TEST_RESULTS.md

# Verification report — EIM+ Real Verification Engineer

Run date: 2026-10-09. Results below are from the local build environment; they do not imply that a Hugging Face Space or a live model API was exercised.

## Tests that actually completed

| Test | Result | What the result establishes |
|---|---:|---|
| Python byte-compilation of project and test modules | PASS | No syntax/byte-compilation errors in these files. |
| `python app.py --selftest` | 13/13 | Core engine, real child-process verification, failure handling, timeouts, policy denial, duplicate-candidate handling and backend selection smoke checks. |
| `python eim_plus.py --selftest` | 67/67 | EIM+ repair-loop, holdout, mutation check, memory/repair-log, retries, ZeroGPU budgeting and cache checks. Exit code was 0. |
| `python chat_app.py --selftest` | 28/28 | Chat/EIM routing, streaming, file context, correction prompt use, report generation and working-code tracking with deterministic test doubles. |
| `python eval_eim.py --selftest` | 7/7 harness checks | Evaluator sanity check with scripted model output; **not** a live-model benchmark. The intentionally poor scripted model's per-task `FAIL` lines are expected and are asserted by the self-test. |
| `python test_terminal_verify.py` | 12/12 | Actual OS subprocesses: stdout/stderr/exit capture, wrong output, exceptions, timeout/process-tree termination, path/safety gates and JSON evidence. |
| `python test_correction_memory.py` | 11/11 | Local correction persistence, deduplication, corrupt-line recovery, bounded file size, and an optional Hugging Face adapter exercised against a fake client. The remote adapter checks are mocked. |
| `python test_eval_benchmarks.py` | PASS | 25 unique benchmark tasks; all 150 reference assertions passed **in-process**. This validates the benchmark references, not model-generated answers. |
| Gradio UI construction | PASS | `build_ui()` successfully constructed a `Blocks` app in local Gradio 6.5.1 and exposed the compatible launch options. The server was not launched on Hugging Face. |
| Simulated ZeroGPU backend selection | PASS | With `SPACE_ID` and `SPACES_ZERO_GPU=1`, `app.py --backend-check` selected `local`. No model weights were loaded in this check. |
| ZIP archive integrity | PASS | `unzip -t` on the delivered ZIP reports no errors. |

## Not verified here

- Live generation from Qwen or any other model: `transformers` and `spaces` are not installed in the local test environment and no valid `HF_TOKEN` is configured.
- A real before/after coding-accuracy benchmark against the same live model/settings: **not run; conclusion is inconclusive**.
- A real push/pull against a private Hugging Face Dataset: adapter tests use a fake client; real token scopes, repository permissions and network sync must be verified in the deployed Space.
- Security-grade isolation: the subprocess verifier has time/resource limits and policy checks but is not an OS sandbox. For public untrusted code, use a disposable container/VM with filesystem restrictions and network disabled.

## Recommended deployment check

After pushing to your eligible ZeroGPU Space, inspect build/runtime logs, run a simple assertion task, then configure a private Dataset repository and a write-scoped `HF_TOKEN` if correction/repair memory must survive Space restarts. Hugging Face's current ZeroGPU eligibility and quota rules are documented at https://huggingface.co/docs/hub/en/spaces-zerogpu.