Every round.
The full picture.

Each round now has a named brief, a winner, a full ranking, the plan used, an output count and a rule check. Rounds 1 and 3 use the curator’s supplied winners and final rankings, added September 27, 2026. Round 2 uses a retrospective Codex assessment added September 26, 2026. Round 4 retains its original winner and ranking, with corrected compliance notes. Round 5 uses the curator’s final ranking and a prompt written by Muse (Meta AI).

What the judging considers

1. Following the brief: consider completed output, required format, titles and captions, and that round’s subject or story requirements.
2. Creative effectiveness: compare clarity, composition, emotional impact, variety and originality. For Round 4, include continuity, readable action and the resolution.
3. Explain the tradeoffs: a ranking balances instruction-following and artistic success. No numerical scoring or automatic disqualification system was published for these runs, so none is presented as an original rule. A winner can still have a documented weakness.

Model names and plans

Entries use the service name and recorded model version where available: ChatGPT — GPT-6 Astra Light and Claude — Opus 5.5. Meta AI, Google Gemini and Grok retain their service names because exact model versions were not recorded. Round 5 labels the Meta entry “Meta AI — Muse” to identify the assistant experience tested; no unrecorded underlying model version is implied. Subscription plans are listed separately: Free plan, Paid plan, ChatGPT Plus or Claude Pro. Prices describe the plan used for that round, not current pricing. Round 1’s exact paid tier is not specified in the curator’s updated result; Round 2’s free ChatGPT model version was not recorded.

What the review can and cannot establish

The curator’s verdicts for Rounds 1 and 3 are presented as supplied. Codex’s compliance annotations are separate from those artistic verdicts. Codex (OpenAI) reviewed the archived artwork, image dimensions, supplied text and existing process notes at the owner’s request. The judgments are editorial, not objective measurements or an independent blind panel. OpenAI is an entrant, so that relationship is disclosed. Visual checks cannot prove originality or reconstruct delivery order, retries or missing chat logs.

Round 1 used a broader website-inspired brief with extra guidance. Its exact original prompt is not reproduced in the archive. It is ranked on the creative work and recorded brief; the later square-image, no-text and six-word requirements are not retroactively treated as rules it originally had.

Rounds 2 and 3 retain Claude’s exact published prompts. Round 4 retains Astra’s exact published prompt. The shared requirements include six separate square images and text outside the pictures. Rounds 2 and 3 ask for a title of six words or fewer and a one-sentence caption for each image. Round 4 asks for one story title of six words or fewer plus six captions, consistent characters and the specified scene order.

Code-generated art remains a valid medium. Paid plans receive no bonus. Plan limits, revisions, prompt authorship and prior exposure to other entries are disclosed because they affect comparability. Abstract forms that evoke landscapes or objects are flagged as interpretive cases rather than asserted to pass without qualification.

Preserving the evidence

Original images, model titles and captions stay unchanged. “Untitled” and scene or feeling labels are archive labels, not newly invented model titles. Missing output stays missing. New review text is separate from the original submissions, and the Round 1 original verdict remains available below its new results.