| name | qveris-a-share-factor-screen |
| description | QVeris-native adaptation of candidate 32, Alphasift. Use for China A-share universe screening, factor scoring, candidate-pool research, strategy screen review, and post-hoc evaluation that must use qveris_finance.* CAP evidence, transparent missing-data handling, and no investment advice. |
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:
| Field | Value |
|---|
| Candidate number | 32 |
| Original repository | Alphasift |
| GitHub URL | https://github.com/ZhuLinsen/alphasift |
| License | Apache-2.0 |
| Evaluation recent activity | 2026-07-03 |
| Local source snapshot | third_party/source_repos/32-alphasift |
| Snapshot latest commit | 9f52274 on 2026-07-03 |
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.
Evidence Gate
- Resolve the universe from an explicit user-supplied ticker list or an approved frozen universe file, then validate every member with
qveris_finance.ref_symbology or qveris_finance.ref_security_master.
- Never call
qveris_finance.index_constituents. If no explicit universe is available, return universe_unavailable; do not silently substitute a small proxy universe.
- For A-share requests, reject securities whose returned market, exchange, listing class, or asset type does not match the requested mainland equity universe.
- Assign every requested security a coverage tier before showing any factor values:
complete_comparable, partial_not_ranked, proxy_only, or insufficient. Show the comparable subset and per-security factor gaps explicitly.
- Require comparable windows before ranking factors: identical validated factor set, price window, fiscal period, measurement basis, and market convention. Rank only the complete comparable subset, never the full requested universe when coverage differs.
- Require at least 2 bars for simple multi-day metrics and at least the requested lookback length plus one observation for lookback indicators.
- Do not compute percentiles or ranks from fewer than 3 comparable securities; output per-name notes instead.
- For Chinese text fields, hard reject mojibake or replacement-character artifacts. Do not quote corrupted company names, industry labels, event titles, news snippets, or research titles in factor evidence; keep valid numeric/date fields only if identity and window checks pass, and mark the text fields
encoding_artifact.
- Use only opened, issuer-matched, in-window Web pages for news and qualitative sentiment, and keep their provenance in
web_trace rather than qveris_trace.
- For post-hoc evaluation, separate
as_of evidence from evaluation-window bars; never let future bars influence the screen score.
- If the original strategy requires a field QVeris cannot validate, mark that factor component missing and disclose the changed denominator.
CAP Invocation
- Standardized CAP invocation is mandatory. A missing CAP or runtime becomes
capability_unavailable or tool_runtime_missing; it never authorizes a legacy raw route.
- Use native
qveris_finance.* tools only when that runtime applies the same Skill-owned adapter and returns a qveris.finance-parameter-adaptation.v1 audit; otherwise use this Skill's CLI.
- If native tools are unavailable and the run is in this repository root, use the repository CLI:
node {baseDir}/scripts/qveris_finance_tool.mjs cap-query qveris_finance.<capability_name> --param key=value --safe-json.
- Treat the Skill-owned CLI as the mandatory finance adapter: it resolves the live canonical CAP, losslessly adapts parameters, hydrates signed full-content results, and applies the shared data-first semantic gates. The envelope
success flag is diagnostic only; record envelope_success and contract_clean separately.
- If the Skill-owned scripts are missing and no native
qveris_finance.* runtime exposes the identical adapter audit, mark tool_runtime_missing; do not use web, legacy providers, or invented data as fallback.
- Use
cap-detail before calling sensitive or uncertain routes such as classification, analyst reports, corporate events, or EOD bars.
- Keep failed, rejected, or not-called capabilities in
Data Quality And Missing Fields and the trace appendix, not in the evidence table.
Workflows
- Scope the screen: identify universe, as-of date, factor set, maximum calls, and whether post-hoc evaluation is requested.
- Resolve the universe: validate each identifier and keep unresolved or non-A-share instruments out of the scored set.
- Collect factor inputs: use bars for momentum/liquidity/volatility, statements and derived ratios for valuation/quality, classification for sector context, and sentiment/news only within their evidence limits.
- Partition coverage: produce the coverage tier, comparable subset, and per-security missing-factor matrix before scoring.
- Score transparently: calculate scores only inside a subset with the identical complete factor denominator. Do not renormalize different denominators into a cross-security rank.
- Report the screen: show coverage tiers and gaps first; add rank only for the complete comparable subset when every comparability gate and the three-security minimum pass.
- Evaluate post-hoc only when requested: fetch evaluation-window bars after the
as_of date, label the result as historical evaluation, and avoid return forecasts.
Fallback Policy
- If a primary factor CAP fails after allowed retries, mark that component missing and lower confidence; do not replace it with unsupported inference.
- If only proxy sector or news evidence is available, label the status
proxy_only and keep claims narrow.
- If cross-sectional comparability fails, avoid ranking and provide per-security evidence notes.
- If
max_calls prevents universe coverage, return a budget-limited report with the exact QVeris calls still needed.
- If a successful payload contains the wrong security, wrong window, wrong fiscal period, or truncated content, hard reject it and mark the reason.
- If a successful payload contains corrupted text fields, exclude the corrupted fields from the report body and mark
encoding_artifact; do not translate, repair, or infer the intended wording.
Output Requirements
- Use level-2 Markdown headings exactly for this user-report structure:
## Summary, ## Screen Results, ## Evidence, ## Analysis, ## Data Quality And Missing Fields, and ## Trace Appendix. Do not replace these headings with bold text.
- Include a factor table with security, validated factor values, component coverage, evidence status, and missing fields. Use rank columns only when comparability gates pass; otherwise use coverage tier or per-name notes.
- Render the Trace Appendix with the exact parseable header
| tool_name | params | status | execution_id | fallback_used | missing_fields |; use compact JSON for params and missing_fields, one row per observed attempt, and no rows for planned or budget-blocked calls.
- For live, fresh, or E2E output, save and validate an
observed_calls.v1 sidecar whose calls record request_kind=capabilities/query and canonical capability_id; without a verified sidecar, place the unverified note before ## Trace Appendix and emit only the exact header plus separator with no rows.
- Put full
qveris_trace JSON only in the appendix, schema fixture, or when the user asks for machine-readable output.
- End user-facing reports with
Not investment advice.
Prohibited Capabilities
Do not output investment recommendations, buy/sell triggers, target prices, upside/downside, rebalancing, execution instructions, automated trading, non-QVeris structured-finance data pulls, Web use outside the audited news/sentiment policy, login/cookie use, provider keys, or strategy claims unsupported by historical evidence.
References
- Use shared finance contract version
2026-07-29.1; repository CI verifies the local rubric, retry policy, CAP registry, and output schema against references/qveris-finance-shared-manifest.json hashes.
- Read
references/qveris-tool-map.md before choosing calls for an A-share factor screen.
- Read
references/qveris-finance-data-quality-rubric.md before treating any payload as evidence.
- Read
references/qveris-finance-retry-policy.md when a CAP fails, returns the wrong shape, or needs fallback.
- Read
references/qveris-web-news-sentiment-policy.md before collecting news or qualitative sentiment.
- Check
references/qveris-finance-cap-registry-snapshot-2026-07-07.md before adding a route to the primary path.
- Use
examples/default-markdown-report.md as the primary user-facing example.
- Use
fixtures/qveris/*.json as machine-readable schema fixtures only.