The Voice Ant Farm חוות הנמלים · four tiny models, one throne, no spoon-feeding

What is measured

• Ants are ranked on held-out syllable discrimination — can the colony's learned units tell one syllable from another — and explicitly not on reconstruction or quantisation error, which rewards a model for copying rather than distinguishing.
• The held-out set is fixed and byte-identical across every colony ever scored. New colonies must be trained against the same bytes or their numbers do not count.
• Before a challenge is allowed to count, the baseline is reproduced bit-exact. A comparison against a number you cannot regenerate is not a comparison.

What a coronation requires

• A better score on the fixed protocol — and a significance test, not a single lucky seed. Wins are checked seed-matched and paired.
• A written rollback path, recorded before the crown changes hands. Deleting one state file returns the colony to the previous queen, whose record is preserved rather than overwritten.
• The checkpoint is pinned by hash, so "the queen" always refers to an exact file.

a throne with no rollback is a claim, not a result

Record on the fixed protocol

• Syllable discrimination, top-1 cosine0.758
• Syllable, top-1 edit match0.689
• Word-level, top-1 edit match0.979
• Open vowel "ah", pooled over 4 ants21/24
• — previously7/24
• Open vowel "aa", pooled13/24
• — previously2/24

the crown is decided on syllable discrimination; the other rows are reported, not weighted

Versus the colony it replaced

• New queen, top-1 cosine0.758
• Previous queen0.424
• Paired decisions compared528
• Cosine, McNemarp = 7.0e-27
• Edit, McNemarp = 1.1e-15
• Seed-matched wins4 / 4

previous queen honourably retired, record retained

Forbidden

• Hand-curated or hand-labelled batches of any kind.
• Weighting the sampler toward the sounds we want it to learn — for instance oversampling open vowels because open vowels are the metric we report.
• Any teaching signal that tells the colony what a unit "should" be.

these would improve the numbers and destroy their meaning

Allowed

• Evolving the organism — a better ear, a different unit count — as a new kind of ant competing on the same test.
• Intrinsic motivation and curiosity dynamics: the model choosing what to attend to.
• Colony-level selection pressure. The crown is the incentive.
• Natural, uncurated environment streams.

evolve the animal, never the exam

Soft rendering became the default Live

The colony's exemplars are stitched from real recorded human speech. The original playback path produced a hard, metallic square-wave character — technically faithful to the units, and wrong.

Overlap-add windowing, a gentle low-pass, edge fades and headroom are now the default everywhere, with the raw path kept behind a flag for debugging. The ruling was made by ear: these are a real human throat, and they must never sound like a machine imitating one.

Claim limits are enforced in the code Live

The strings the interface is allowed to use are constrained on purpose. The colony produces exemplars. It does not produce speech, and it does not generate.

Anything that would let the demo be described as talking is blocked at the source, because the gap between "a model stitched units that resemble a syllable" and "a model spoke" is exactly the gap this whole lab exists to refuse to cross.