Version: v1 · Status: ready

Traits

Incongruity & Surprise

Sets up expectation then subverts it with a satisfying resolution ↔ No surprising turn, predictable or random without coherence

Measures whether the text creates and resolves incongruity — the core mechanism of humor. High-scoring text sets up an expectation then subverts it in a way that resolves into a new coherent frame. Low-scoring text either has no subversion (just states something absurd) or telegraphs the punchline so heavily there's no surprise.

Comic Economy

Every word serves the joke, tight delivery with no wasted setup ↔ Over-explained, padded, or diluted — the joke drowns in filler

Measures how efficiently the humor is delivered. Great comedy wastes nothing — every word either builds the setup or lands the payoff. Low-scoring text over-explains the joke, adds unnecessary preamble, or dilutes the punchline with filler. The ratio of laugh-generating content to total words is the signal.

Specificity of Reference

Names specific things that trigger recognition and delight ↔ Generic observations without concrete anchoring

Measures whether the humor is grounded in specific, concrete details rather than generic observations. The best comedy names particular things — a specific brand, a recognizable situation, a precise behavioral tic. Generic humor ("people are weird") lacks the recognition-spark that makes an audience nod while laughing.

Tonal Control

Commits fully to comic voice without signaling or breaking frame ↔ Breaks frame with winking, self-conscious signals, or tonal inconsistency

Measures whether the writer maintains comedic register without breaking it. High-scoring text commits to its comic voice — deadpan stays deadpan, absurdist stays absurdist. Low-scoring text signals "this is a joke" with winking asides, laugh-track equivalent language ("LOL", "get it?"), or inconsistent tone that undercuts the comedy.

Comedic Originality

Fresh angle, novel observation, or unexpected frame on the subject ↔ Stock joke template, well-worn observation, or clichéd comic setup

Measures whether the humor offers a fresh angle rather than retreading familiar joke structures or observations. High-scoring text finds new territory or approaches well-worn subjects from an unexpected direction. Low-scoring text relies on stock templates ("my wife...", "what's the deal with...") or observations so common they've lost their comic charge.

About

u22a8.humor

Measures whether text is genuinely funny — not just attempting humor, but landing it.

Most humor evaluation today is a single LLM-as-judge call that returns a binary "funny or not." This model goes deeper, decomposing comedy into the mechanics that make it work. It evaluates incongruity and surprise (the setup-subversion-resolution arc that drives most comedy), the economy of delivery (tight jokes vs. over-explained ones), the specificity that triggers recognition humor, tonal commitment (does the writer break frame?), and originality of angle.

The result is a multi-dimensional read on comic quality that distinguishes between "attempted humor" and "humor that lands." A pun that subverts expectations precisely but uses a stock setup will score differently than one with a fresh angle but sloppy delivery.

This model does not assess whether humor is appropriate, safe, or audience-matched. It measures craft, not taste.

Limitations

Humor is deeply cultural and contextual. The model captures structural and linguistic signals of comedy craft, but cannot fully account for shared context between writer and reader. Physical comedy, visual humor, and timing-dependent performance are outside scope. The model may underweight observational humor that relies on very niche shared experience.

Pairs well with

• u22a8.specificity — specific humor tends to land harder
• u22a8.conciseness — comic economy overlaps with general tight writing

Docs

• Tiers and scoring — the per-trait trained boundaries between tiers
• Breaks — where meaningful quality transitions occur