📄 Must φ is felicitous only if φ is not known

Summary

The paper asks what licenses English epistemic must. One influential view (von Fintel & Gillies 2010) says must φ is felicitous only when the evidence for φ is indirect. Goodhue constructs minimal pairs where both speakers perceive exactly the same things, but only one of them can naturally say "It must be raining." He argues that the indirectness view can handle these pairs only by stipulating that what counts as "direct" shifts with context. A knowledge-based condition explains them in a principled way: must φ is fine only if the speaker does not know φ, as Giannakidou & Mari (2016) proposed. Theories of knowledge such as Lewis 1996 then account for why one speaker knows and the other does not. Combined with Kratzer's ordering-source semantics, this supports a weak reading of must, where stereotypical assumptions are needed to reach φ. The author also sketches how the condition could be derived as a conversational implicature. He shows that must in deductive conclusions puzzles every account, and suggests treating it as a separate modal flavor.

Key points

• The two rival felicity conditions are:
  • indirectness: no directly known proposition entails φ (von Fintel & Gillies 2010; Matthewson 2015; Lassiter 2016; Mandelkern 2016);
  • non-knowledge: the known propositions ∩Epi do not entail φ (Giannakidou & Mari 2016).
• The semantic backbone is Kratzer 1991:
  • a modal base Epi is ordered by a stereotypical source Norm;
  • must quantifies over the resulting Best worlds;
  • so must φ can be true while φ is false, which captures Karttunen's 1972 weakness intuition.
• The paper argues that G&M's piano examples do not refute indirectness, because trustworthy reports count as direct evidence for von Fintel & Gillies. New identical-perception pairs are offered instead.
• The new evidence for the epistemic account also favors a weak, non-entailing semantics for must (in line with Karttunen, Kratzer, Lassiter and G&M).
• Deductive conclusions are a puzzle for both accounts. Cross-linguistic hints of a distinct deductive modal:
  • Spanish debe de vs. tiene que;
  • Portuguese deve vs. tem de (Marques 2020).

Interesting ideas

• I1. Hold the perceptual evidence fixed and vary only the background beliefs (e.g. a scheduled fake-rain shoot) to isolate whether a cue depends on evidence type or on knowledge state; this controlled minimal-pair design generalizes to any pragmatic diagnostic.
• I2. Treat must φ as signalling that the speaker does not fully know φ but reached it via default assumptions, not as a claim that φ holds.
• I3. Derive the "not known" signal as a quantity-style implicature: had the speaker known φ, the bare assertion would have been the expected choice.
• I4. Distinguish epistemic must from deductive must (conclusion of explicit premises), using cues like a preceding therefore/then or enumerated premises, mirroring the Spanish/Portuguese lexical split.
• I5. Model must with a modal base plus an ordering source (known facts vs. normal-course assumptions), keeping the two sets explicit.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): Build negative-control test pairs where only the speaker's background knowledge differs, and check that the modal/certainty analysis flips accordingly.
• I2 (world): Store must φ in the text world as a speaker-attributed, defeasible belief tagged "inferred via assumptions" rather than as an asserted fact, so later contradictions do not break consistency.
• I3 (frames): In crate prag, add an implicature rule that must/probably triggers "speaker lacks knowledge of φ", which is cancellable and can be explained by pointing to the stronger unused alternative.
• I4 (answer): In logic or word-problem readers, treat "therefore it must be X" after premises as a certain derived conclusion rather than a hedge, avoiding false uncertainty in answers.
• I5 (world): Represent known facts and stereotypical defaults as separate layers, so every modal inference can be explained as "facts + these defaults ⇒ φ".