The overarching taste check compares each intended intent-vector field against a perceived estimate within tolerance

The intent-vector comparison asks whether the output matches what was intended. For each intent field (energy, density, section, reactivity_target, timbre/mood) a perceived estimate is derived from the output and compared to the intended value within a tolerance. The energy field compares normalized 4-bin FFT magnitude (audio) and inter-frame motion (visual) within about ±0.2. Density needs onset rate and is not-possible on this rig; section checks whether the band profile fits the role; reactivity_target is checkable now by confirming the thunk’s FFT index matches the intended band; timbre/mood has no perceptual extractor. A gap beyond tolerance feeds the policy table: behind the arc means build, ahead means subtract, and an inert edit means try a different move.

Examples

Intended energy 0.7, but the code shows only one muted voice with no active gains above 0.3 — the proxy signals behind the arc, and the policy response is to add a layer.

Assessment

For the energy field, what is the pass/fail tolerance, and what does a gap beyond it trigger in the policy table?