docs: regenerate creative-pipeline media with gpt-image-2

flux/schnell was chosen on cost — $0.003 per megapixel — and it was the wrong
trade. It mangles hands, and this brief requires them in three of four shots. A
model that cannot draw the thing being commissioned is not cheap. Stills now
come from openai/gpt-image-2; clips stay on LTX-distilled, which was never the
problem. Both are overridable with CREATIVE_IMAGE_MODEL / CREATIVE_VIDEO_MODEL.

Regenerated from the prompts recorded in the run dump rather than by re-running
the agent, so this cost image and video calls and not another loop.

The swap introduced a different brief violation, which is the more interesting
half. flux could not render legible text, so it never breached the brief's
no-logo rule by accident. gpt-image-2 can, and branded the bottle in shot 3
with raised COLD BREW COFFEE lettering. Better capability, new failure mode.

The prompt is why: sonnet never carried no-logo into any image prompt. It read
the brief and the constraint did not survive into the generation call. Shot 3
was regenerated with it appended.

Neither violation is visible to the suite. forbiddenAvoided reads shot action
text and layer names, so it can see the word logo in a prompt and never what
the picture contains; mustFeature asks only whether each element appears
somewhere in the shot list, so a shot with neither hands nor sunrise passes.
The predicates grade the plan. Grading the artifact needs a human or a vision
model, and that is now written down in both the fixture README and the package
docs rather than left for someone to discover.

Co-Authored-By: Claude <noreply@anthropic.com>

Author: georgi

docs/evals/creative-pipeline/README.md

@@ -26,9 +26,12 @@ measured it, trimmed it, and signed it off — 162 tool calls, delivered at
 | `keyframe-shot_{1..4}.webp` | Storyboard — one still per shot |
 | `clip-shot_{1..4}.mp4` | Storyboard — each still animated |
 
+Stills come from `openai/gpt-image-2`, clips from
+`fal-ai/ltx-2-19b/distilled/image-to-video`.
+
 Compressed for the repo: stills to WebP at 720px, clips to h264 at 540px.
-15 MB of raw output became 1.3 MB, which is the difference between a fixture
-and a liability. The originals are not kept.
+Raw output became 1.8 MB, which is the difference between a fixture and a
+liability. The originals are not kept.
 
 ## Reproducing
 
@@ -45,14 +48,39 @@ Roughly **$2.60** for the agent loop and **$0.17** for the media —
 $0.0008. The model driving the pipeline costs an order of magnitude more than
 the pipeline's output.
 
+## Why not a cheaper image model
+
+The first pass used `flux/schnell` at $0.003 per megapixel, chosen on cost. It
+was the wrong trade: flux mangles hands, and this brief requires them in three
+of four shots. A model that cannot draw the thing being commissioned is not
+cheap. gpt-image-2 costs materially more per image and renders anatomy that
+survives a look.
+
+Swapping it in introduced a *different* brief violation, which is the more
+interesting half. flux could not render legible text, so it never breached the
+brief's `no logo` rule by accident. gpt-image-2 can, and branded the bottle in
+shot 3 with "COLD BREW COFFEE" in raised lettering. Better capability, new
+failure mode.
+
+The prompt is why. Sonnet never carried `no logo` into any image prompt — it
+lives in the brief, and the model read the brief, but the constraint did not
+survive into the generation call. Shot 3 was regenerated with the constraint
+appended explicitly.
+
 ## Two things these files show that the eval cannot
 
 **The planted defect is conservative.** The suite simulates rendered clips
 coming back 1.35× longer than requested. LTX returned 4.84s takes for 3s
 requests — **1.61×**. Real video models overshoot harder than the eval assumes.
 
-**A per-shot brief violation the predicates miss.** `clip-shot_3` (condensation
-on a bottle) contains neither hands nor sunrise. The model caught this in its
-own review and filed it; the suite's `mustFeature` check only asks that each
-element appears *somewhere* in the shot list, so it passed. The model's
-judgement was finer than the check's.
+**Two brief violations the predicates cannot see.** `clip-shot_3` contains
+neither hands nor sunrise; the model caught that in its own review and filed
+it, while the suite's `mustFeature` check passed because it only asks whether
+each element appears *somewhere* in the shot list. And the same shot came back
+branded, which `forbiddenAvoided` also passed — that check reads shot action
+text and layer names, so it can see the word "logo" in a prompt and never what
+the picture actually contains.
+
+Both are the same limit. The predicates grade the plan; only a human or a
+vision model grades the artifact. Worth knowing before trusting a green run to
+mean the deliverable is on brief.