# Codex HTML animation reference runs — 2026-09-16

Three first attempts, one per requested reasoning effort, generated for the homepage comparison. The exact prompt is in `prompt.txt` (Pelican Test `pelican-bicycle-cli-v1`); it differs from the historical static SVG prompt. No output was redrawn, selected from retries or edited. Original HTML, its SHA-256, the prompt hash, requested settings and individual static analysis reports are preserved here.

| Requested model | Requested effort | Static score / 100 |
| --------------- | ---------------- | ------------------ |
| gpt-6-astra     | medium           | 81                 |
| gpt-6-astra     | high             | 90                 |
| gpt-6-astra     | xhigh            | 84                 |

Codex CLI 0.154.0 ran three concurrent, independent ephemeral sessions in read-only temporary directories. The command disabled `memories` and `multi_agent` and set `approval_policy="never"`; other user/runtime configuration was inherited. The prompt prohibited tools; event logs reported no completed command/MCP/web-search calls. Requested model names are not independent verification of a backend snapshot. This is a three-sample observation, not evidence of degradation or a general effort ranking.

PNG previews were captured at 1200 × 900 in a Chromium opaque sandbox 1.5 seconds after loading each HTML document. Network requests were blocked. They are screenshots of the actual HTML, not redrawn illustrations.

Analysis used the published `@ryqdev/pelican-test@0.1.0` CLI, ruleset `html-svg-heuristic-v2`:

```sh
npx --yes --registry=https://registry.npmjs.org @ryqdev/pelican-test@0.1.0 analyze docs/benchmarks/2026-09-16/medium.html docs/benchmarks/2026-09-16/high.html docs/benchmarks/2026-09-16/xhigh.html --json
```

The current homepage uses these references until a matching-effort, published community result is available. Community scores remain self-reported. References and community uploads preserve separate provenance.