A-Share Evidence Gate: Point-in-Time Research for AI Agents

Where this fits

You're building an agent that answers "What was the revenue growth for 贵州茅台 as of March 2023?" or "Did the Q3 report beat expectations when it was actually disclosed?" Without strict point-in-time discipline, your agent will mix future restatements, wrong announcement dates, or unverified Tushare snapshots into its reasoning. This skill forces every downstream claim to pass through an evidence ledger first — so before your agent computes a valuation or writes a summary, it must prove what was knowable at that exact moment and where each number came from.

Why agents benefit

• Stops data contamination: Enforces as_of + require_pit=true for historical queries; any evidence dated after the research cutoff is flagged as forbidden_future and blocks the entire manifest.
• Builds auditable evidence trails: Each record carries source URL, announcement timestamp, report period, unit, currency, and page/section locator — so every number can be traced back to the original 巨潮 or exchange filing.
• Detects silent conflicts: Automatically compares merged vs. standalone, single-quarter vs. cumulative, TTM, and restated vs. original figures. Unresolvable conflicts are recorded side-by-side, never averaged.
• Prevents hallucinated disclosure times: Distinguishes report period, planned disclosure date, and actual publication moment — a key failure mode for agents that treat 年报日期 as the day data became public.
• Generates a structured verdict: Outputs ready, partial, or blocked status plus a safe-to-use field list, so downstream analysis only touches verified numbers.

Practical scenarios

Backtesting a factor based on historical financials

Your quant agent pulls ROE from AKShare for a 2018 backtest. Without point-in-time logic, it uses 2023 restated data. This skill forces all 财务数据 to be visible only after ann_date or f_ann_date, and keeps original restatement versions — your factor no longer leaks future information.

Building a company research brief for a specific date

You ask for a snapshot of 宁德时代 as of June 2024. The agent must gather quarterly reports, announcements, and analyst consensus — but only those actually published before that date. The evidence manifest marks any post-cutoff source as forbidden_future, so the brief reflects only the information landscape of that day.

Resolving conflicting numbers across sources

Tushare says Q3 net profit is 12.4亿, while a news article claims 12.8亿. The skill checks merger scope, tax treatment, and announcement timestamps. If the conflict can't be resolved, both values are listed separately with their sources — never merged into a misleading average.

Add it to your agent workflow

Install the skill and add it as a pre-processing step for any A-share research task:

npx skills add cyijun/china-financial-services --skill a-share-research-evidence

Then in your agent prompt:

Use a-share-research-evidence to build a point-in-time evidence pack for 贵州茅台 as of 2024-03-31. Research question: "What was the 2023 annual revenue and net profit?" Output the evidence manifest and a ready/partial/blocked status.

The skill will generate a JSON manifest via scripts/build_evidence_manifest.py, download original PDFs (read-only, with SHA-256), and verify each record against the as_of cutoff. See more details on the OpenAgentSkill skill page.

Compare before adopting

Before committing, compare this skill against:

• Quality signals: Does it enforce original-source verification (交易所/巨潮) over convenience APIs? Does it handle restatements and timezone-aware timestamps?
• Maintenance freshness: The repo is small (18 stars) — check recent commits and whether the build script still works with current Tushare/AKShare APIs.
• Alternatives: Generic data-validation skills lack A-share-specific rules like ann_date visibility and 复权 handling. This one is tailored to China's disclosure ecosystem.
• Workflow fit: If your agent only does real-time quotes and never needs historical point-in-time research, this adds overhead. Use it when backtesting, compliance, or audit trails matter.

Why it is worth tracking

18 stars is early, but the design is unusually rigorous for a niche skill: it explicitly forbids search summaries as evidence, requires HTTPS original URLs, and blocks future data. Even if you don't adopt it today, its output contract — four evidence categories and a blocked state — is a good pattern to steal for any domain where data provenance and timing are critical. Watch for updates on handling more exchange file types and integration with china-market-data; that's when it becomes a must-have for production A-share agents.