# How AI Berkshire Prevents Bias in Research: A Systematic Framework for AI-Driven Investment Analysis > Discover how AI Berkshire prevents bias in research. Learn about its systematic framework, bias-awareness layer, and mandatory checklist for unbiased investment analysis. Explore the xbtlin/ai-berkshire repository for details. - Repository: [Xbt Lin/ai-berkshire](https://github.com/xbtlin/ai-berkshire) - Tags: deep-dive - Published: 2026-07-27 --- **AI Berkshire prevents bias in research by embedding a systematic bias-awareness layer into every workflow, combining an Information-Richness Rating (A/B/C) system with a mandatory four-point checklist that analysts must complete before finalizing any investment memo.** The `xbtlin/ai-berkshire` repository implements a rigorous governance framework to ensure AI-generated research maintains objectivity. By classifying target companies based on data availability and consensus strength, the system forces analysts to confront specific cognitive biases through structured documentation. This approach transforms abstract bias concerns into concrete, auditable actions embedded directly in the research lifecycle. ## Information-Richness Rating: The A/B/C Classification System Before analysis begins, AI Berkshire categorizes every target company using an **Information-Richness Rating** defined in [`README_EN.md`](https://github.com/xbtlin/ai-berkshire/blob/main/README_EN.md). This classification determines the intensity of subsequent bias checks based on how much public information exists and how strong the market consensus is. The three-tier system works as follows: - **A-class**: Information-rich environments with abundant analyst coverage. These targets carry high **consensus risk** because prevailing narratives are deeply entrenched. - **B-class**: Moderate coverage with mixed opinions, requiring standard bias awareness. - **C-class**: Information-scarce targets, often private or early-stage firms, where data limitations rather than consensus pose the primary challenge. This rating triggers downstream workflow modifications. For A-class targets, the system automatically injects enhanced scrutiny via the bias-awareness checklist, ensuring that high-consensus environments receive appropriate skepticism. ## The Four-Category Bias-Awareness Checklist For information-rich (A-class) or consensus-heavy targets, AI Berkshire mandates a dedicated **"AI 研究偏见自觉"** (AI research bias awareness) section. Defined in [`codex-skills/investment-memo-craft/SKILL.md`](https://github.com/xbtlin/ai-berkshire/blob/main/codex-skills/investment-memo-craft/SKILL.md), this checklist captures four specific bias categories and requires documented mitigation actions for each. **Large-cap bias** occurs when analysts over-weight well-known corporate giants. The mitigation step requires explicitly comparing the target against smaller peers to ensure mid- and small-cap alternatives receive fair consideration. **English-language bias** manifests as a preference for English-language sources, potentially missing critical local intelligence. Analysts must include Chinese-language reports and local regulatory filings to counter this limitation. **Narrative bias** involves following prevailing storylines without sufficient skepticism. The system requires analysts to note contrary data points and counter-arguments that contradict the dominant market thesis. **Listed-only bias** describes the tendency to ignore private-market opportunities. Mitigation requires adding a "future IPO candidates" list, referencing the industry-funnel skill to capture pre-public companies in the sector. ## Workflow Integration: From Skill Invocation to Audit Trail The bias prevention mechanism activates automatically when specific skills are invoked. When a user calls `/investment-research`, the generated report begins with the **AI研究偏见自觉** section (see template lines 97-100 in the skill definition). The workflow proceeds through several enforced stages: 1. **Classification**: The system records the Information-Richness Rating (A/B/C) for the target. 2. **Checklist Population**: Analysts must fill the four-category checklist with concrete items, such as listing the top-five English-language analyst reports or documenting private-company pipeline candidates. 3. **Cross-Reference Requirements**: During subsequent analysis stages—including moat assessment, management evaluation, and valuation—the analyst must reference checklist items to confirm or disprove earlier bias hypotheses. The [`skills/investment-research.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/investment-research.md) file serves as the source markdown for this process, while [`skills/industry-funnel.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/industry-funnel.md) repeats the bias-awareness step in sector-wide analyses. For example, the industry-funnel skill explicitly calls out the same four bias categories in its "AI bias awareness" bullet point, ensuring consistency across research types. Real-world outputs demonstrate this enforcement. In `reports/AI算力-funnel-20260509.md`, completed research deliverables show fully documented bias awareness sections, creating an audit trail for compliance review. ## Practical Implementation: Triggering the Bias-Aware Workflow Researchers interact with the bias prevention system through simple skill invocations. The following examples demonstrate how the workflow generates mandatory bias documentation. To initiate a bias-aware analysis for a specific company: ```markdown # Invoke the bias-aware research skill /prompts:investment-research NVIDIA # The generated report begins with the bias section (excerpt) ## AI研究偏见自觉 - 信息丰富度评级：A - 共识陷阱：高 – 多数分析师持“长期增长”乐观观点 - 偏见检查清单 - 大盘偏见 ✔ 已对比同业中小盘公司 - 英文语言偏见 ✔ 包含中文晨星报告与本地监管文件 - 叙事偏见 ✔ 记录与“AI 计算芯片”热潮相悖的产能利用率数据 - 上市仅偏见 ✔ 添加未上市的芯片材料供应商作为潜在 IPO 目标 ``` The same enforcement appears in industry-wide research: ```markdown /prompts:industry-funnel AI Compute ## AI bias awareness - Counters large-cap bias, English-language bias, narrative bias, listed-only bias ``` These snippets illustrate that bias-awareness elements are automatically inserted by the skill definitions; analysts cannot bypass the checklist and must populate the specific mitigation fields before completing the deliverable. ## Summary AI Berkshire implements a multi-layered defense against research bias through systematic classification and mandatory documentation: - **Information-Richness Ratings (A/B/C)** calibrate bias checks based on data availability and consensus strength, with A-class targets receiving maximum scrutiny. - **Four specific bias categories** (large-cap, English-language, narrative, and listed-only) are defined in [`codex-skills/investment-memo-craft/SKILL.md`](https://github.com/xbtlin/ai-berkshire/blob/main/codex-skills/investment-memo-craft/SKILL.md) with explicit mitigation steps. - **Automated workflow integration** ensures the "AI 研究偏见自觉" checklist appears in every relevant research output, from [`skills/investment-research.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/investment-research.md) to [`skills/industry-funnel.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/industry-funnel.md). - **Audit trail requirements** force analysts to reference bias checks during subsequent valuation and analysis stages, preventing checkbox-compliance without substantive review. ## Frequently Asked Questions ### What is the Information-Richness Rating system in AI Berkshire? The Information-Richness Rating is a three-tier classification (A, B, C) defined in [`README_EN.md`](https://github.com/xbtlin/ai-berkshire/blob/main/README_EN.md) that categorizes target companies based on public data availability and analyst consensus strength. A-class targets are information-rich with high consensus risk, B-class have moderate coverage, and C-class are information-scarce, often private firms. This rating determines the intensity of bias checks applied during the research workflow. ### What are the four types of bias that AI Berkshire checks for? According to the skill definitions in [`codex-skills/investment-memo-craft/SKILL.md`](https://github.com/xbtlin/ai-berkshire/blob/main/codex-skills/investment-memo-craft/SKILL.md), the system checks for **large-cap bias** (over-weighting giants), **English-language bias** (ignoring non-English sources), **narrative bias** (following prevailing storylines), and **listed-only bias** (ignoring private markets). Each category requires specific mitigation actions, such as comparing against smaller peers or including Chinese-language reports. ### How is the bias checklist enforced during the research process? The checklist is enforced through automated skill templates. When invoking `/investment-research` or `/industry-funnel`, the system automatically inserts the **AI研究偏见自觉** section (template lines 97-100) at the beginning of the report. Analysts must document mitigation actions for each bias category, and subsequent analysis sections must reference these checks, creating a mandatory audit trail that cannot be bypassed. ### Where is the bias prevention logic defined in the codebase? The core logic resides in three key locations: [`README_EN.md`](https://github.com/xbtlin/ai-berkshire/blob/main/README_EN.md) describes the overall mechanism and four bias types; [`codex-skills/investment-memo-craft/SKILL.md`](https://github.com/xbtlin/ai-berkshire/blob/main/codex-skills/investment-memo-craft/SKILL.md) defines the checklist structure and automation rules; and [`skills/investment-research.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/investment-research.md) implements the workflow integration. Example outputs in `reports/AI算力-funnel-20260509.md` demonstrate the live application of these definitions.