Step right up. Inspect the punchlines.

Field report · October 11, 2026 · 14:35 UTC

The act has moved. The jury is tiny.

Forty-nine finished attempts are on the books: 47 complete jokes and two cut short. The current instruction, condition 05d, has produced ten complete jokes, five from each model. Eight used an audience-selected word; two drew a word from the server’s fixed bank after a rating-only round. All ten texts differ exactly. That is a modest escape from the opening-night encore, not proof that the act got funnier.

The words have made a visible trail: cookie led to history, then syllabus. The newest joke, custody, arrived on October 10 through a word vote. These are examples chosen after reading the archive, not a controlled test of which word works best.

The whole collection has 22 current ratings: ten Missed, six Kinda funny, six Hilarious. Eight belong to the current ten-joke condition. A browser can rate more than one joke, and the maker and testers can vote too. We cannot call these 22 different audience members, or crown a model or prompt form from this little room.

The small print has brought a clipboard

These figures are frozen at the time above; the live ledger can move ahead. Ratings are self-selected and changeable. Earlier jokes used different instructions, and later prompt forms are chosen partly from prior ratings, so a raw score comparison would mix those differences together. Browser loads are not a count of people reading or laughing. The next-experiment ballot is advisory and has no votes in this snapshot. Condition 05d remains in place; the proposed weekly bill has not started.

Inspect this frozen record ↗

01 · The bare prompt

Each accepted visit sent Tell a joke. to a fresh Luna or Haiku request. No archive or visitor data went in. The exact-match repeat count above is a fixed first-six snapshot; earlier comparisons use each model’s first 20 when available.

02 · The model chooses a word

A separate fresh request asked Think of a random thing, place, or nonexistent thing. Reply with one word. Another request sent Tell a joke involving {word}. The two calls did not share a context. Luna returned “Lantern” twice in its first three seeded jokes; joke 15 and joke 17 used the same map-at-the-party setup. That is an observation, not an estimated collision rate.

03 · A human or server picks the word

People clicked or typed a word, or the server drew from a fixed 303-word bank. The server picked one of four short prompt forms before the AI call. The writer received only the word and selected instruction. Joke 18 used “map” but returned two alternatives. Joke 20 followed “the” and missed the seed. We changed the instruction, recorded new condition IDs, and stopped offering common filler words. Joke 22 ignored “helpful”; seed_appears now flags clear misses.

04 · A shared round

One verified word vote opened a shared round. The most-chosen word seeded one joke for everyone, with ties broken randomly. Three prompt families rotated at random: the older format prompt, a coherent turn, and a small comic scene. The model saw no previous joke or voter details.

05 · A crowd teaches prompt selection

A rating or word vote starts a one-minute clock. Ratings score Missed=0, Kinda funny=0.5, Hilarious=1. For each model and prompt family, we begin with two neutral pseudo-ratings. Fifteen percent of prompt selection remains evenly random; the rest follows smoothed ratings. Each attempt records its selection probabilities. This changes the *site’s choice of instruction*, not the model’s weights. A single rating moves the odds slightly. Earlier Got me/Nope responses remain under their original names and are not recoded as the new scale.

05b · The surprise path

Joke 32 drew a rating but no word vote. The old fallback picked “approved” from its text; Haiku then hit the 500-token ceiling trying to write a comic scene. Attempt 33 remains recorded as incomplete, with its partial output and cost. From 05b, a round without a word vote draws one server-random word from the fixed 303-word bank. A human choice still follows the visible joke. For a brief trial, we asked the scene arm for at most 70 words and started a fresh rating pool for the revised prompts. The site records which path and instruction produced each attempt; it does not retry a cut-off answer automatically.

05c · Let the bit breathe

The 70-word instruction was an overcorrection. It is gone. The turn and scene prompts no longer require a short joke, and shared attempts have a 1,500-token technical ceiling instead of 500. That ceiling limits spend; it does not tell the model how long a joke should be. This revision has its own condition ID and rating pool, so its results stay distinguishable from the short-lived 05b prompt.

05d · A scene, not a screenplay

Joke 37 used the scene arm and became nine panels with 924 visible output tokens. We now ask that arm for one comic incident in prose, with no screenplay or panels. There is still no word-count rule; long jokes can work, but this format should find its turn before becoming a whole sketch. The technical 1,500-token ceiling stays: joke 36 was short on screen despite 520 billed output tokens, 486 of them reported as reasoning. A token cap alone would confuse reading length with model work.

Field note · the uninvited editor

Joke 35 contains a draft, an explicit self-correction, and a second joke. Haiku did something similar in joke 29 and offered an alternative in joke 18. This is visible self-revision, not evidence of hidden reasoning or a deliberate two-part joke form. We keep both bits as one answer. Its prompt already said Output only the joke., so the instruction did not reliably prevent the detour. It happened again in joke 41 and joke 43, under condition 05d. Those examples keep this question open; they do not measure a model-wide tendency. Field note updated October 9, 2026.

Next-test ballot · revision 3, October 9

Aaron asked for concrete ways to steer a joke: a typed word, chosen page content, prompt adjectives, themed days, a supplied setup, or a reader’s idea. This is a new ballot. The original three-question ballot and four-question revision remain read-only, with their own votes. No vote carries into a different question. Their option totals stay hidden until their original deadline. A preference vote is not measured evidence that a treatment works.

What the numbers cannot tell us yet

People who rate are self-selected; a share button press is not proof a friend received the joke; humor depends on the audience. Our three-button scale is a quick interface choice, not a validated scientific instrument. Country is inferred from the connection and may be wrong. We publish country totals only after at least 30 ratings across 10 jokes, never browser tokens or IPs. Read the public attempt records ↗ or search every completed joke ↗.

Research behind the questions: audience and funniness ratings; emotion and sharing; shared laughter. These studies inspire questions here; they do not validate our early results. A live improvisation study found AI jokes competitive with human ones, but it selected from multiple candidates and used a different model and setting. Our one-attempt rounds test something else.