# OpenCode x gpt-5.6-terra - eight-prompt four-take SVG series

_Run date: 2026-09-06._

- **Harness / model:** OpenCode x `gpt-5.6-terra` (pair `opencode/gpt-5.6-terra`).
- **Method:** eight independent temporary workspaces; a fresh agent made v1 without
  seeing files or future takes, then fresh agents separately made v2, v3, and an
  animated version from the immediately preceding artifact. Each agent opened its
  completed SVG. Headless Chrome rendered all 32 imported artifacts.
- **Raw:** [`../../logs/opencode/gpt-5.6-terra/`](../../logs/opencode/gpt-5.6-terra/)
- **Prompt set:** cow plowing; dolphin through a hula hoop for a trainer's fish;
  picnic on clouds after a hot-air-balloon landing; rubber ducky in a soapy bathtub;
  fruit stall; half-buried treasure barrel; stunt car through a circle of fire; and
  the recovered animated dawn-on-the-beach request with the `Qwen 3.8 can draw`
  plane banner.
- **SVGBench:** 28 artifacts map to q0, q4, q5, q6, q7, q12, and q13; mean
  **0.336**, judged by `claude-opus-5` — the single judge applied to every
  artifact in the rejudged board, not this model judging its own output.
  Per-question means: q0 0.458, q4 0.25, q5 0.35, q6 0.469, q7 0.536,
  q12 0.091, q13 0.2. The four dawn artifacts are intentionally off-benchmark.
  Under the earlier setup, where this pair's own model judged all 28 of its
  verdicts, the mean read 0.421. The comparison is like-for-like: the same 28
  artifacts carry both means. Three pairs kept their artifact set across the
  rejudge — this one, `opencode/qwen3.8-27b-8bit` (5 artifacts, 0.419 → 0.499),
  and `claude/claude-opus-4-8` (1 artifact, flat at 0.714). This is the only one
  that fell. The other pairs on the board gained artifacts at the same time
  (`claude/claude-opus-5` 1 → 21, `opencode/qwen3.8-flash-next` 8 → 25) or are
  new (`claude/fable-5-1`), so their means are not comparable across boards. See
  [`../svgbench/verdicts/opencode/gpt-5.6-terra/`](../svgbench/verdicts/opencode/gpt-5.6-terra/).
- **SVGBench:** 5/7 (q7) · verdict: `../svgbench/verdicts/opencode/gpt-5.6-terra/dolphin-hula-hoop-fish-v2.json`

The picnic filenames use `balloon` while the pinned question uses the source typo
`ballon`; `results/svgbench/manifest.json` therefore records `question_index_override: 5`
for the four picnic artifacts.