Documentation

Linglib.Phenomena.Gradability.Studies.CobrerosEtAl2012

Cobreros, Egré, Ripley & van Rooij 2012: 4-Element Sorites Model #

@cite{cobreros-etal-2012}

The paper's worked 4-element example (described in the body text on p. 354, immediately above footnote 4), formalized against the substrate types in Theories/Semantics/Supervaluation/TCS.lean.

The model #

Four individuals a, b, c, d arranged in a similarity chain a ~ b ~ c ~ d (with reflexivity and symmetry, no other pairs). The classical extension of tall is {a, b}. The paper uses this model to illustrate:

Cross-framework note #

tcs_supervaluation_disagree_concrete instantiates the Theory-file existence theorem tcs_vs_supervaluation_borderline_contradiction at (soritesModel, .tall, .b). Comparator caveat: the paper's headline borderline-contradiction contrast (p. 356) is actually with Hyde 1997 subvaluationism, not with Fine 1975 supervaluation. Subvaluationism agrees with TCS that P ∧ ¬P holds at borderline cases — they disagree on framework foundations. The Lean theorem here formalises the formal-validity-side contrast (p. 359) where TCS at t mode says ✓ and supervaluation says ✗. Subvaluationism is not formalised in linglib; a proper "borderline-verdict comparator" theorem awaits its formalisation.

The asymmetric-engagement reciprocation: LassiterGoodman2017PMF.lean::lg_literal_borderline_bounded proves L&G's literal-meaning rule predicts ≤ 25% acceptance for borderline tall ∧ ¬ tall, well below the @cite{alxatib-pelletier-2011} empirical 44.7%. That paper's docstring cites TCS as the alternative that handles the data; this file's b_tolerant_contradiction shows the TCS prediction direction matches the data.

Architecture #

This file is anchored on the Cobreros-Egré-Ripley-van Rooij 2012 paper itself. Per CLAUDE.md "no bridge files" + "Theory examples to Phenomena" discipline, the worked example was extracted here from a previous embedding inside Theories/Semantics/Supervaluation/TCS.lean.

Four individuals — the elements of the paper's chain.

Instances For
    Equations
    • One or more equations did not get rendered due to their size.
    Instances For
      @[implicit_reducible]
      Equations
      @[implicit_reducible]
      Equations
      • One or more equations did not get rendered due to their size.

      One vague predicate — "tall".

      Instances For
        Equations
        • One or more equations did not get rendered due to their size.
        Instances For
          @[implicit_reducible]
          Equations
          @[implicit_reducible]
          Equations
          • One or more equations did not get rendered due to their size.

          The similarity chain a ~ b ~ c ~ d with reflexivity + symmetry, nothing else. Implemented via a Bool-valued helper so the Decidable instance reduces through Bool equality, keeping the model fully computable for decide-based extension proofs.

          Equations
          Instances For

            The similarity relation as a Prop, for direct integration with the modal-logic substrate's boxR/diamondR.

            Equations
            Instances For

              The paper's running 4-element model: I(tall) = {a, b}, similarity chain a ~ b ~ c ~ d.

              Equations
              • One or more equations did not get rendered due to their size.
              Instances For
                @[implicit_reducible]

                Per-pair decidability of the model's sim relation. Needed for decide-style proofs of the extension theorems below. Since soritesModel.sim P ignores P (definitionally soritesSim), the instance reduces to soritesSim's DecidableRel. The DecidablePred (soritesModel.sim P x) shape needed by toleranceSpace is auto-derived from this per-pair instance.

                Equations
                @[implicit_reducible]

                StrictAt is decidable on the sorites model: it's a finite universal over the 4 elements.

                Equations
                @[implicit_reducible]

                TolerantAt is decidable on the sorites model: it's a finite existential.

                Equations

                b is borderline-tall: tolerantly tall (witness: a), but not strictly tall (counter-witness: c, which is similar to b but not classically tall).

                c is borderline-tall: tolerantly tall (witness: b), but not strictly tall (counter-witness: d).

                d is NOT borderline: not even tolerantly tall, hence not in tolerant \ strict.

                Cross-framework divergence theorem, instantiated at (soritesModel, .tall, .b).

                On the same tolerance neighborhood {d | b ~ d}:

                • TCS says tall(b) ∧ ¬ tall(b) is tolerantly true
                • Fine 1975 supervaluation says it is super-FALSE (= Truth3.false)

                The frameworks disagree on the formal-validity-side verdict. Caveat (per the file docstring): the paper's headline contrast on borderline contradictions (p. 356) is with Hyde 1997 subvaluationism, which agrees with TCS on the verdict. Fine 1975 supervaluation is the comparator we can formalise here because it's the only one of the two with a substrate in linglib.

                Non-transitivity in the full vocabulary with similarity predicates Iₚ. Each one-step Pa, aIb ⊨ˢᵗ Pb is valid (Fact 2, p. 354); each two-step Pa, aIb, bIc ⊨ˢᵗ Pc is valid (Fact 3, p. 354); but three-step Pa, aIb, bIc, cId ⊨ˢᵗ Pd is invalid because d requires three similarity steps from the classical extension {a, b}, exceeding Fact 3's two-step limit.

                Attribution: paper §3.4.1 (p. 372) demonstrates non-transitivity for
                `⊨ᶜᵗ` (the chain `Pa_1, a_1I_Pa_2, a_2I_Pa_3 ⊭ᶜᵗ Pa_3`), and
                footnote 14 formalises the simple-transitivity statement. The `⊨ˢᵗ`
                sorites version is in §3.6 (Version 1, p. 376). The structural
                pattern (each step valid, chain fails) is §3.4.1's contribution; the
                `st`-instantiation here transfers it to the relation the paper
                advocates. 
                

                Three-step st-INVALIDITY: tall(a), aIb, bIc, cId ⊭ˢᵗ tall(d).

                Each step is individually st-valid (above), but the chain of three exhausts the two-step tolerance limit (Fact 3, p. 354), so the conclusion fails to hold even tolerantly at d. Structural pattern of paper §3.4.1; paper itself states the ⊨ᶜᵗ version there and the ⊨ˢᵗ sorites Version 1 in §3.6 (p. 376).

                Two Lean-checkable cross-framework theorems that close gaps flagged by the cross-framework reconciler:

                1. **TCS-vs-product-rule expressivity gap**: TCS predicts borderline
                   contradictions *categorically* (the conjunction is true, not
                   fractional); any framework that predicts them as a product of
                   complementary probabilities is *fractionally bounded* by 1/4
                   (AM-GM). The architectural divergence between TCS and L&G's
                   literal rule is structural, not parametric.
                2. **Klein-delineation tension**: in `soritesModel`, two adjacent
                   individuals are tolerantly similar AND both borderline, yet
                   classically separated — exactly the configuration where any
                   sound Klein delineation would over-impose a scalar relation. 
                

                TCS predicts borderline contradictions categorically. Helper: for any T-model M, predicate P, and individual a that is borderline-P, the conjunction P(a) ∧ ¬P(a) is tolerantly true. Extracted from the TCS half of tcs_vs_supervaluation_borderline_contradiction.

                TCS-vs-product-rule framework expressivity gap. A framework-level contrast theorem with two simultaneous halves, witnessing that TCS and any "literal rule = product of complementary probabilities" framework (paradigmatically L&G: Phenomena/Gradability/Studies/LassiterGoodman2017PMF.lean::lg_literal_borderline_bounded) diverge structurally on borderline contradictions:

                • TCS half: the conjunction P(a) ∧ ¬P(a) is tolerantly TRUE (not fractional) at every borderline a. This is a categorical framework prediction — independent of any parameter.
                • Product-rule half: any product q · (1 - q) for q ∈ [0, 1] is bounded above by 1/4 (AM-GM). This is the framework-level ceiling on any literal-meaning rule that multiplies a probability by its complement. L&G's lg_literal_borderline_bounded is the PMF-typed instance of this generic bound.

                The empirical Alxatib-Pelletier 2011 acceptance rate (44.7% — see Phenomena/Gradability/Vagueness.lean::alxatibPelletier2011Tall) exceeds the product-rule's 25% ceiling, refuting the literal-rule framework empirically. TCS's categorical prediction is consistent with the empirical direction without committing to any specific fractional value. The structural divergence is what's interesting: L&G's framework architecture cannot represent the empirical direction (any literal-rule PMF is bounded ≤ 25%); TCS's framework architecture predicts it categorically.

                Klein-delineation tension witness: in soritesModel, the individuals b and c are both borderline-tall AND tolerantly similar (b ~_tall c), yet classically separated (b is classically tall, c is classically not-tall). Any Klein-style sound delineation derived from M.interp .tall would impose a scalar relation R b c on this pair (per Theories/Semantics/Comparison/Delineation.lean::IsSoundDelineation), over-generating where TCS says b and c are both equally borderline and the scalar relation should be unwarranted.

                This is the Lean-checkable witness for the prose tension flagged at Comparison/Delineation.lean lines 566-571.