news-scalping-lab

name	news-scalping-lab
description	Workflows for research import, brain rebuild, blind daily analysis, evaluation, and audits in this repository.

news-scalping-lab Skill

Use this skill for:

• research episode import
• brain update or full rebuild
• brain audit
• daily blind analysis
• postmortem evaluation
• lookahead leak audit
• hardcoding audit

Core Rules

• Do not add stocks, tickers, themes, regions, or beneficiary mappings to source code.
• Research knowledge belongs in research/, memory/, and brain/.
• Exact keyword retrieval is never a candidate gate.
• Blind inference cannot use D-day prices or cutoff-after evidence.
• Exhaustive mode must include every accepted episode in the context manifest.

Commands

Initialize:

nslab init
nslab doctor

Import and accept research:

nslab research import path/to/research.md
nslab research validate <episode_id>
nslab research accept <episode_id>
nslab brain rebuild --mode llm-full
nslab brain audit

Daily blind analysis:

nslab analyze --news path/to/news.csv --trade-date YYYY-MM-DD --cutoff YYYY-MM-DDT08:59:59+09:00 --mode exhaustive --web-search

Evaluation:

nslab evaluate --trade-date YYYY-MM-DD

Audits:

nslab audit hardcoding
nslab audit lookahead --trade-date YYYY-MM-DD
nslab audit provenance
nslab audit coverage

Expected Outputs

• predictions/YYYY-MM-DD.json
• reports/YYYY-MM-DD_preopen.md
• runs/manifests/<run_id>.json
• brain/current/brain_manifest.json
• brain/current/coverage_manifest.json

Quality Gates

python -m ruff check .
python -m mypy src/news_scalping_lab
python -m pytest
make full-check

Recovery

• If brain audit fails, run nslab brain rebuild --mode llm-full.
• If exhaustive analysis reports coverage errors, check research/accepted/ and brain/current/coverage_manifest.json.
• If hardcoding audit fails, move domain knowledge out of source code and into research or memory data.
• If lookahead audit fails, inspect the manifest price_snapshot.allowed_through and cutoff-after web exclusions.