Interactive clean-room workbench for search-query taxonomy discovery, classification, evaluation, and human review. GitHub source is canonical; GitHub Pages and Molab are derived, commit-backed views.
Implemented local MVP:
- deterministic browser-safe TF-IDF and latent-semantic analysis
- synthetic sample and bounded CSV upload
- cluster discovery and transparent taxonomy classification
- confidence-based abstention and HITL review
- canonicalized supplied-label metrics with explicit denominators
- session-bound, hash-chained human-review evidence
- safe CSV and JSON audit downloads with runtime and input identities
- exact browser dependencies and an auto-built local package wheel
No employer source, private taxonomies, client data, or owner-funded API calls are included. See CLEAN_ROOM.md.
uv sync --locked
uv run marimo run src/app.pyRun validation:
uv run pytest -q
uv run ruff check .
uv run ruff format --check .
uv run marimo check --strict src/app.py
uv run marimo export session src/app.py --force-overwrite --no-continue-on-error
uv run python scripts/validate_session_snapshot.py \
src/__marimo__/session/app.py.json
uv run marimo export html-wasm src/app.py \
--output output/wasm --mode run --no-show-code --execute --force
uv run python scripts/validate_wasm_export.py output/wasm
uv run python scripts/browser_smoke.py output/wasmUploaded CSV data remains in the active notebook runtime. Molab may host that
runtime; a /wasm deployment executes in the visitor browser. Neither mode
persists uploads through this app.
uv.lock is the executable dependency authority. See
DEPENDENCIES.md for direct-license evidence; release CI also
generates an SPDX SBOM and blocks critical vulnerability findings.
Status: public. Pinned CI, critical-vulnerability scan, SPDX SBOM, executable WASM build, browser analysis/review journey, Pages, and Molab routes pass for the published source.
MIT. Public fixture provenance remains separately documented.