qveris-a-share-factor-screen

QVeris A-Share Factor Screen

Use this skill to preserve Alphasift's research-screening workflow while replacing local market data packages, model-provider assumptions, and stock-picking language with QVeris structured-data evidence, audited Web news/sentiment evidence, and auditable factor notes.

Source record:

Source Adaptation

• Preserve Alphasift's core shape: strategy catalog, screen, hard filters, factor scoring, risk/source-health fields, saved-run metadata, reports, and T+N post-hoc evaluation.
• Replace original data packages and provider paths (efinance, akshare, baostock, tushare, yfinance, HTTP source fallbacks) with QVeris CAP evidence.
• Remove operational LLM-provider requirements and external analyzers from the runtime contract. Original litellm, DSA, and deep-analysis fields are migration context only.
• Suppress or rename fields that imply actions, such as operation_advice, invalidators used as trading instructions, buy/sell wording, target prices, and position actions.
• Treat strategy output as a transparent research candidate pool. Only output a rank when the same factor set, price window, fiscal period, and market convention are comparable across at least 3 securities.

Runtime Contract

• Use only qveris_finance.* CAP tools and QVERIS_API_KEY.
• Execute every finance data call through this Skill's scripts/qveris_finance_adapter.mjs, or through a native wrapper that runs the byte-identical adapter; never call /capabilities/query directly from the workflow.
• Default natural-language output to a Markdown user report, not a JSON object.
• Accept dry_run, max_calls, max_age, and budget_note; if omitted, default to dry_run=false, no hard max_calls limit, max_age=P1D, and a conservative budget note, then echo those controls.
• Read references/qveris-finance-data-quality-rubric.md before using QVeris payloads as factor evidence.
• Use references/qveris-finance-retry-policy.md for failed calls, invalid capabilities, payload truncation, and semantic mismatches.
• Build trace, call counts, retries, and timestamps only from saved observed_calls. Never invent an execution ID, retry, timestamp, per-security call, or result from the intended workflow; use execution_id=null when an observed call returned no ID.
• Sanitize every output surface, including Evidence, Sources, prose, params, responses, and Trace. Strip provider names, provider API URLs, raw route/tool IDs, candidates, failover, credentials, and routing metadata recursively; the Trace row remains exactly tool_name, params, status, execution_id, fallback_used, and missing_fields.
• Treat screening output as a research candidate pool, not as investment advice or an action list.
• Suppress target prices, upside/downside, ratings, buy/sell wording, rebalancing instructions, and trade execution plans even if present in QVeris payloads.
• Read and follow references/qveris-web-news-sentiment-policy.md. Never call qveris_finance.news_fin_tagged or qveris_finance.sentiment_text_signals; use its audited Web lane in every run mode, including benchmark and replay.