An agent without the L3 perception bridge must not claim to have evaluated perceived audio or visual output

Until the L3 perception bridge (get_fft, screenshot, did-output-change diff) exists, the agent can only evaluate intent versus code structure, not intent versus perceived output, and it must state this limitation explicitly rather than imply it heard or saw the result. This is an honesty constraint with a practical consequence: the agent’s predictions of perceptual effect are models, not observations. The corollary is that the agent defaults toward fewer edits when perception would be needed to confirm an action was correct — since it cannot verify the effect, it cannot trust its own assessment of it.

Examples

Correct: ‘I predict this filter sweep will open the sound over 16 cycles — I cannot confirm this without L3.’ Incorrect: ‘The filter sweep opened the sound nicely.’ — which implies perception that does not exist.

Assessment

Why must the agent surface its L3 limitation explicitly rather than staying silent? What is the practical consequence for edit frequency when perception is unavailable?