Hook Manifold
A research instrument that models a song as an Umwelt world changing in time, to test whether coherence-sensitive trajectories predict hook recognition better than classical baselines. Built to be able to publish a no.
IN BUILD
What it is
Hook Manifold ingests time-aligned audio evidence, evolves an Umwelt world as the track plays, and exposes trajectories for listener-response hypotheses: hook salience, attention, expectation, familiarity, novelty, memorability and replay intent, each with a confidence and its provenance.
The question it exists to answer is narrow and falsifiable: do coherence-sensitive world-state trajectories predict human hook recognition, delayed recall or replay intent better than equivalent classical and coherence-disabled models? The project must produce an answer even when the answer is no. It is not a viral score.
The estimator ladder is explicit: a cohort prior, persistence, handcrafted features with a linear model, a classical echo-state reservoir, the engine with coherence disabled, and the engine as shipped. The last two are pinned to agree exactly where no coherence can arise and to differ where it can.
What exists today
Private repository, last pushed 19 July 2026. Phases 0 to 4 complete: contracts, a causal audio-to-events extractor, blank replay with snapshot parity, labels, baselines and the coherence-ablation seam; 61 tests at phase 3. The phase 5 machinery is built: a pre-declared pilot protocol, an annotator web app and a content-blind split.
Not done
No real corpus yet: the pilot needs 20 to 50 legally usable tracks and 5 to 20 listeners. Until then nothing about real music is claimed.
Capacity and compute accounting between arms is owed before any superiority claim.
Next
Collect the pilot corpus and run the pre-declared protocol.
Built with
Python Umwelt numpy