boundary

Это особенно важно перед будущими global темами.

MASS

Heaviest

Максимальная contextual mass в declared comparison set.

HIERARCHY

Most important

Semantic/visual hierarchy требует отдельного owner и дополнительных signals.

BALANCE

Balanced

Нужны positions, moments/distribution и scene structure, а не только ranked masses.

heaviest object MAY be focal MAY be secondary MAY be background block Mass relation alone cannot decide.

Relation correction, не mutation

11 / correction requests

4.03 может сформулировать desired relation, но не знает, какой параметр объекта нужно менять.

MassRelationCorrectionRequest {
  comparisonSetRef,
  objectRefs[],

  requestedRelation:
    "INCREASE_A_RELATIVE_TO_B" |
    "DECREASE_A_RELATIVE_TO_B" |
    "BRING_TO_EQUIVALENCE_BAND" |
    "SEPARATE_FROM_EQUIVALENCE_BAND",

  currentRelationRef,
  targetOwner: "DEFERRED / solver",
  constraints[],
  provenance
}

NO FOREIGN MUTATION. Чтобы изменить A/B, будущий solver может выбрать size, alpha, contrast, position или другой owner-controlled parameter. 4.03 сам ничего из этого не меняет.

Visual Lab

12 / visual lab

Relations показаны как факты внутри sets, без verdict «хорошо/плохо».

Reference relation tests

13 / deterministic tests

Проверяем ratios, symmetry, zero cases, scale invariance, comparability, shares и deterministic ordering.

Tests not run yet.

DATA FOR LAYOUT ENGINE

14 / machine contract

Machine-readable contract 4.03.

{
  "module": "Compo 4.03",
  "name": "mass_ratios_ordering_and_relations",
  "version": "1.1-corrective",
  "scope": "PAIR_AND_SET_RELATIONS_BETWEEN_COMPARABLE_VISUAL_MASS_MAGNITUDES",
  "consumes": [
    "Compo4.00.VisualMassDescriptor / MassEvidenceStage",
    "Compo4.02.CrossObjectContextMassResult as preferred comparison input",
    "Compo4.01.ShapeGeometricMassResult? for explicitly intrinsic-only comparisons",
    "ComparableMassRef[]",
    "MassRelationPolicy",
    "StylePreset.massRelationTolerance?"
  ],
  "provides": [
    "ComparableMassRef",
    "PairMassRelation",
    "MassOrderingResult",
    "MassEquivalenceBand",
    "MassShareDescriptor[]",
    "MassRelationSet",
    "MassRelationCorrectionRequest[]"
  ],
  "owns_metrics": [
    "visualMass.relation.logRatio",
    "visualMass.relation.ratio",
    "visualMass.relation.relativeGap",
    "visualMass.relation.massShare",
    "visualMass.relation.orderClass",
    "visualMass.relation.equivalenceClass",
    "visualMass.relation.confidence"
  ],
  "descriptors_only": [
    "visualMass.relation.pairOrder",
    "visualMass.relation.rankIndex",
    "visualMass.relation.tieGroup",
    "visualMass.relation.comparisonSetRef",
    "visualMass.relation.inputStage",
    "visualMass.relation.calibrationDomainRef",
    "visualMass.relation.measurementDomainRef",
    "visualMass.relation.status"
  ],
  "comparable_mass_schema": {
    "objectRef": "string",
    "massRef": "owner-qualified metric ref",
    "massValue": "finite >=0",
    "stage": "INTRINSIC|CROSS_OBJECT_CONTEXTUALIZED",
    "calibrationDomainRef": "required; all compared items must match",
    "measurementDomainRef": "required when source may be full/clipped/custom; all compared items must match",
    "confidence": "0..1",
    "provenance": "object",
    "calibrationProfileRef": "required profileId@version; all compared items must match"
  },
  "pair_relation_schema": {
    "aRef": "object ref",
    "bRef": "object ref",
    "ratio": "A/B; N/A when B≈0 unless both≈0",
    "logRatio": "ln((A+epsMass)/(B+epsMass)) for stable signed comparison descriptor",
    "relativeGap": "(A-B)/max(A,B,epsMass)",
    "pairOrder": "A_GT_B|A_LT_B|EQUIVALENT|ZERO_ZERO|N_A",
    "equivalenceClass": "SAME_BAND|DIFFERENT_BAND|N_A",
    "confidence": "min/combined source evidence confidence",
    "status": "OK|PARTIAL|N_A|INVALID",
    "calibrationProfileRef": "uniform profileId@version"
  },
  "set_relation_schema": {
    "comparisonSetRef": "string",
    "orderedObjectRefs": "descending mass; stable lexical tie-break inside equivalence band",
    "massShares": "mass_i / sum(mass); N/A when total≈0",
    "tieGroups": "equivalence-band groups",
    "pairRelations": "optional bounded pair relations",
    "stage": "uniform across set",
    "calibrationDomainRef": "uniform across set",
    "measurementDomainRef": "uniform across set",
    "confidence": "aggregate evidence confidence",
    "status": "OK|PARTIAL|N_A|INVALID",
    "calibrationProfileRef": "uniform profileId@version"
  },
  "relation_math": {
    "ratio": "A/B",
    "logRatio": "ln(A/B) when both A and B are above massZeroEpsilon; otherwise N/A except ZERO_ZERO state",
    "relativeGap": "(A-B)/max(A,B,epsMass)",
    "massShare": "A/sum(set)",
    "order_by": "massValue descending",
    "equivalence_band": "|logRatio| <= equivalenceLogEpsilon",
    "scale_invariance": "for positive masses that remain above the declared zero threshold, multiplying every compared mass by the same positive k leaves ratio/logRatio/relativeGap/share/order unchanged"
  },
  "policy_schema": {
    "equivalenceLogEpsilon": "positive finite tolerance",
    "massZeroEpsilon": "nonnegative finite numerical threshold",
    "allowIntrinsicComparison": "boolean",
    "requireSameStage": "true by default",
    "requireSameCalibrationDomain": "true",
    "requireSameMeasurementDomain": "true",
    "maxPairRelations": "bounded implementation budget",
    "confidencePolicy": "MIN|GEOMETRIC_MEAN"
  },
  "semantics": {
    "ratio": "magnitude relation only",
    "order": "visual mass ordering only",
    "equivalent": "within declared tolerance, not geometrically identical",
    "massShare": "share of mass inside the declared comparison set only; not scene balance",
    "dominant_mass": "if used descriptively, means highest mass in set only; not semantic hierarchy or focal priority",
    "tie_break": "stable lexical tie-break is implementation ordering, not perceptual superiority"
  },
  "hard_constraints": [
    "4.03 consumes mass magnitudes and must not reapply contrast/isolation/position/overlap",
    "4.03 does not recompute 4.01 area/density or 4.02 contextual mass",
    "compared masses must share calibrationDomainRef",
    "compared masses must share measurementDomainRef unless an explicit external normalization adapter has already reconciled them",
    "mixed INTRINSIC and CROSS_OBJECT_CONTEXTUALIZED stages are incomparable by default",
    "ratio/order/share are descriptors, not aesthetic scores",
    "massShare is relative only to declared comparison set and is not global balance",
    "highest visual mass does not imply semantic hierarchy, importance or focal correctness",
    "equivalence tolerance has provenance and is not a universal perceptual law",
    "zero-vs-zero pair ratio is N/A with ZERO_ZERO order state, not ratio=1 by fiat",
    "zero denominator with positive numerator yields ratio=+INF descriptor only if representation supports it; canonical JSON result should use ratio=null plus order/logRatio evidence rather than non-finite JSON",
    "N/A != 0",
    "negative space remains Topic5",
    "global balance/hierarchy/eye-flow remain future owners",
    "target-aware mass scoring remains Compo4.04",
    "no beautyScore",
    "compared masses must share calibrationProfileRef/profile version, not only canonicalDomainRef",
    "positive-mass logRatio uses ln(A/B) without additive epsilon; zero-involved logRatio is N/A to preserve honest scale semantics",
    "scale invariance is claimed only while zero/nonzero classification remains unchanged under the declared numerical threshold"
  ],
  "initial_heuristics": {
    "equivalence_log_epsilon": {
      "value": 0.08,
      "provenance": "INITIAL_HEURISTIC"
    },
    "mass_zero_epsilon": {
      "value": 1e-12,
      "provenance": "IMPLEMENTATION_HEURISTIC"
    },
    "max_pair_relations": {
      "value": 256,
      "provenance": "IMPLEMENTATION_HEURISTIC"
    }
  },
  "correction_policy": {
    "foreign_mass_mutation": "FORBIDDEN",
    "incomparable_calibration_domain": "request 4.04 calibration adapter/profile",
    "mixed_stage": "request upstream stage alignment via 4.00/4.02",
    "mixed_measurement_domain": "request source-domain alignment before comparison",
    "mass_value_issue": "route to 4.01/4.02/source local-weight owner",
    "desired_mass_relation_change": "emit relation correction request; actual object/color/spacing/transform change belongs to relevant owner",
    "balance_issue": "route to future balance/global owner",
    "hierarchy_issue": "route to future hierarchy/semantic owner"
  },
  "deferred_inputs": [
    "Compo4.04 calibration/target-aware scoring/regression",
    "Topic5 positive/negative space",
    "future global balance",
    "future hierarchy/focal-point owner",
    "future scene/group selection policies"
  ],
  "recompute_when": [
    "any compared massValue changes",
    "input stage changes",
    "calibrationDomainRef changes",
    "measurementDomainRef changes",
    "comparison set membership changes",
    "equivalence tolerance changes",
    "confidence policy changes"
  ],
  "finality": "VISUAL_MASS_RELATIONS_ONLY"
}

ARCHITECTURE CHECKPOINT · 4.03 OWNS ONLY PAIR/SET RELATIONS BETWEEN COMPARABLE MASS MAGNITUDES: RATIO, LOG-RATIO, RELATIVE GAP, EQUIVALENCE BANDS, ORDERING AND SET-RELATIVE MASS SHARE. IT DOES NOT RECOMPUTE 4.01 AREA/DENSITY OR 4.02 CONTEXTUAL MASS. MIXED CALIBRATION/STAGE/MEASUREMENT DOMAINS ARE INCOMPARABLE BY DEFAULT. MASS ORDER ≠ SEMANTIC HIERARCHY. MASS SHARE ≠ GLOBAL BALANCE. 4.04 OWNS TARGET-AWARE CALIBRATION/SCORING. TOPIC5 OWNS FREE SPACE. NO BEAUTYSCORE.

Сравниваем массу, не композицию

ComparableMassRef

ComparableMassRef

Comparable iff

PairMassRelation

Ratio, log-ratio, relative gap

Ratio

Log ratio

Relative gap

Order и equivalence band

Mass share внутри comparison set

MassOrderingResult

Global scale invariance

Zero mass cases

Confidence

Mass dominance ≠ hierarchy

Heaviest

Most important

Balanced

Relation correction, не mutation

Visual Lab

Reference relation tests

DATA FOR LAYOUT ENGINE