the write-up

Findings

Cross-experiment lessons, taken from the run logs rather than theory. The raw records are in data/.

One frozen Krea2-turbo workflow, one dumb Python runner, three loop topologies. Everything below was observed in the logs (\1), not theorized.

Headline results

stage	question	answer
01_hillclimb	can agent prompt-mutation + fixed rubric climb reliably?	yes: 27→44/50 in 10 iters, plateau detector fired correctly
02_morph	can an i2i chain travel studio→highway holding identity?	yes: arrival in 7 clean beats, identity 10/10 throughout
03_evolve	does population search beat single-lineage refinement?	yes: 48/50 vs 44/50, +4 in 5 generations (20 images)

Seed policy depends on loop topology (the biggest finding)

\1 Every visual change is then attributable to the prompt edit. Non-negotiable for judging the judge.
\1 Fixed seed re-applies the identical noise pattern each step; artifacts correlate and lock in (our speckle storm, frames 0–3). One seed rotation broke a two-frame stall instantly.
\1 A seed reroll alone matched a 44 champion and deleted its seed-locked defect (gen 0, n2). Treat reroll as a mutation operator.

Artifact recovery in i2i chains

\1 Denoise defends existing pixels, junk included. Our frozen "drop to 0.30 to recover" policy made speckle categorically worse in one step.
What works, in order of preference: (1) roll back to the last clean ancestor and re-branch, (2) repaint at 0.50+ with fresh seed, (3) full chain restart. The car body always self-repaired before the background did — the model defends the subject, junk survives in low-attention regions.
Artifacts trace to prompt wording: "floor beginning to lighten" birthed the wet speckle. Transitional/ambiguous surface language is a collapse seed.

Layout attractors: relabel, don't fight

Compositions with strong structure (dark surround + bright center: our garage doorway, then the bridge) survive any denoise you can afford. Negating them ("no overpass") does nothing. What works: \1 — garage mouth → highway underpass mouth, in one 0.55 step, best beat of the run. Describe what things ARE, not what should disappear.

Prompt physics: cause, not effect

"fine metal flake catching the light" → first 10/10 paint (the model has a physical reason for rich speculars). WON the evolution run.
"reflections modeling every curve" / "a single sweep" → jagged, calligraphic junk, twice (n13, n17 in 03_evolve; n9 in 01_hillclimb). Art-directing the exact reflection shape over-constrains; describing material/light-source physics lets the sampler solve it.
Style keywords move composition, not just quality: "professional advertising photograph" redesigned the entire car (hillclimb n6). Budget for that.
Surface-quality clauses can hijack framing: "smooth uncluttered hood" pulled the camera high and tight, twice.

Loop mechanics that proved out

\1 (agent writes step JSON, dumb runner executes, agent judges files): zero failures across 42 generations, fully agent-agnostic.
\1 kept 10-iteration score curves comparable; the plateau rule (spread < 2 over last 4) fired exactly when marginal gains died. Policy errors were handled as logged amendments, never silent edits.
\1 made evolution monotonic — fitness never regressed in 5 generations.
\1: both 46 and 47 champions were crossovers, fusing independently discovered wins (chrome grille line × clean hood line × low angle).
Judge blind spots to watch: brand-identity drift (the model relaxes toward real production designs — BMW/Maserati/Aston) is invisible to this rubric.

Numbers worth remembering (5090, Krea2-turbo fp8, 1664×928)

~7.5–8s per image; 8 steps, cfg 1, er_sde/sgm_uniform. Hillclimb run ≈ 9 min, morph ≈ 2 min of generation, full 20-image evolution ≈ 3 min. Iteration is effectively free — design experiments accordingly.