# Test evidence and limits

## Version 0.3.4 — recognition and tool output safeguards

Final release regression: **183 Python tests passed in68.664s; 13 Node interface/package tests passed**, exit0. Includes pagination bounds, full cue text/detail provenance, CPU float32/no-previous-context contract and non-destructive repetition flags. Real recognition accuracy is a separate concern, not implied by these tests.

Authorized local small-model acquisition verified all required model files. A first5-minute reference trial with CPU/int8 and previous-text conditioning produced75 identical short cues out of76; a joint CPU/float32 + conditioning-off comparison removed the runaway repetition, but terms and numbers remained erroneous. This is not an accuracy certificate or proof that either changed parameter alone caused the improvement. Separate level-adjustment testing left original audio untouched and was not made a plugin default.

Segment results now page overview grids (default4/max16), keep all cue text and times, and link complete word evidence instead of overflowing tool output. ASR records effective decoding parameters and non-destructive repetition alarms; quality_status always requires_review. VAD disabled does not guarantee all spoken words were recognized, since decoder no-speech rules can skip windows.

## Version 0.3.3 — real-course reliability

Release staging full regression: **174 Python tests passed in30.477s; 12 Node interface/package tests passed**, exit0. Tests cover nine-segment allocation under candidate caps, short-tail/tiny budgets, source anchors, retained ASR raw timestamps, valid expanded cue overlaps and strict invalid-time rejection. Unit tests use synthetic media and mocked inference; they do not establish ASR recognition accuracy.

The normal plugin manager installed0.3.3 from its TGZ and reported application=applied. Existing GUI status and actual doctor confirmed0.3.3/six tools, configured isolated Python, and local dependency availability. Runtime file filtering now retains bounded_http.py. Authorized actual fetch ofBV1jy4y1Y7mi first part succeeded again at852×480/2497.216s with audio and no manual subtitle. A previously cached base model was validated and loaded using CPU/int8 with CTranslate24.4.0 and setuptools80.10.2; the system VC++ runtime was not changed. Full-course transcription and balanced extraction are undergoing live verification; do not infer full content understanding from these tests.

## Prior installed development version
The predecessor 0.3.0 passed 127 Python tests + 8 Node contract tests on Windows, using synthetic media, mocked network responses and real yt-dlp/FFmpeg native pipelines. Actual DSH tool registration, Skill loading, doctor, prepare, segment and detail were smoke-tested on synthetic media. This is a historical verification record, not a claim that every machine or website is supported.

## Patched version 0.3.2

Bounded decoding and HTTP integration added32+4 tests; development regression163 Python tests passed (22.076s), and release staging163 Python +9 Node tests passed (22.683s for Python). Authorized live Bilibili first-part acquisition succeeded with original852×480 video plus audio, duration2497.216s,58,162,505bytes. Manual Chinese subtitles were unavailable. Model actually read an introductory grid and a lesson frame at120.004s, not the whole course; no ASR model was configured. These live results are separate from offline tests.

## Prior distribution version 0.3.1

Release staging tests verified on the reference Windows environment: **127 Python tests passed (23.816s) with the optional pinned fetch dependency available; 9 Node contract tests passed**. Without optional yt-dlp, 88 Python test entries completed with 4 acquisition-suite skips (20.431s); that run is not counted as full acquisition verification. No real platform download or actual Whisper recognition was performed.

The release includes test source files but excludes generated media/artifacts. Tests create their own temporary synthetic media on the test machine. Optional acquisition suites skip if yt-dlp is not configured; core Pillow is required. A skipped acquisition suite is not a passed download test.

```powershell
node --test tests/test_plugin.mjs
& '<Python executable>' -m unittest discover -s tests -p 'test_*.py' -v
```

Set the same interpreter used by the plugin. To exercise acquisition without real network requests, explicitly set up the pinned fetch dependency first. Real Whisper transcription remains unverified without a local model; mock tests validate API and timestamp behavior only.

No test initiates a real Bilibili/YouTube video download. Live site access, full course restoration quality and model recognition quality require additional authorized evaluation.

Tested reference environment: Windows, DSH Desktop0.2.1-rc.3, Node24, Python3.12.14, Pillow12.3.0, FFmpeg/FFprobe9.0.1 full, optional yt-dlp2026.10.26. No promise is made about older versions. Output directories must use local NTFS and not traverse reparse points.

Package verification additionally checks file allowlists, no personal configuration/backups/media/models, no absolute author machine paths, valid package entry/patch/Skill, no install lifecycle scripts, and archive hashes. Packaging is not a security certification or plugin-market approval.