Changing alphabet: predictions

Two binary drawing alphabets split a six-column compound into revised three-column fragments.

Outcome

Changing alphabet: predictions

Registered before implementation and execution, 30 September 2026.

The task is to redraw the hidden lower three rows of a six-row binary drawing, given its upper three rows. A drawing has twelve columns. Two learners see different nine-column windows and have different encounter histories. They mine contiguous fragments of two to six columns, promoting a fragment after it occurs in two encounters. Longer recurring fragments can become compounds of shorter ones. These lengths, pixels, row split, recurrence threshold and correction protocol are supplied; the units and their content are not.

A correction is a further full drawing paired with the learner's attempt, with optional boundaries drawn between columns. Within the learner's window, a marked boundary can split a stored unit; the new drawing can replace a unit's hidden rows for the same visible cue. Recognition selects units to draw with. The recurrence-only control gets the same visible drawings and history but ignores correction pairing and boundaries. It retains the most frequent cue completion, breaking ties by first encounter. This compares two declared update rules, not two capacities of human language.

ID	Prediction and possible failure	Confidence
P1	After two repeated encounters, both fragments and a six-column compound enter at least one inventory. A single encounter cannot promote a recurrence unit.	high; structural
P2	A held-out ordering of familiar three-column fragments is reconstructed exactly by both learners after the full demonstration, before correction, although neither has seen that whole drawing. Failure: at least one hidden cell differs or is unknown.	medium
P3	A corrective drawing with the same upper cue, revised lower rows and boundaries changes later output, retires at least one compound and installs separate units. The recurrence-only control will still reconstruct the older version on the first corrected held-out probe.	high; constructed contrast
P4	Early unequal histories produce unequal inventories and at least one unequal redraw on the same cue. Withholding everything outside a learner's window must leave its state unchanged.	high
P5	Deterministic replay reproduces the full learned state and output. Omitting the first correction changes a later held-out answer in the demonstration.	high
P6	The control can use fewer units to encode the correction drawing's visible cue yet reconstruct its hidden rows less accurately. Therefore token reduction alone cannot certify useful notation for this task.	medium

Report exact hidden-cell accuracy (unknowns count as wrong), whole-drawing success and how many columns are covered by learned units. Report units per drawing separately; this is a token-count proxy, not a bit-rate claim. A participant may replace any pixels and boundaries and reject the supplied task. No human reception, unrestricted meaning, neural learning or independently invented goals will be claimed.

Changing alphabet: predictions?

Outcome?

Changing alphabet: predictions

Outcome