I plan to create a series of markets with the aim of crowdsourcing a Haiku on Manifold Markets. It will work in a similar style to Large Language Models, with manifold itself providing the “weights” of the top predicted next words. The markets will pay out based on whether the resulting final “output” is a valid haiku. A valid haiku has:

The correct grammar requirement will be judged loosely, particularly if departures from correct grammar contribute meaningfully to the aesthetics of the haiku.

The words will be chosen with a weighted random number based on their market weights at the time of market resolution.

1. The final valid/not-valid decision. (This market)

2. Vocabulary Market. This will be a free-response/bounty market in which the set of words available for generation is established. In this market, submissions that end up being used in a valid haiku will pay out.

All markets will resolve when either 1) a valid haiku has been created or 2) the process has failed to produce a valid haiku, in which case the “failure” options in each market will pay out.

This market will resolve YES if I decide that a valid haiku has been created through the process I describe. It will resolve NO if it becomes clear that the process has failed irretrievably, or it has completed and the generated output is not a valid haiku. All the other markets are essentially a derivative of this one.

I will not trade in these markets. I reserve the right to modify the text of the markets to avoid unintended consequences/gaming, preserving as much as possible the original intent and mechanism.

Edit 1: It turns out this mechanism leads to quite a high "temperature" which in LLMs can generate nonsensical output.

 I am lowering the temperature slightly by only taking the top 20 scoring words in each market. 

Also, to reduce possible gaming of the "Haiku fails" outcome, I want to set the initial probability of each word to zero or close to zero. However this is not possible with the API, so I will instead buy up the "Haiku fails" shares in each word market.

Edit 2: The temperature is too darn high! We are still alive, but it's clear that the danger from selecting lower-weighted words is only increasing. So I am invoking my "I may change things" clause and changing the word selection algorithm as follows: The words will be sorted in order from highest to lowest weight. For each word the algo will either select it as the next word, or skip it and move on to the following word, with 50% probability.