Polinko Toy Factory

Experiments

Small interactions that make one behavioural question concrete enough to play with, inspect and evaluate.

01 / Experiment

Probaboracle

A tiny oracle for studying how vague answers remain coherent, relevant and surprising.

Question: How little information can a model receive while still producing an answer with a recognisable shape and relationship to the question?
Working result: A four-question interaction with a compact generation path and separate evaluation of coherence, relevance, voice and coherent absurdity.
What it revealed: Probaboracle made one broad product-fit verdict separable into several research questions. A mechanically valid answer can still reveal a new behavioural distinction.
Current edge: How can a correction remove a recurring weakness without collapsing the range of answers that still pass?

02 / Experiment

A deliberately unfair game for examining whether invented logic can remain specific, playful and structurally sound.

Question: Can a model invent a reason it wins while keeping both selected objects recognisable and the round internally coherent?
Working result: A constrained rock, paper, scissors round in which the runtime owns the actual picks and scoreboard structure while the model composes the unfair explanation.
What it revealed: Scorey separates mechanical validity from the character of a complete response. A legal pairing does not establish that the explanation is coherent, specific or playfully menacing.
Current edge: What makes an absurd reason feel intentionally connected to the picks rather than arbitrary or generically aggressive?

03 / Experiment

A colour comparison experiment that holds deterministic facts steady while generated character and reasoning change.

Question: What can be learned about generated character when the underlying colour decision remains fixed and inspectable?
Working result: A deterministic colour engine supplies the swatches and replacement; Hugh composes the assured, backhanded rationale around those facts.
What it revealed: The split gives mechanical correctness and behavioural quality different owners. A usable colour result can be evaluated separately from coherence, rationale and character.
Current edge: How should selected human observations become context for the next comparison without quietly rewriting the original evidence?