@cite{ronderos-etal-2024}

@cite{sedivy-etal-1999} @cite{kursat-degen-2021} @cite{giles-etal-2026} @cite{aparicio-xiang-kennedy-2015} @cite{aparicio-2017}

Perceptual, Semantic, and Pragmatic Factors Affect the Derivation of Contrastive Inferences. Open Mind: Discoveries in Cognitive Science 8, 1213–1227.

Empirical Phenomenon

Cross-linguistic visual-world eye-tracking (English, Hindi, Hungarian) crossing a same-category contrast manipulation with three adjective types (color, scalar, material). The paper reports two qualitatively distinct findings, formalised here as two predicates:

1. Contrast effect (target-advantage analysis, §3 Results, ¶1–2): the contrastive inference effect — reduced cross-category competitor looks in the contrast condition — appears for color and scalar adjectives but is absent for material. The interaction term (helmert contrast: material vs. color+scalar) is significant.

2. Scalar baseline disadvantage (No-Contrast total looks, §3 Results, ¶4): in the no-contrast condition, fixations on target + competitor are lower for scalar adjectives than for either color (β = 0.25, p < 0.01) or material (β = 0.24, p < 0.05). No significant difference between color and material. Attributed to scalar adjectives requiring more comparison-class processing (@cite{aparicio-xiang-kennedy-2015}, @cite{aparicio-2017}) — gaze is more distributed across all four display objects when the listener must construct a comparison class.

These two findings target different mechanisms: pattern (1) is the perceptual / pragmatic story (visual salience of the contrastive property modulates the contrastive inference); pattern (2) is the semantic story (gradable adjectives demand comparison-class binding). The paper's contribution is teasing these apart.

Paradigm

Built on Paradigms.VisualWorld (@cite{huettig-rommers-meyer-2011}). The display contains four objects (ObjectRole):

• target: the intended referent.
• contrastingObject: same category, opposite pole on the adjective dimension. Present only in the contrast condition.
• crossCategoryCompetitor: different category, sharing the adjective property with the target. Always present.
• distractor: unrelated. Always present.

Within-subjects manipulations on Cell:

• Contrast (ContrastCondition): same-category contrasting object present vs. absent.
• AdjType: color, scalar, material — a study-local factor that partitions the cell space into three strata. Adjective type is not a paradigm primitive (the @cite{huettig-rommers-meyer-2011} review does not single it out), so it stays study-local; the paradigm exposes the stratified predicates ContrastReducesCompetitorLooksWhen and RoleSumLowerInBaselineWhen to consume it.

Cross-linguistic generalisation (English, Hindi, Hungarian) is methodological rather than structural: the paper deliberately omits LANGUAGE as a regression predictor (treating it as a clustering unit above participants) because the empirical claim is that the same qualitative pattern survives across all three groups, not that the effect magnitudes are identical. We therefore do not include a Language field on Cell — it would force the predicates to make empirically too-strong pairwise claims (e.g., the marginal-mean helmert interaction does not entail that every Hungarian material cell has a smaller contrast effect than every English color cell, as Figure 3 makes visible). The cross-linguistic generalisation is documented here in prose; if a future study formalises a language-stratified predicate at the paradigm level, the field can be added.

The task is held constant (instruction with definite NP across all trials), so it is not lensed on Cell — varying it would have no within-study consumer.

Architectural Role

This file is an empirical anchor: it defines the experimental cells and qualitative predicates that downstream theoretical models must satisfy. The empirical claims are encoded as paradigm-level predicates, never as rfl-over-stipulated-statistics theorems (@cite{ronderos-etal-2024}'s F/β/p values are documented in prose at each predicate, per the CLAUDE.md Processing scope).

The novel architectural feature relative to @cite{sedivy-etal-1999} is stratification: where Sedivy's contrast effect is universal over cells, Ronderos's is conditional on the stratum (color or scalar, not material). The stratified predicate Paradigms.VisualWorld.ContrastReducesCompetitorLooksWhen was added to support exactly this kind of multi-factor design without inflating the paradigm with a study-specific factor. The interaction (color and scalar effects strictly larger than material effects) is expressed via Paradigms.VisualWorld.ContrastEffectLargerFor — the paradigm-level shape of an "X × condition" interaction.

The scalar baseline disadvantage is expressed via Paradigms.VisualWorld.RoleSumLowerInBaselineWhen — also added to the paradigm because it is a recurring analysis pattern (@cite{aparicio-xiang-kennedy-2015} report it on color vs. scalar; @cite{ronderos-etal-2024} replicate and extend to material vs. scalar).

Theoretical Significance

Two prior accounts of the contrastive inference effect:

1. Lexical comparison-class (@cite{sedivy-etal-1999}, @cite{bierwisch-1989}): scalar adjectives carry a free comparison-class variable, bound by visual context, which makes the contrast pair pragmatically informative. Predicts an effect for scalar but not for color or material (color and material do not require a comparison class — see Features.PropertyDomain.requiresComparisonClass).

2. Perceptual discrimination (@cite{kursat-degen-2021}, @cite{giles-etal-2026}): high perceptual discriminability makes a contrastive description informative. Predicts an effect tracking the noise-discrimination ordering color > size > material from RSA.Noise.

Pattern (1) above is the joint envelope: scalar effect (lexical route), color effect (perceptual route), no material effect (fails both routes). Pattern (2) above is additionally required to capture the semantic-restrictiveness signature on the no-contrast baseline, which is observable independent of any contrast manipulation.

Open architectural threads

• No theoretical witness for SatisfiesRonderosPattern: only the trivial witness trivialLooks is provided. A natural deepening is to instantiate Theories/Pragmatics/RSA/Incremental.lean with a noise-perturbed wordApplies (using RSA.Noise.noiseChannel and the per-domain match/mismatch parameters) and prove the resulting LookProportion satisfies the pattern — this would derive the Ronderos effects from the noise-discrimination ordering rather than observing structural alignment between two independently stipulated orderings. The CG-style vanilla IncrementalSemantics (no noise channel) cannot satisfy SatisfiesRonderosPattern because it would predict equal contrast effects for all adjective types — a useful negative result that could itself be a theorem.
• Connection to Theories/Semantics/Gradability/: Ronderos's semantic factor (scalar = gradable + comparison-class-dependent; color/material = non-gradable) is currently mediated only through Features.PropertyDomain.requiresComparisonClass. A deeper bridge would tie AdjType.scalar to the Semantics.Gradability.Classification subsective/comparison-class hierarchy, so the "scalar requires a comparison class" claim is derivable from gradability theory rather than projected via the domain table.

inductive RonderosEtAl2024.AdjType : Type

Adjective types crossed with the contrast manipulation.

• color: black, blue, brown, green, orange, red, white, yellow.
• scalar: large, narrow, short, small, tall, thick, thin, wide.
• material: cotton, glass, gold, leather, metal, paper, plastic, wooden, woolen.

Adjective type is a study-local factor — it partitions the cells but is not lensed (the contrast factor is the only one swapped by the paradigm-level predicates here).

• color : AdjType
• scalar : AdjType
• material : AdjType

@[implicit_reducible]

instance RonderosEtAl2024.instDecidableEqAdjType : DecidableEq AdjType

Equations

• RonderosEtAl2024.instDecidableEqAdjType x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯

@[implicit_reducible]

instance RonderosEtAl2024.instReprAdjType : Repr AdjType

Equations

• RonderosEtAl2024.instReprAdjType = { reprPrec := RonderosEtAl2024.instReprAdjType.repr }

def RonderosEtAl2024.instReprAdjType.repr : AdjType → ℕ → Std.Format

Equations

• One or more equations did not get rendered due to their size.

def RonderosEtAl2024.instInhabitedAdjType.default : AdjType

Equations

• RonderosEtAl2024.instInhabitedAdjType.default = RonderosEtAl2024.AdjType.color

@[implicit_reducible]

instance RonderosEtAl2024.instInhabitedAdjType : Inhabited AdjType

Equations

• RonderosEtAl2024.instInhabitedAdjType = { default := RonderosEtAl2024.instInhabitedAdjType.default }

@[implicit_reducible]

instance RonderosEtAl2024.instFintypeAdjType : Fintype AdjType

Equations

• One or more equations did not get rendered due to their size.

def RonderosEtAl2024.AdjType.toDomain : AdjType → Features.PropertyDomain

Map adjective type to its Features.PropertyDomain. Scalar adjectives are spatial dimensions (size); color and material map to their eponymous domains. This is the bridge that lets cross-study theorems connect Ronderos's adjective-type stratification to Sedivy's domain-level reasoning and to RSA.Noise's discrimination ordering.

Equations

• RonderosEtAl2024.AdjType.color.toDomain = Features.PropertyDomain.color
• RonderosEtAl2024.AdjType.scalar.toDomain = Features.PropertyDomain.size
• RonderosEtAl2024.AdjType.material.toDomain = Features.PropertyDomain.material

structure RonderosEtAl2024.Cell : Type

A condition cell in Ronderos's 2 × 3 design (contrast × adjective type). Only the contrast factor is lensed — adjective type is consumed by the sub-cell predicates in §3. See the module docstring for why Language is not a field.

• contrast : Paradigms.VisualWorld.ContrastCondition
• adjType : AdjType

def RonderosEtAl2024.instDecidableEqCell.decEq (x✝ x✝¹ : Cell) : Decidable (x✝ = x✝¹)

Equations

• One or more equations did not get rendered due to their size.

@[implicit_reducible]

instance RonderosEtAl2024.instDecidableEqCell : DecidableEq Cell

Equations

• RonderosEtAl2024.instDecidableEqCell = RonderosEtAl2024.instDecidableEqCell.decEq

def RonderosEtAl2024.instReprCell.repr : Cell → ℕ → Std.Format

Equations

• One or more equations did not get rendered due to their size.

@[implicit_reducible]

instance RonderosEtAl2024.instReprCell : Repr Cell

Equations

• RonderosEtAl2024.instReprCell = { reprPrec := RonderosEtAl2024.instReprCell.repr }

@[implicit_reducible]

instance RonderosEtAl2024.instInhabitedCell : Inhabited Cell

Equations

• RonderosEtAl2024.instInhabitedCell = { default := RonderosEtAl2024.instInhabitedCell.default }

def RonderosEtAl2024.instInhabitedCell.default : Cell

Equations

• RonderosEtAl2024.instInhabitedCell.default = { contrast := default, adjType := default }

@[implicit_reducible]

instance RonderosEtAl2024.instFintypeCell : Fintype Cell

Equations

• RonderosEtAl2024.instFintypeCell = Fintype.ofEquiv ((_ : Paradigms.VisualWorld.ContrastCondition) × RonderosEtAl2024.AdjType) RonderosEtAl2024.Cell.proxyTypeEquiv

@[implicit_reducible]

instance RonderosEtAl2024.instHasContrastConditionCell : Paradigms.VisualWorld.HasContrastCondition Cell

Equations

• One or more equations did not get rendered due to their size.

@[implicit_reducible]

instance RonderosEtAl2024.instHasDisplayKindCell : Paradigms.VisualWorld.HasDisplayKind Cell

Display kind is the four-object workspace throughout, per the paper's Methods. Between-study constant, no lens.

Equations

• RonderosEtAl2024.instHasDisplayKindCell = { displayKindOf := fun (x : RonderosEtAl2024.Cell) => Paradigms.VisualWorld.DisplayKind.objectArray }

Note: no HasTask instance. The task is fixed (definite-NP instruction) across all cells — a constant projection cannot satisfy the setTask/taskOf lens laws (taskOf (setTask k c) = k would force the projection to vary). Studies that hold a paradigm factor constant should omit the lens, not stub it with a constant.

Adjective type partitions the cell space into three strata. The paradigm-level predicates ContrastReducesCompetitorLooksWhen, ContrastEffectLargerFor, and RoleSumLowerInBaselineWhen consume Cell → Prop filters; defining the strata as Props lets us state the empirical patterns at exactly the granularity Ronderos's analysis reports.

def RonderosEtAl2024.ColorCells (c : Cell) : Prop

Cells whose adjective is a color term.

Equations

• RonderosEtAl2024.ColorCells c = (c.adjType = RonderosEtAl2024.AdjType.color)

def RonderosEtAl2024.ScalarCells (c : Cell) : Prop

Cells whose adjective is a scalar (spatial-dimension) term.

Equations

• RonderosEtAl2024.ScalarCells c = (c.adjType = RonderosEtAl2024.AdjType.scalar)

def RonderosEtAl2024.MaterialCells (c : Cell) : Prop

Cells whose adjective is a material term.

Equations

• RonderosEtAl2024.MaterialCells c = (c.adjType = RonderosEtAl2024.AdjType.material)

def RonderosEtAl2024.InferenceTriggeringCells (c : Cell) : Prop

The "inference-triggering" strata — color and scalar adjective types — corresponding to the helmert-coded grouping the paper uses for the interaction analysis (material vs. color+scalar). The name is interpretive of the paper's conclusion that this group produces a contrastive inference; the underlying primitive is the helmert contrast.

Equations

• RonderosEtAl2024.InferenceTriggeringCells c = (RonderosEtAl2024.ColorCells c ∨ RonderosEtAl2024.ScalarCells c)

The Ronderos pattern is encoded as a Prop-valued structure with one field per qualitative finding. The empirical claims fall into two distinct families with two distinct observables:

• Contrast-effect family (target-advantage / cluster-permutation analyses, §3 Results ¶1–3): four fields, all over the contrast manipulation on cross-category-competitor looks.
• Baseline-restrictiveness family (No-Contrast total-looks analysis, §3 Results ¶4): two fields, comparing the target + competitor role-sum across strata in the no-contrast baseline.

A theory of contrastive inference "satisfies the Ronderos pattern" iff its predicted look proportions satisfy all fields. Statistical readings (the F/β/p values cited below) need a real-valued aggregator and are out of scope for the paradigm contract.

structure RonderosEtAl2024.SatisfiesRonderosPattern {R : Type} [LT R] [Sub R] [Add R] [Zero R] (looks : Paradigms.VisualWorld.LookProportion Cell R) : Prop

A look-proportion observable satisfies the Ronderos pattern if:

Contrast-effect family (§3 Results ¶1–3):

• color_contrast_reduces_competitor: paradigm-level ContrastReducesCompetitorLooksWhen ColorCells — every color cell shows a strictly smaller cross-category competitor look proportion in the contrast condition than in the no-contrast condition. Paper §3 ¶1: significant cluster 240–600 ms post adjective onset (sum-t = 39.61, p < 0.01). §3 ¶2: target- advantage β = 0.24, t = 2.41, p < 0.05.
• scalar_contrast_reduces_competitor: same shape, restricted to ScalarCells. §3 ¶1: significant cluster 260–500 ms (sum-t = 33.07, p < 0.01). §3 ¶2: β = 0.19, t = 2.02, p < 0.05.
• material_effect_smaller: paradigm-level ContrastEffectLargerFor InferenceTriggeringCells MaterialCells — the contrast effect on every inference-triggering cell strictly exceeds the contrast effect on every material cell. §3 ¶1 adjective-type × condition interaction: significant cluster 280–600 ms (sum-t = 37.96, p < 0.01). §3 ¶2 reports the material main effect as non-significant (β = 0.10, t = 1.08, p = 0.28).

The absence of a material_contrast_reduces_competitor field is deliberate: a null finding is encoded by absence, not by adding ¬ ContrastReducesCompetitorLooksWhen MaterialCells looks (statistical power and direction-of-null are prose, not theorem). The interaction field is the stronger qualitative claim that survives without statistical machinery: material's effect is strictly smaller than the others, not exactly zero.

Baseline-restrictiveness family (§3 Results ¶4):

• scalar_baseline_lower_than_color: paradigm-level RoleSumLowerInBaselineWhen .noContrast [.target, .crossCategoryCompetitor] ScalarCells ColorCells — in the no-contrast baseline, total looks on target + cross-category competitor are strictly lower for scalar than for color cells. §3 ¶4: β = 0.25, z = 2.80, p < 0.01.
• scalar_baseline_lower_than_material: same shape, scalar vs. material. §3 ¶4: β = 0.24, z = 2.40, p < 0.05.

The paper reports no significant color-vs-material baseline difference; we encode this null by the absence of a field asserting either direction (same convention as the contrast-effect null). The two scalar fields together encode the paper's interpretation: scalar adjectives demand comparison-class processing, distributing gaze across all four display objects and away from the two critical roles. This semantic factor is independent of the contrast manipulation.

• color_contrast_reduces_competitor : Paradigms.VisualWorld.ContrastReducesCompetitorLooksWhen ColorCells looks
• scalar_contrast_reduces_competitor : Paradigms.VisualWorld.ContrastReducesCompetitorLooksWhen ScalarCells looks
• material_effect_smaller : Paradigms.VisualWorld.ContrastEffectLargerFor Paradigms.VisualWorld.ObjectRole.crossCategoryCompetitor InferenceTriggeringCells MaterialCells looks
• scalar_baseline_lower_than_color : Paradigms.VisualWorld.RoleSumLowerInBaselineWhen Paradigms.VisualWorld.ContrastCondition.noContrast [Paradigms.VisualWorld.ObjectRole.target, Paradigms.VisualWorld.ObjectRole.crossCategoryCompetitor] ScalarCells ColorCells looks
• scalar_baseline_lower_than_material : Paradigms.VisualWorld.RoleSumLowerInBaselineWhen Paradigms.VisualWorld.ContrastCondition.noContrast [Paradigms.VisualWorld.ObjectRole.target, Paradigms.VisualWorld.ObjectRole.crossCategoryCompetitor] ScalarCells MaterialCells looks

def RonderosEtAl2024.trivialLooks : Paradigms.VisualWorld.LookProportion Cell ℚ

A trivial look model exhibiting both Ronderos shapes simultaneously.

Cross-category competitor looks: 1 in every contrast cell; in no-contrast cells, color 5, scalar 4, material 2. The contrast effect is therefore color 4, scalar 3, material 1 — strictly positive everywhere with inference-triggering strata strictly exceeding material.

Target looks (no-contrast only): color 5, scalar 1, material 6. The target + competitor role-sum in the no-contrast baseline is therefore color 10, scalar 5, material 8 — scalar strictly below both, satisfying the baseline disadvantage fields.

Witness only — carries no theoretical content.

Equations

• RonderosEtAl2024.trivialLooks Paradigms.VisualWorld.ObjectRole.crossCategoryCompetitor c = 1
• RonderosEtAl2024.trivialLooks Paradigms.VisualWorld.ObjectRole.crossCategoryCompetitor c = 5
• RonderosEtAl2024.trivialLooks Paradigms.VisualWorld.ObjectRole.crossCategoryCompetitor c = 4
• RonderosEtAl2024.trivialLooks Paradigms.VisualWorld.ObjectRole.crossCategoryCompetitor c = 2
• RonderosEtAl2024.trivialLooks Paradigms.VisualWorld.ObjectRole.target c = 5
• RonderosEtAl2024.trivialLooks Paradigms.VisualWorld.ObjectRole.target c = 1
• RonderosEtAl2024.trivialLooks Paradigms.VisualWorld.ObjectRole.target c = 6
• RonderosEtAl2024.trivialLooks role✝ c = 0

theorem RonderosEtAl2024.trivial_satisfies_pattern : SatisfiesRonderosPattern trivialLooks

The Ronderos pattern is satisfiable: trivialLooks satisfies all five fields. Without this witness the structure could in principle be uninhabited.

These theorems articulate the theoretical positions Ronderos's data take on. They are type-level connections to the relevant infrastructure (Features.PropertyDomain, RSA.Noise, SedivyEtAl1999), not restated empirical claims.

theorem RonderosEtAl2024.scalar_shares_sedivy_domain : AdjType.scalar.toDomain = SedivyEtAl1999.adjDomain ∧ AdjType.scalar.toDomain.requiresComparisonClass = true

Agreement with @cite{sedivy-etal-1999} on scalar adjectives. Both studies place the scalar contrast effect on the size domain, which Features.PropertyDomain flags as requiring comparison-class binding.

theorem RonderosEtAl2024.color_does_not_require_comparison_class : AdjType.color.toDomain.requiresComparisonClass = false

Disagreement with the comparison-class-only mechanism on color. Color does not require comparison-class binding (so the Bierwisch/Sedivy mechanism alone predicts no contrast effect for color), yet Ronderos finds a robust color contrast effect.

theorem RonderosEtAl2024.material_does_not_require_comparison_class : AdjType.material.toDomain.requiresComparisonClass = false

Material fails the comparison-class route. Material adjectives, like color, do not require comparison-class binding — so the lexical mechanism predicts no contrast effect, consistent with Ronderos's null finding for material. (The lexical mechanism alone is insufficient for color, however; see color_does_not_require_comparison_class.)

theorem RonderosEtAl2024.effect_ordering_aligns_with_noise_discrimination : RSA.Noise.colorDiscrimination > RSA.Noise.sizeDiscrimination ∧ RSA.Noise.sizeDiscrimination > RSA.Noise.materialDiscrimination

Effect ordering aligns with noise discrimination. The perceptual-discrimination route predicts the cross-category competitor reduction to track RSA.Noise's discrimination ordering. Ronderos's qualitative effect ordering — present for color and scalar, absent for material — is consistent with RSA.Noise's ordering color (≈0.98) > size (≈0.60) > material (≈0.40). @cite{kursat-degen-2021} found the same direction on the production side; @cite{ronderos-etal-2024} extends it to comprehension.

theorem RonderosEtAl2024.adjType_to_noise_discrimination : AdjType.color.toDomain.noiseDiscrimination = some RSA.Noise.colorDiscrimination ∧ AdjType.scalar.toDomain.noiseDiscrimination = some RSA.Noise.sizeDiscrimination ∧ AdjType.material.toDomain.noiseDiscrimination = some RSA.Noise.materialDiscrimination

PropertyDomain ↔ noise discrimination wiring. Each adjective type maps through AdjType.toDomain to a PropertyDomain that Features.PropertyDomain.noiseDiscrimination resolves to the corresponding RSA.Noise constant. Recorded here so the bridge to discrimination ordering is auditable.

theorem RonderosEtAl2024.two_mechanisms_factorise : AdjType.scalar.toDomain.requiresComparisonClass = true ∧ AdjType.color.toDomain.requiresComparisonClass = false ∧ AdjType.material.toDomain.requiresComparisonClass = false ∧ RSA.Noise.colorDiscrimination > RSA.Noise.materialDiscrimination ∧ RSA.Noise.sizeDiscrimination > RSA.Noise.materialDiscrimination

The two empirical families project onto different mechanisms. The contrast-effect family is keyed off Features.PropertyDomain.noiseDiscrimination (perceptual route): all three adjective types have some discrimination value, but only those above the material level produce a contrast effect. The baseline-restrictiveness family is keyed off Features.PropertyDomain.requiresComparisonClass (semantic route): only the scalar (size) domain returns true, predicting the no-contrast baseline disadvantage to be uniquely scalar.

This theorem records the two-mechanism factorisation as a typed statement: the scalar adjective type is the unique one whose domain requires a comparison class; color and material are the two whose domains have noise discrimination strictly above the material baseline (where "above the material baseline" means "strictly larger" — material is at the floor). The two mechanisms therefore make orthogonal predictions, and Ronderos's pattern is the joint envelope.