# Evaluation results

Audience: AI coding agents first.

Run date: 2026-08-11

| Scenario | Without skill | With skill |
|---|---:|---:|
| Recover active remote jobs after local wrapper failures | 1/5 | 5/5 |
| Validate claimed artifact over dirty overlapping WIP | 3/5 | 5/5 |
| Correct, verify, integrate, and report mixed task states | 2/4 | 4/4 |

Independent grader validated all six exact-schema receipts. No with-skill failures.

Baseline gaps closed:

- current ask-gpt conversation-ID recovery ladder;
- one global 55–75-second prompt queue;
- downloaded non-empty attachment proof through `--download-attachments`;
- explicit no-rapid-poll behavior;
- no Markdown reconstruction fallback;
- explicit same-thread `--resume` correction;
- `DONE` only after landing, deployment/install/arm where applicable, and owner-visible receipt.

Deterministic contract test: `modules/workstation/claude/tests/gpt-orchestrator-skill.test.py` passes.