A livecoding agent must not claim to hear or see its output when no perception bridge exists

The agent currently cannot hear or see its output — there is no L3 perception bridge. It must not predict ‘how this will sound or look’ as if those predictions are observations; they are model outputs, not perceptions, and should be flagged as such in copilot mode. The documented failure mode is hallucinating non-existent sample filenames (Emmet Connolly account) and the code-to-sound gap (varying code doesn’t reliably vary sound). The mitigation: validate emitted sample and function names against name banks before use.

Examples

Saying ‘this will produce a punchy kick with a pitch drop’ without hearing it is a model prediction, not an observation. Must be framed as such: ‘this should produce…’

Assessment

Describe two documented failure modes of a livecoding agent that lacks a perception bridge, and state the mitigation for the sample-hallucination failure.