← ClaudeAtlas

alpha-reviewlisted

Open the alpha review screen to mark judge keepers as gold or miss, read per-miner precision, and print the eval state. Use when tuning a miner's accuracy. Thin launcher over the existing harvest engine and local review server. Triggers: "/alpha-review", "alpha screen", "open the review screen".
tonydzi/second-brain-starter-kit · ★ 6 · Code & Development · score 78
Install: claude install-skill tonydzi/second-brain-starter-kit
# /alpha-review — the alpha selection screen in one command The engine is already built (2026-06-18/20). This skill is launch + summary; it duplicates nothing. ## Steps 1. **Is the server alive?** `netstat -ano | findstr :8772` (PowerShell) / `netstat -ano | grep :8772` (bash). - Listening → jump to step 3. 2. **Start it** (a fresh harvest runs inside): ``` cd /e/Obsidian/_imports/alpha && PYTHONIOENCODING=utf-8 python alpha_review_server.py --no-browser ``` in the background (`run_in_background`). Manual alternative for the operator: double-click `$IMPORTS_ROOT/alpha/alpha-review.cmd`. 3. **Hand over the link**: http://127.0.0.1:8772 — open in a browser (local only, nothing leaves the machine). 4. **Eval summary** (0 LLM tokens): ``` PYTHONIOENCODING=utf-8 python $IMPORTS_ROOT/alpha/alpha_tune.py ``` Show the operator: how much is labelled / per-miner precision / what to label first (uncertainty sampling: PARTIAL first). At >=8 labels per miner, tune names a concrete detector fix. 5. **Remind them of the loop**: labels -> `alpha_tune.py` -> adjust the detector threshold/filter -> re-harvest -> re-label. Cards carry a batch badge (🆕 = fresh nightly batch). ## Pitfalls - The DB is a cumulative INBOX of every nightly batch (not just the latest judged file) — "extra" items are not junk, they are an unlabelled backlog. Do not "fix" it. - An empty screen is not the same as no data: first check that the harvest actually ran (`alpha_harvest.py` prints