Setup, turn, reveal
How premises establish expectations; how incongruity, misdirection, escalation, and reversal reorganize them; and how punchlines resolve or rupture the emerging pattern.
BombShelterEnter the practice roomBombShelter Research
Mapping the
structure
of comedy.
BombShelter's research team investigates how language, timing, delivery, context, and expectation combine to make comedy work. We are building new tools for understanding the live mechanics of a laugh—and translating that understanding into better daily practice for performers.
Comic
response
response
Language
Delivery
Context
Timing
The central question
Why does
this moment
land?
Comedy is more than funny text. The same line can explode in one room and vanish in another. A pause can sharpen a reversal. A slight change in stress can reveal the premise. A delayed callback can transform a detail the audience heard minutes earlier.
BombShelter Research treats comedy as a dynamic structure unfolding through time. Our work examines not only what was said, but when it was said, how it was delivered, what the audience already knew, and how expectation changed from beat to beat.
The goal is practical: create better measurement, better training signals, and better tools for comedians—without pretending that a model can replace the judgment of a performer or the complexity of a real room.
Areas of investigation
The anatomy
of a laugh.
Our research program explores the interacting systems that shape comic response.
The shape of timing
How pauses, speaking rate, beat boundaries, interruption, overlap, and callback distance affect comprehension, anticipation, and release.
The joke beyond the words
How pitch movement, emphasis, energy, elongation, rhythm, and vocal contrast help audiences locate the comic frame and recognize the turn.
Memory makes meaning
How characters, motifs, rules, and earlier details accumulate across a story so that a delayed punchline can activate context built many beats before.
No single universal room
How different evaluative models interpret the same material, where they agree, where they diverge, and why model choice must remain explicit rather than silently averaged.
From signal to better reps
How performers can use recordings, transcripts, reaction timelines, and repeated practice to test revisions while preserving their own artistic judgment.
Research method
One clock.
Many signals.
Clear limits.
Canonical performance timeline
Speech, acoustic features, transcript revisions, model decisions, and audience actions are aligned to a single sample-accurate clock.
Minimized multimodal evidence
Research features describe timing and delivery without requiring raw microphone audio to be sent to an AI provider.
Reproducible offline comparison
Versioned replay tools let the team compare decision policies against the same canonical recording rather than drawing conclusions from mismatched performances.
Human judgment stays central
Model reactions are measurements and hypotheses—not objective truth. Performer review and appropriately authorized evaluation data define the meaningful standard.
Research principles
Rigorous about data.
Humble about funny.
Privacy by architecture
Raw microphone audio stays on the performer's device. Any future research collection must be explicit, consented, minimized, and reviewable.
No hidden consensus
Different models embody different judgments. BombShelter does not silently compare them during a live show or blur disagreement into a single unexplained score.
Tools, not taste
Comedy remains cultural, contextual, and human. The research team builds instruments for inquiry and practice—not an authority that declares what is universally funny.
From research to rehearsal
Better questions.
Better sets.
Every research system in BombShelter exists to help performers examine the work, test the next version, and return to the stage with sharper instincts.
Discover BombShelter