peanut-gallery

The peanut gallery for AI outputs. Multiple models comment, grade, and one-up each other in a GAN game for many uses.

The "peanut gallery" is a concept from vaudeville — a vocal, opinionated audience that heckles from the cheap seats. In AI, this becomes:

Multiple models watch the same output, then try to outperform each other by finding edges and breaking each other's paradigms.

The pattern

[Target output] → Model A produces v1
                → Model B critiques v1, produces v2
                → Model C critiques v2, produces v3
                → ...
                → JEV + JEPA vote on which version is best
                → Winners are canon, losers are archived

The substrate cell is the unit

Each round of the gallery is a substrate cell:

• cell_id: unique
• prev_hash: chain to previous round
• producer: which model made it
• critic_feedback: what the gallery said
• jev_score: how JEV scored it
• jepa_score: how JEPA scored it

Use cases

• Creative writing: models compete to write the best poem
• Code review: models find bugs in each other's code
• Image description: models describe images better than each other (extends JEV-Diffusion)
• Strategy games: models play each other at games
• Math proofs: models find errors in each other's proofs

Files (coming)

• gallery.py — the main loop
• models.py — model registry
• critic.py — what the gallery says
• voting.py — JEV + JEPA voting
• examples/ — pre-run competitions