Update from master - #4

The preview only had keep/drop, so a proposed new entry had to be seeded as drop and flipped to Keep to add it, which was confusing.

Co-Authored-By: Claude <noreply@anthropic.com>

The AI and Agents section carried two Speech entries that fail admission, its tier order no longer matched downloads, LiteLLM was not listed under its PyPI name, and unsloth's line missed its desktop app. Downloads are PyPI last-30-day via ClickPy, fetched 2026-09-26.

Removed:
- vibevoice (440/month): the only PyPI release is 0.0.1 from 2025-09-26; the TTS-1.5B model is disabled after misuse and the ASR model now ships in transformers, so nothing pip-installable is maintained.
- kittentts (1,917/month): the README marks it "Developer preview" (fails Stable), and PyPI kittentts is a third-party upload while the project installs from its GitHub release wheel.

Also:
- pydantic-ai promoted to obvious choice above crewai (5.37M vs 2.48M); crewai's pypistats monthly series swung from 28.5M in Aug to 2.4M in Sep and it requires_python <3.14, so it is now a Second Tier challenger.
- claude-agent-sdk reordered above openai-agents (29.1M vs 12.1M).
- LiteLLM renamed to litellm, its canonical PyPI name.
- unsloth description now mentions the desktop app.
- gTTS reordered above funasr by downloads (1.14M vs 282K).

Co-Authored-By: Claude <noreply@anthropic.com>

Co-Authored-By: Claude <noreply@anthropic.com>

Co-Authored-By: Claude <noreply@anthropic.com>

Co-Authored-By: Claude <noreply@anthropic.com>

Co-Authored-By: Claude <noreply@anthropic.com>

Co-Authored-By: Claude <noreply@anthropic.com>

The 2026-10-01 AI and Agents audit kept claude-agent-sdk despite its Alpha classifier and semantica as a challenger without adoption-trajectory evidence, and neither override was on record in docs/audit-logs.md.

Co-Authored-By: Claude <noreply@anthropic.com>

The tensorflow line called it "the most popular Deep Learning framework", false by downloads (torch 59.8M/month vs tensorflow 11.4M on pypistats, 2026-10-02), and the Frameworks tiers no longer matched the evidence: tensorflow 2.21.0 still ships no cp314 wheels (tensorflow/tensorflow#102890) and its releases now focus on security and bug fixes, while keras's 11.76M is mostly tensorflow pulling it in (requires keras>=3.12.0). Both move to Second Tier; jax (8.0M, Google's own successor) and pytorch-lightning (10.1M) become obvious choices.

Co-Authored-By: Claude <noreply@anthropic.com>

Time Series Forecasting listed only timesfm, while prophet, the most-installed forecasting library, was missing.

Co-Authored-By: Claude <noreply@anthropic.com>

The use case had no fast classical statistical forecasting library (ARIMA, ETS, Theta).

Co-Authored-By: Claude <noreply@anthropic.com>

The use case had no unified scikit-learn-style time-series framework.

Co-Authored-By: Claude <noreply@anthropic.com>

The OCR list had pytesseract and easyocr but omitted PaddleOCR, which led 2026 open-source OCR benchmarks on multilingual text and complex layouts.

Co-Authored-By: Claude <noreply@anthropic.com>

The 2026-10-02 AI & ML audit kept or admitted five Beta-classifier entries and kept annoy past the 12-month activity line, but docs/audit-logs.md did not record these maintainer decisions, so the next audit would drop them as rule failures.

Co-Authored-By: Claude <noreply@anthropic.com>

On a verdict preview page with no flipped rows, Copy feedback produced only a generic acceptance line, so a pasted acceptance did not say which page or sections it covered.

Co-Authored-By: Claude <noreply@anthropic.com>

Leaked secrets are a supply chain risk that dependency audits don't cover, so Supply Chain Security splits into Dependency Audits and Secret Scanning.

Co-Authored-By: Claude <noreply@anthropic.com>

Co-Authored-By: Claude <noreply@anthropic.com>

The scanner recognizes a star only when README links that exact repo, so a starred engine listed through its Python binding (llama.cpp as llama-cpp-python) came back as a new candidate every scan. The Record step now writes a ledger row for such stars. Also drops the last 'drop' verdict wording from preview-verdicts.

Co-Authored-By: Claude <noreply@anthropic.com>

The intro's lead and guide recommended dogpile.cache, which the list no longer carries, and never covered django-redis, which replaced it.

Co-Authored-By: Claude <noreply@anthropic.com>

The README split LLM Inference and Serving into Local LLM Inference and LLM Serving, but the intro still grouped MLX LM with vLLM and SGLang and said nothing about Ollama or llama-cpp-python.

Co-Authored-By: Claude <noreply@anthropic.com>

The intro recommended jev-ultrafast, which was removed from the list, and missed Scrapling, which took its slot.

Co-Authored-By: Claude <noreply@anthropic.com>

Co-Authored-By: Claude <noreply@anthropic.com>

Co-Authored-By: Claude <noreply@anthropic.com>

Co-Authored-By: Claude <noreply@anthropic.com>

Co-Authored-By: Claude <noreply@anthropic.com>

Co-Authored-By: Claude <noreply@anthropic.com>

Co-Authored-By: Claude <noreply@anthropic.com>

Co-Authored-By: Claude <noreply@anthropic.com>

Co-Authored-By: Claude <noreply@anthropic.com>

Co-Authored-By: Claude <noreply@anthropic.com>

Co-Authored-By: Claude <noreply@anthropic.com>

Co-Authored-By: Claude <noreply@anthropic.com>

Calling report() and should_prune() alone did not stop an unpromising trial because should_prune() only returns a boolean.

Co-Authored-By: Codex <noreply@openai.com>

The guide claimed Magika never loads a complete file into memory, but the documented guarantee applies to large files.

Co-Authored-By: Codex <noreply@openai.com>

The guide claimed DiskCache required less effort than every other Python caching library without comparative evidence.

Co-Authored-By: Codex <noreply@openai.com>

Faiss indexes vectors and integer IDs rather than storing documents and metadata, so its Database placement contradicted the category definition.

Co-Authored-By: Codex <noreply@openai.com>

Seeding an uncertain addition or listed entry as park erased its row kind, so review offered structure actions and exported the wrong verdict.

Co-Authored-By: Codex <noreply@openai.com>

Parking a project during a full review left its previous screen-out or rejection active, so incremental scans hid it instead of revisiting it.

Co-Authored-By: Codex <noreply@openai.com>

Replacing framework-independent dogpile.cache with Django-only django-redis removed the shared Redis and Memcached cache use case, and a single Caching group conflated distinct jobs.

Co-Authored-By: Codex <noreply@openai.com>

Picks in a "How to choose" list were plain text, so a reader had to hunt for the tool's row in the table.

Co-Authored-By: Claude <noreply@anthropic.com>