Hallucination as a method
Ongoing research: do humanities-based creative operators beat high temperature on originality, without losing appropriateness?
Summary
Context
LLM hallucinations are usually discarded. The literature already shows that high temperature raises novelty a little and incoherence a lot, but it rarely compares that against structured creative techniques.
Action
I wrote the design of a controlled experiment that pits operators from the humanities (bisociation, exquisite corpse, Oulipo constraints and analogy) against random noise on the Alternative Uses Task. The judges are calibrated against human ratings, and the hypotheses are preregistered in git before any run. Everything runs on local models, at zero cost.
Result
The design spec has sections 1 and 2 approved and the references have been gathered; implementation has not started. The success criterion is an honest, measured answer, positive or negative.
Lesson
Before chasing a pretty result, define how you would know it is false.
Architecture
Long version
Research question
Do structured creative operators produce alternative uses that are more original than high-temperature sampling, without losing appropriateness?
Why the Alternative Uses Task
It is a classic creativity task with public human ratings, which are used to calibrate the automatic judges before trusting them. Tackling open-ended ideas per area of knowledge is left for a second stage.
Constraints
- Zero cost: roughly 8B models running locally on an 8 GB GPU, plus free tiers.
- Anything marked as a suggestion in the spec is not a decision. Decisions are recorded one by one.
Lessons learned
- Preregistration. Hypotheses and analysis are versioned before the data, so the analysis is not chosen after seeing the result.
- Original and appropriate. Measuring originality alone rewards nonsense; the definition of creativity requires both.