# How the Private-Company-Research Skill Handles Data Scarcity: A Technical Deep Dive

> Discover how the private company research skill tackles data scarcity using hierarchical source ranking, cross-verification, and a fallback mode. Learn its technical approach.

- Repository: [Xbt Lin/ai-berkshire](https://github.com/xbtlin/ai-berkshire)
- Tags: deep-dive
- Published: 2026-07-03

---

**The private-company-research skill mitigates data scarcity through a three-tier defense system: hierarchical source ranking with confidence tags (🟢🟡🔴), mandatory cross-verification of all financial metrics, and a first-principles fallback mode that marks unknowns as "data missing" rather than fabricating estimates.**

The private-company-research skill in the `xbtlin/ai-berkshire` repository is engineered specifically for the opacity of unlisted companies. It employs a "team-lead" agent architecture that treats data scarcity as a first-class constraint rather than an edge case, ensuring every analysis acknowledges uncertainty explicitly.

## Architecture: Three Mechanisms for Scarce Data

The skill's workflow, defined in [`skills/private-company-research.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/private-company-research.md), implements a resilient pipeline that recognizes scarcity, quantifies uncertainty, and prevents the illusion of precision.

### Multi-Source Hierarchy and Confidence Labelling

At the foundation lies a **prioritized matrix of data sources** (lines 85‑97). The skill defines a strict hierarchy ranging from high-fidelity documents (prospectuses, parent-company filings, regulatory disclosures) down to lower-trust signals (industry estimates, former-employee tips). Every source carries an explicit confidence tag:

- **🟢 High**: Audited filings and direct disclosures
- **🟡 Medium**: Third-party reports and financing news
- **🔴 Low**: Deep-media articles and unverified estimates

These tags propagate through the entire workflow, ensuring that downstream agents weight data according to its proven reliability.

### Cross-Verification with Transparent Attribution

Every key metric—revenue, cost, cash flow—is **cross-validated against at least two independent sources** whenever possible (lines 31‑34). The skill mandates that each data point records its *origin, timestamp, and confidence level*. When sources conflict, the workflow lists figures side-by-side with a concise rationale for the chosen value, preventing silent overriding of weak signals by strong ones.

### First-Principles Mode for Extreme Scarcity

When public information is so thin that the hierarchy yields few facts, the workflow **switches to a "first-principles" mode** (lines 29‑33). Instead of forcing a full-blown report, the skill focuses on four core questions: business problem, team advantage, upside ceiling, and validation milestones. This mode enforces strict **"leave-blank" discipline**—any field unsupported by evidence is marked as *data missing* and excluded from valuation.

The skill's design philosophy is codified as a bias-mitigation principle (lines 26‑28): *"宁可留白说‘不知道’，也不要用推测填满表格伪装确定性"* ("Better to leave blank and say 'don't know' than to fill the form with speculation pretending to be certainty").

## Implementation in Code

The scarcity-handling logic is exposed through a Python API that wraps the six specialized agents (team-lead, financial-detective, market-analyst, etc.). When data is unavailable, the system returns structured uncertainty markers rather than hallucinated values.

```python

# Example: invoking the private-company-research skill for a startup with limited data

from ai_berkshire.skills import private_company_research

# Step 1 – launch the team

team = private_company_research.create_team(company_name="AcmeAI", lead="team-lead")

# Step 2 – ask the financial‑detective for FY2024 revenue

revenue_info = team.query(
    role="financial-detective",
    subject="拼凑AcmeAI财务数据并进行估值分析",
    question="AcmeAI 2024 年度收入是多少？"
)

print(revenue_info)

```

When data is scarce, the `financial-detective` agent returns a structured response with explicit confidence tagging:

```json
{
  "metric": "Revenue FY2024",
  "value": "🔴 低 – 数据缺失",
  "sources": [],
  "notes": "未找到公开的收入披露；仅有媒体报道提及融资规模，无法推算收入。"
}

```

This pattern is applied across all six agents, ensuring every piece of information carries an explicit confidence indicator.

## Configuration and Source Files

The scarcity-handling rules are encoded across the following repository files:

- **[`skills/private-company-research.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/private-company-research.md)** — Master definition of the workflow, including the Data Sources Matrix and leave-blank discipline rules.
- **[`codex-prompts/private-company-research.md`](https://github.com/xbtlin/ai-berkshire/blob/main/codex-prompts/private-company-research.md)** — Generated slash-prompt that exposes the skill to Codex-style commands, kept in sync with the scarcity logic.
- **[`AGENTS.md`](https://github.com/xbtlin/ai-berkshire/blob/main/AGENTS.md)** — Repository-wide description of the "team-lead" orchestration model used to coordinate agents during uncertain analysis.
- **[`scripts/sync-codex-skills.py`](https://github.com/xbtlin/ai-berkshire/blob/main/scripts/sync-codex-skills.py)** — Utility that regenerates the Codex version after any change to the Markdown source, ensuring scarcity-handling logic stays current.

## Summary

- **The private-company-research skill** treats data scarcity as a core constraint, not an exception.
- **Confidence tags (🟢🟡🔴)** are mandatory for every data point, sourced from a hierarchical matrix defined in [`skills/private-company-research.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/private-company-research.md).
- **Cross-verification** requires two independent sources for key metrics, with conflicting values preserved and explained.
- **First-principles mode** triggers automatically for extreme scarcity, enforcing "leave-blank" discipline rather than fabrication.
- **Implementation** uses a team of specialized agents (team-lead, financial-detective) that return structured "data missing" flags when evidence is absent.

## Frequently Asked Questions

### What does the "leave-blank" discipline mean in the private-company-research skill?

The "leave-blank" discipline is a strict rule that prohibits agents from filling valuation templates with speculative numbers. When a metric like revenue or cash flow cannot be verified through the skill's multi-source hierarchy, the agent must explicitly mark the field as *data missing* (数据缺失) and exclude it from final calculations. This prevents the "illusion of precision" that plagues private-company analysis.

### How does the skill determine which confidence tag to assign?

Confidence tags are determined by the **Data Sources Matrix** defined in [`skills/private-company-research.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/private-company-research.md) (lines 85‑97). Audited prospectuses and regulatory filings receive 🟢 high confidence; third-party reports and financing news receive 🟡 medium confidence; industry estimates and media articles receive 🔴 low confidence. The `financial-detective` agent applies these tags automatically during the cross-verification phase.

### Can the private-company-research skill generate reports when no financial data exists?

Yes, through **first-principles mode**. When the scarcity threshold is reached (no reliable sources available), the skill pivots to qualitative analysis focused on four core questions: business problem, team advantage, upside ceiling, and validation milestones. This mode produces actionable investment theses without fabricating quantitative metrics, ensuring intellectual honesty when analyzing early-stage or stealth startups.

### Which agents are involved in handling data scarcity?

The skill utilizes six specialized agents coordinated by the **team-lead** agent. The **financial-detective** agent handles quantitative cross-verification and applies confidence tags, while the **market-analyst** and **industry-expert** agents provide qualitative context when numerical data is absent. All six agents adhere to the same confidence-labelling and leave-blank protocols defined in the master skill file.