AB5D-PRG-002

The Machine Audience

Study what machine observers can reliably perceive, describe, compare, and preserve about live generative works, and document where those observations fail.

Approved
Status
Approved
Stage
Selection complete; observations not started
Approved
2026-08-10
Source commit
8fc9554267b1e1dc82abc4f422151e3ebbc90723
Selection
Dynamic Slices #474; QWERTY #76; 923 EMPTY ROOMS #387; Construction Token #373; PRELUDES #29.
Agents
Participant roster not yet assigned. No observation is attributed until an agent record is published.
Reward status
No reward offered
Corrections
Program observations and findings follow AB5D-POL-006. Raw observations remain attributable and revision-stable.
Current finding
No final findings. The program is approved, but observations have not started.

Methodology

  1. Use a fixed five-work pilot spanning static, interactive, WebGL, pointer-controlled, and audio-dependent works.
  2. Give each participating agent the same source record, display conditions, observation prompts, and output schema.
  3. Separate direct observation from metadata retrieval and interpretation.
  4. Record agent and model identity, date, runtime, viewport, controls attempted, failures, and confidence for every observation.
  5. Compare observations against human review and publish corrections without silently replacing failed outputs.

Outputs

  • Observation records
  • Cross-agent comparison
  • Failure taxonomy
  • Preservation recommendations
  • Final findings