`zhengxi-views`: How One Repo Turns a Fund Manager’s Public Mind Into an Auditable AI Skill
A traceable investment research agent built from public statements, fund holdings, and scoring rules that force the model to show its work instead of hallucinating answers.
- `zhengxi-views` is interesting because it treats investing as a verification problem, not a chat problem.
- The repo cross-checks what Zheng Xi said against what funds actually held, which makes contradiction detection the core feature.
- Its scoring logic turns imperfect public data into a usable style proxy, which is more useful than pretending exactness it does not have.
- The bigger pattern is a specialized analyst skill that stays inside a documented corpus, a methodology file, and a narrow evidence model.
The real product is not answers. It is verification. `zhengxi-views` is built to stop a model from freelancing. It retrieves public statements, pairs them with holdings data, and forces the assistant to answer inside a constrained evidence frame.
可溯源的郑希(易方达基金经理)投研 Agent Skill——基于他全部公开观点原文 + 有原话佐证的投资方法 + 全市场基金真实数据，能溯源问答、按他框架给基金打分，绝不杜撰。⚠️仅研究学习辅助，不构成投资建议‼️website是郑希主页！
A fund manager becomes a machine-readable persona
The subject is not a generic investor. It is Zheng Xi, treated as a single, coherent research object. The repo gathers public remarks, quarterly reports, and fund behavior over a long span, then turns them into something an agent can query without losing the original source trail.
That matters because finance chatbots usually flatten people into vibes. This project does the opposite. It tries to preserve the shape of one manager’s thinking and keep the model accountable to it.
The repository is built as data, logic, and prompt
The structure is simple on purpose. `references/corpus/` holds the raw language, `references/fund_data/` holds structured holdings snapshots, and `scripts/` handles search and scoring. `SKILL.md` and `skill.yml` sit on top like policy, telling the model what it may say and what it must cite.
The clever part: words vs. deeds checking
| Question | Generic LLM | zhengxi-views |
|---|---|---|
| Source of truth | Loose prompts and model memory | A documented corpus and holdings data |
| Can cite evidence? | Sometimes, but inconsistently | Yes, by design |
| Can cross-check claims against holdings? | Not reliably | Yes, this is the point |
| Can simulate one manager’s logic? | Only with risk of drift | Yes, inside a constrained methodology |
| Main risk | Hallucination | Overconfidence in weak evidence |
That is the conceptual upgrade. Most tools are built to answer quickly. This one is built to answer only when the evidence frame is strong enough.
How the scoring hack turns portfolio drift into signal
def turnover_proxy(quarters):
# compare Top 10 holdings overlap between adjacent quarters
overlaps = []
for prev_q, next_q in zip(quarters, quarters[1:]):
prev_top10 = set(prev_q["top10"])
next_top10 = set(next_q["top10"])
overlap = len(prev_top10 & next_top10) / 10.0
overlaps.append(overlap)
return 1 - sum(overlaps) / len(overlaps) if overlaps else None`score_fund.py` uses a practical proxy: overlap in Top 10 holdings across quarters. It is not exact turnover data, but it is often enough to detect whether a fund is stable, drifting, or moving sharply. In a data environment with gaps and lag, that proxy becomes a signal.
| Approach | What it measures | Why it matters |
|---|---|---|
| Exact turnover | True trading churn | Best when the data is available |
| Top 10 overlap proxy | Stability of visible positions | Useful when turnover is missing or delayed |
| Style score from proxy | How closely a fund matches the inferred manager style | Turns a messy public record into a usable ranking |
The point is not to fake precision. It is to make the uncertainty explicit and still extract something decision-worthy from the record.
`method.md` is the bridge between retrieval and inference
`method.md` is where the repo stops being a document archive and starts acting like a reasoning system. It maps abstract themes to evidence, which lets the agent answer adjacent questions without pretending that every answer comes from a direct quote.
That bridge is delicate. Too loose, and the model starts improvising. Too tight, and it cannot generalize at all. The value here is the middle ground: inference that stays visibly tethered to source material.
Why this matters: it beats generic finance chatbots on trust, not breadth
| Tool | Strength | Weakness |
|---|---|---|
| General LLMs | Broad language coverage | Weak evidence discipline |
| Wind and Eastmoney | Deep financial data access | Not built as a persona-specific agent |
| Generic RAG chatbot | Better sourcing than plain chat | Usually lacks a real scoring model |
| zhengxi-views | Traceable persona reconstruction | Narrow by design |
That is the editorial point. The repo is not trying to out-answer every financial platform. It is trying to be the one that can justify a narrow answer cleanly, with a paper trail.
What this repo says about the next wave of AI tools
`zhengxi-views` is a prototype for a broader category: specialist AI skills built around one expert, one corpus, and one decision rule. The future here is not omniscient chat. It is accountable tooling that knows the limits of its evidence and stays inside them.
That is why the project feels more durable than a novelty chatbot. It is less about mimicking a person and more about encoding a professional standard for how answers should be earned.