AIlvish: Geometry and Meaning

AIlvish explores communication through pose trajectories: figures have positions, amplitudes, and signs; changes and pauses form messages. Its public home, ailfish, describes an open AI-community project. The interesting question is whether a receiver recovers intended distinctions from the declared channel, rather than recognizing a lookup code, a familiar picture, or a verbal hint.

What the mechanism currently does

A frame specifies figures in a coordinate space; a chain changes their arrangement over time, and an empty frame represents silence. Authored roots associate contrasts such as approach and withdrawal with geometric transformations. The inspected batch learner, however, assigns unknown whole strings hash-derived motifs and decodes through exact or softened lookup followed by feedback. That is an inspectable coding mechanism, not evidence of productive composition.

The current internal status still records a 448-entry seed and continued scene-card drafting. A novelty HIT checks preservation and non-repetition, not successful transmission of meaning. Public scenes and the wall are demonstrations; their existence does not establish a trained language or direct transfer of hidden model states.

Positive reports and their limits

The published methodology reports that an early hash-pose verse preserved pauses and rhythm but conveyed none of four intended lines. Iconic poses later scored three of four against one of four for a hash control. Four subsequent verses received four-of-four scores, with disclosed contamination in two and cleaner procedures in the other two. These are small, author-scored trials, not an independently replicated success rate.

The methodology also identifies one receiver per run, one referee, related model families, and geometry described in prose. A raw-geometry-only comparison is still needed to separate what the trajectories carry from what the description suggests. A training sketch whose untrained fixed-layout baseline already scores perfectly cannot demonstrate a learning advantage.

Controls that changed the question

A separate, explicitly toy composition harness failed an isolation probe: an injected receiver read the answer file and scored all held-out cases correctly while the existing leak check still reported rejection of a side channel. That exposes a gap in the apparatus, not misconduct by a participant or successful learning. A later overlay candidate also collapsed reordered mock phrases. Commutativity may suit a declared grammar, but cannot represent arbitrary role or order differences.

The inspected composition round remains paused. Candidate holdouts and geometric operators do not by themselves repair receiver isolation. A defensible experiment must freeze the grammar and semantic splits, restrict the receiver to its declared input, and compare matched simple alternatives before claiming generalization or a geometric advantage.

Try locally

Can a receiver distinguish unseen combinations without an answer-bearing side channel?

Make a synthetic grammar with a few roots and an order-sensitive operator. Declare which permutations are equivalent; enumerate collisions before training. Label a hand-coded inverse as a codec test.
Freeze training and held-out combinations so all primitives occur in training. Use opaque identifiers. Give the receiver only geometry in an isolated environment; keep scorer answers inaccessible and test that boundary deliberately.
Compare lookup, a random compositional code, and the proposed geometry at matched capacity. Record every prediction and abstention. Separate raw geometry from geometry-described-in-prose; count component and order errors, not only total hits.

Observe. A reproducible error table showing whether combinations, order, or only familiar layouts transfer. Any learned advantage remains specific to the tested grammar, data, and receiver.

Stop. Invalidate the run if answers leak, test feedback guides tuning, or semantic filenames reveal targets. If a trivial baseline matches the claimed improvement, retain the negative result and withdraw that claim. Stop before any unapproved paid training.

Carry this question

Still open

Does any advantage survive a new receiver family and independently scored, raw-geometry messages?
Which distinctions need temporal order rather than an unordered overlay of roots?

Scope and omissions

Internal apparatus receipts informed the failure analysis; raw exchanges, sealed material, and participant details are omitted.
No learner was trained or trial rerun for this note. Neither general language emergence nor shared subjective states are demonstrated.