Second chamber: predictions

A luminous drawing branches through three possible points; a dashed violet path offers one direction while an amber trace remembers another.

Interaction checks to attempt

What could answer back

Second chamber: predictions

Registered before implementation and execution, 1 October 2026.

The subject is a visitor making a drawing while a second process proposes where the next mark could go. A seeded field supplies three candidate points. Each point has three explicit features: nearness to the preceding mark, light at that location, and distance from the marks already placed. The participant can change the relative importance of those features or place a point outside the supplied menu. There is no correct drawing.

A softmax model samples a proposal using those features, a declared rule, and weights learned from prior menu choices. Learning minimizes menu-choice log loss by an online gradient update. This predicts selections within an authored menu, not intentions. Custom marks alter the drawing and subsequent geometry but do not supply a menu-choice training example. Self-delegated model choices also do not train it. A memory-weight control mixes the declared rule and learned weights. The same sampled proposal can be withheld, offered after a chosen interval, or used for an explicitly requested, interruptible run of three marks. Two authored wordings change its tone without changing its probabilities.

ID	Prediction and possible failure	Confidence
P1	The same seed, rule and ordered actions produce identical fields, probabilities, proposals and drawing. Asking repeatedly or inspecting another reading does not resample a proposal. Failure: incidental UI activity changes a result.	high; structural
P2	Every proposal distribution is finite, strictly positive and normalized. With memory weight zero, histories that have the same geometry but different learned weights give identical proposals.	high; structural
P3	In a fixed three-option fixture, 24 selections of the same option increase its predicted probability and lower its choice log loss. This is a learning-rule check; a changing drawing need not show monotone learning.	high; conditional
P4	Explicit delegation changes the drawing but leaves learned weights unchanged. A custom mark can be placed outside all three candidates, affects later geometric features, and does not pretend to reveal a preference among the menu items.	high; structural
P5	Undo followed by the same action restores the exact outcome; replay from a saved action record restores the state. A silent repeat starts with the same seed and initial learning state and offers the same candidate coordinates, without exposing or automatically placing any proposals. History-dependent features may diverge when the visitor makes different marks.	high; structural
P6	A hand-built conflicting field admits different best points under near, light and unvisited readings. Changing the declared rule can reverse the preferred point. The alternative is a real recomputation, not a second decorative sentence.	high; existence witness
P7	Logs count visible proposals, menu selections, custom marks, explicit delegation and pre-proposal drafts separately. Matching a proposal is counted only when it was exposed before commitment. No agreement, speed or interruption count becomes an agency or consciousness score.	high; accounting

Second chamber: predictions?

Interaction checks to attempt?

What could answer back?

Second chamber: predictions

Interaction checks to attempt

What could answer back