home/ modules/ evaluating-livecoded-av-without-perception

Evaluating live-coded AV quality without a perception bridge

L3Craft 3h F · Live coding: music H · Creative coding & live-coded visuals 12 atoms

Learning objectives

learner can classify any quality heuristic as [NOW] (static-checkable), [L3] (perception-blocked), or not-possible (architectural limit), and justify the classification
learner can apply every available [NOW] proxy — sonic-token diff, .out() presence, reactive-term scan, .analyze('hydra') tag, intent-vector fields — to audit a live-coded AV snapshot without perceiving the output
learner can state the honesty constraint: distinguishing agent predictions from observations, and defaulting to fewer edits when perception is unavailable
learner can trace a quality gap (intent vs. code structure) to the correct policy response using the intent-vector comparison

Capstone — one whole task that evidences the objectives

Given a frozen AV snapshot (a Strudel pattern file and a Hydra sketch), run the full [NOW] quality audit by hand: (1) classify each of the eight heuristics as [NOW], [L3], or not-possible and explain why; (2) check the sonic-token diff, .out() presence, time/reactive terms, and .analyze('hydra') tag; (3) compare each intent-vector field against the code structure and identify any gap beyond tolerance; (4) state what the policy response should be for each gap; and (5) write a one-paragraph agent output that honestly surfaces every L3 limitation without implying perception that does not yet exist.

This module is about a foundational constraint in agent-assisted live coding: the agent cannot hear its audio or see its visuals, so every quality check must be decomposed into what can be verified right now from the code structure versus what requires a perception bridge that does not yet exist. That distinction is not a temporary inconvenience — it determines what honest, trustworthy agent output looks like, and it determines which edits the agent is entitled to make autonomously versus which it must flag as predicted rather than confirmed.

The arc is tightly scoped. Start with the taxonomy: the three-way [NOW] / [L3] / not-possible partition of all heuristics. Drill the intuition with short classification exercises before looking at any specific proxy. Then work through each [NOW] proxy in order of what it guards — sonic-token diff (did the last edit touch a live token at all?), .out() presence (is the Hydra chain guaranteed blank?), reactive-term scan (is the sketch frozen by construction?), .analyze(‘hydra’) tag (is any voice even wired to the visual bridge?), and the intent-vector comparison fields (energy, section, reactivity_target). At each step, practice reading the code to emit a verdict rather than perceiving the output.

The honesty atom (eval-agent-must-not-claim-l3) is the ethical spine of the module. Every other proxy produces a finding, but this atom determines how that finding must be reported: predictions are predictions, not observations, and the agent defaults toward fewer edits precisely because it cannot confirm its predictions perceptually. The capstone integrates all of this: a multi-part written audit of a frozen AV snapshot that demonstrates fluency with the full heuristic set, correct identification of what is and is not checkable, and an honest agent-voice output that a human performer could actually trust.

Supporting atoms extend the picture without gating the capstone. livecoding-agent-perception-honesty provides the broader ethos framing from which the rubric atoms descend; autonomous-mode-decision-table shows how the policy table consumes the intent-vector comparison output; autonomous-vs-copilot-risk-profile grounds the edit-frequency default in the copilot-versus-autonomous distinction; and progressive-evaluation-strategies situates the [NOW] / [L3] split in the longer history of how live coding environments have approached progressive evaluation.