How the /private-company-research Skill Handles Data Confidence Levels (🟢🟡🔴) for Unlisted Companies

The /private-company-research skill enforces a strict three-tier confidence-badge system that tags every metric with 🟢 (High), 🟡 (Medium), or 🔴 (Low) reliability grades, ensuring transparent sourcing and cross-validation during private company analysis.

The ai-berkshire repository provides specialized research tools for analyzing non-public entities where regulatory disclosures are limited. The /private-company-research skill addresses this opacity by treating data confidence levels as mandatory metadata, embedding a color-coded verification framework directly into the research workflow to prevent unverified assumptions from contaminating investment analyses.

The Three-Tier Confidence Classification System

The skill defines a strict hierarchy of source reliability that every researcher must apply. Each confidence level corresponds to specific source types and verification standards.

High Confidence (🟢): Primary Sources

Green-badge data represents information from primary, verifiable disclosures. According to the source matrix defined in skills/private-company-research.md (lines 86-96), this includes regulatory filings, official prospectuses, and audited financial statements. These sources undergo minimal interpretation and carry the highest weight during conflict resolution.

Medium Confidence (🟡): Secondary Sources

Yellow-badge data originates from reputable secondary sources such as established analyst reports, trusted media publications, and verified industry databases. While credible, these sources involve some degree of interpretation or compilation, warranting a middle-tier classification that signals "use with contextual awareness."

Low Confidence (🔴): Inferred or Estimated Data

Red-badge data encompasses metrics that are inferred, estimated, modeled, or based on unverified rumors. The skill explicitly flags these items as speculative, requiring clear documentation of the estimation methodology and triggering "信息缺失" (information gap) alerts in final reports.

Workflow Integration: Three Enforcement Stages

The confidence system is not merely advisory—it is enforced through mandatory workflow checkpoints documented in the skill specification.

Initial Data-Source Matrix

Before analysis begins, researchers must define a prioritized source matrix that maps each potential input type to its corresponding confidence level. This matrix is explicitly documented in skills/private-company-research.md (lines 86-96), establishing ground rules that prevent subjective grading during later stages.

Per-Metric Annotation Requirements

Every computed metric must include a structured annotation recording four elements: source, time, confidence, and estimation method. As specified in lines 99-104 of the skill file, revenue figures, user counts, and valuation estimates must each carry these metadata tags to maintain auditability.

Global Enforcement Rules

The skill mandates that "每个关键数据必须标明：来源、时间、置信度、推算方法" (every key data point must indicate source, time, confidence, and calculation method). This rule appears in the "数据标注规范（严格执行）" section (lines 25-31) and is reiterated in the badge schema definition (lines 26-28), ensuring researchers cannot bypass confidence tagging.

Cross-Validation and Conflict Resolution

When multiple sources report conflicting values for the same metric, the system employs confidence-weighted arbitration. According to lines 31-34 of skills/private-company-research.md, the workflow requires listing every conflicting source, highlighting discrepancies, and giving precedence to higher-confidence entries following the hierarchy 🟢 > 🟡 > 🔴.

For example, when two high-confidence sources contradict each other, the skill recommends presenting both values alongside a judgment call:

| 指标 | 来源 A | 时间 A | 置信度 A | 来源 B | 时间 B | 置信度 B | 采信判断 |
|------|--------|--------|----------|--------|--------|----------|----------|
| 2025 Q4 收入 | $3.1 B | 2025-12-01 | 🟢 | $3.3 B | 2025-12-03 | 🟢 | 采用🟢均值 $3.2 B |

Transparency in Final Reporting

The confidence system directly shapes the final deliverable. The consolidated report includes a dedicated "置信度" (confidence) column for every key figure, allowing readers to instantly assess reliability. Additionally, low-confidence items (🔴) automatically populate the "信息盲区" (information blind spots) section, ensuring analysts explicitly acknowledge uncertainty rather than hiding it behind point estimates.

Implementation Example

Researchers implement these standards using structured markdown tables that conform to the badge schema. Below is a minimal example of metric documentation:

| 指标 | 数据 | 来源 | 时间 | 置信度 |
|------|------|------|------|--------|
| 2025 Q4 收入 | $3.2 B | 2025 Q4 财报（招股书） | 2025-12 | 🟢 |
| 2025 Q4 用户数 | 45 M | 2025 Q4 内部招聘报告 | 2025-11 | 🟡 |
| 2025 Q4 EBITDA | $0.5 B (推算) | 运营访谈 + 行业基准 | 2025-12 | 🔴 |

This format ensures that every data point carries its provenance and reliability grade, fulfilling the skill's requirement for transparent source attribution.

Summary

• The /private-company-research skill employs a mandatory 🟢🟡🔴 badge system to classify source reliability for unlisted company research.
• High-confidence (🟢) data derives from primary disclosures, medium-confidence (🟡) from reputable secondary sources, and low-confidence (🔴) from estimates or inference.
• Three enforcement stages—initial source matrix, per-metric annotation, and global rules—ensure compliance with confidence-tagging requirements.
• Cross-validation logic prioritizes higher-confidence sources when conflicts arise, with explicit documentation of resolution methods.
• Final reports surface confidence levels in dedicated columns and highlight low-confidence data in "information blind spots" sections.

Frequently Asked Questions

What happens when two sources have the same confidence level but different values?

When two sources carry identical confidence badges (e.g., two 🟢 sources) but report conflicting figures, the skill requires documenting both values in the table alongside the resolution logic. Researchers typically calculate a mean or select the more recent timestamp, but must explicitly state this judgment in the "采信判断" (adoption decision) column to maintain transparency.

Can low-confidence (🔴) data ever be used in investment decisions?

Yes, but with strict documentation requirements. The skill explicitly mandates that 🔴 data must include the estimation method and trigger an "information blind spot" flag. This ensures analysts acknowledge uncertainty while allowing the inclusion of modeled figures—such as revenue projections based on industry benchmarks—when primary sources are unavailable for private companies.

Where is the confidence-badge schema defined in the repository?

The master definition resides in skills/private-company-research.md, specifically in the "数据标注规范" (data annotation standards) section (lines 25-31) and the source matrix (lines 86-96). Lines 26-28 explicitly define the three-emoji schema, while lines 99-104 mandate the four-element annotation structure (source, time, confidence, method).

How does the system prevent researchers from misclassifying source quality?

The workflow enforces classification through mandatory table structures and validation rules. The "Initial Data-Source Matrix" (lines 86-96) pre-maps source types to confidence levels before research begins, eliminating subjective grading during analysis. Additionally, the global enforcement rules (lines 25-31) require peer-review-style documentation that makes misclassification auditable in the final markdown output.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}