# GPT-6 Astra replacement v1 run

- Date: 2026-09-28.
- Dispatched model: `gpt-6-astra`.
- Dispatched reasoning: **medium**, for both conditions.
- Scope: all five briefs in [../prompts.md](../prompts.md), with five baseline pages here and five skill pages in `skills/`.
- Isolation: separate fresh agents for baseline and skill; no inherited conversation and no reading earlier implementations. The five briefs within each condition share a session. Shared filesystem access is not technically restricted.
- Skill: [../benchmark/skills/anthropic-frontend-design-v1/SKILL.md](../benchmark/skills/anthropic-frontend-design-v1/SKILL.md), SHA-256 `d91970639e9f5c37682ac7ab60094d35f1c7c1f38d731bd56396563aee10c1d3`.
- Provenance: [baseline-run.json](baseline-run.json) and [skills/run.json](skills/run.json) describe the generating agents' work. The coordinator records requested runtime settings independently in [../benchmark/v1-2026-09-28/run.json](../benchmark/v1-2026-09-28/run.json).
- Validation: see [../benchmark/v1-2026-09-28/checks.json](../benchmark/v1-2026-09-28/checks.json) for independent common smoke checks. These do not establish complete business-flow correctness or score design quality.
- Repairs: any check-driven repairs and preserved initial files are recorded under `../benchmark/v1-2026-09-28/`.
- Limits: one sample per cell; no reported token usage; applicable personal instructions differ by model. Earlier providers' entries were not rerun. No ranking or statistical skill effect is claimed.

This run replaces the previous v1 implementations at the same URLs. Full protocol: [../benchmark/v1-2026-09-28/PROTOCOL.md](../benchmark/v1-2026-09-28/PROTOCOL.md).