How Earnings-Review Differs from Traditional Equity Research by Using Primary Sources Only

The earnings-review skill enforces a strict "primary-source-only" philosophy that requires every analysis to start with original SEC filings, earnings transcripts, or exchange disclosures, cross-validating all metrics against independent sources before generating a standardized eight-section report.

The ai-berkshire repository implements a specialized equity research workflow through its earnings-review skill that fundamentally departs from conventional analysis methods. Unlike traditional research that blends primary filings with secondary summaries and consensus estimates, this approach treats original corporate disclosures as the sole trusted foundation, applying rigorous verification gates at every step.

The Primary-Source-Only Philosophy

Traditional equity research typically mixes 10-K/10-Q filings with news clips, analyst reports, and third-party aggregators. The earnings-review skill rejects this hybrid approach in favor of direct provenance. As defined in skills/earnings-review.md, the workflow begins with a "前置步骤：资料可得性评级" (pre-step: data availability rating) that categorizes sources into tiers: A (full original filing), B (partial), or C (news only)【/cache/repos/github.com/xbtlin/ai-berkshire/main/skills/earnings-review.md#L24-L31】.

When a ticker receives a C rating, the analysis weight is adjusted accordingly or halted until primary documents are obtained. This enforced hierarchy ensures that management discussion and analysis (MD&A) sections, footnotes on related-party transactions (关联交易), and accounting policy changes originate directly from the issuer rather than interpreted through secondary channels.

Six Key Differences from Traditional Research

Data Origin and Direct Acquisition

Traditional workflows often rely on data vendors that scrape and repackage filings. The earnings-review skill instead implements parallel agents to fetch documents directly from authoritative sources: SEC EDGAR for US listings, HKEX 披露易 for Hong Kong equities, A-share annual reports (年报 PDFs) for mainland China, and raw earnings-call transcripts【/cache/repos/github.com/xbtlin/ai-berkshire/main/skills/earnings-review.md#L24-L31】.

# Fetch the latest 10-Q directly from SEC EDGAR

import subprocess

subprocess.run([
    "task", "fetch-10q", 
    "--ticker", "AAPL"
])

This direct acquisition eliminates interpretation layers that typically distort metrics when passed through intermediary databases.

Source Hierarchy and Availability Rating

The skill enforces a tiered availability rating system that dictates analytical rigor. An A rating permits full analysis using original documents; B ratings trigger partial-weight analysis; C ratings restrict output to data-availability warnings【/cache/repos/github.com/xbtlin/ai-berkshire/main/skills/earnings-review.md#L24-L31】.

Traditional research rarely exposes its source-hierarchy logic, often defaulting to whatever data is "good enough" for modeling. The earnings-review workflow explicitly surfaces this metadata, allowing consumers to understand the evidentiary foundation of every conclusion.

Cross-Validation Requirements

Where traditional research accepts consensus numbers at face value, earnings-review mandates cross-validation against two independent sources as stipulated in skills/financial-data.md【/cache/repos/github.com/xbtlin/ai-berkshire/main/skills/financial-data.md#L1-L4】. The tools/financial_rigor.py utility automates this verification:

# Cross-validate revenue figures between EDGAR and macrotrends

import subprocess

subprocess.run([
    "python3", "tools/financial_rigor.py", 
    "cross-validate",
    "--metric", "revenue",
    "--values", "274.5e9", "275.1e9",
    "--sources", "SEC EDGAR", "macrotrends"
])

Error thresholds are strictly enforced:

• Discrepancies > 1% trigger flags for manual review
• Discrepancies > 5% force an immediate return to the original filing for reconciliation

Analytical Focus: Footnotes Over Headlines

Traditional reports emphasize headline numbers and valuation multiples. The earnings-review skill redirects attention to MD&A tone, management commitments, and hidden footnote items including balance-sheet anomalies (资产负债表异常) and related-party disclosures【/cache/repos/github.com/xbtlin/ai-berkshire/main/skills/earnings-review.md#L94-L118】.

This methodological pivot recognizes that material risks and alpha-generation opportunities often reside in accounting policy changes and nuanced management language rather than top-line revenue beats.

Multi-Step Pipeline vs. Single-Pass Workflow

Conventional equity research typically follows a linear path: read → model → output. The earnings-review skill implements a gated, multi-step pipeline defined in skills/earnings-review.md【/cache/repos/github.com/xbtlin/ai-berkshire/main/skills/earnings-review.md#L32-L176】:

1. Material acquisition with availability rating
2. Core-financial extraction using standardized parsers
3. MD&A tone analysis for sentiment and commitment extraction
4. Footnote deep-dive for accounting quality assessment
5. Historical comparison against 8-K/6-K predecessor filings
6. Structured report generation with verification checks

Each step must pass validation gates before proceeding, preventing error propagation that plagues traditional single-pass analysis.

Standardized Output Structure

Traditional research outputs variable narrative summaries with subjective valuation recommendations. Earnings-review produces a standardized eight-section report that explicitly answers preset questions such as "超预期/符合预期/低于预期?" (beat/meet/miss expectations)【/cache/repos/github.com/xbtlin/ai-berkshire/main/skills/earnings-review.md#L68-L105】.

The tools/report_audit.py utility enforces "准出" (release) criteria before publication, auditing the generated markdown against data-integrity standards:

# Audit the report before publication

import subprocess

subprocess.run([
    "python3", "tools/report_audit.py", 
    "verdict",
    "--results", '{"status":"pass"}',
    "--report", "reports/Apple-earnings-2024Q4.md"
])

This ensures every published analysis includes a data-audit verdict certifying primary-source verification.

Implementation in the ai-berkshire Repository

The primary-source workflow relies on specific components:

File	Function
skills/earnings-review.md	Core skill definition and step-by-step workflow
skills/financial-data.md	Cross-validation standards and permissible secondary sources
tools/financial_rigor.py	CLI utility for metric verification and market-cap calculations
tools/report_audit.py	Pre-publication auditing and release criteria enforcement
codex-prompts/earnings-review.md	Entry-point for the /earnings-review slash command

These components collectively prevent secondary-data contamination by routing all analysis through financial_rigor.py validation checks before any conclusions reach the final report template in reports/.

Summary

• Primary-source exclusivity: Earnings-review starts exclusively with original filings from SEC EDGAR, HKEX, or exchange-specific disclosure platforms, rejecting news summaries and analyst interpretations.
• Validation gates: Every metric undergoes mandatory cross-validation against two independent sources via tools/financial_rigor.py, with >5% discrepancies forcing filing re-review.
• Structured workflow: A six-step gated pipeline replaces the traditional single-pass read, emphasizing MD&A tone and footnote analysis over headline numbers.
• Standardized output: The skill generates an eight-section markdown report with explicit data-audit verdicts, unlike the variable narrative formats of traditional research.

Frequently Asked Questions

What makes earnings-review different from reading analyst reports?

Analyst reports often synthesize primary data through secondary interpretation layers and consensus estimates. Earnings-review bypasses these intermediaries entirely, requiring direct access to original 10-K/10-Q filings or earnings transcripts before any analysis begins. The workflow also mandates cross-validation against independent financial databases like macrotrends, stockanalysis, 东方财富, and FinMind to verify accuracy.

How does the availability rating system work in earnings-review?

The system assigns an availability rating during the "前置步骤：资料可得性评级" (pre-step data availability assessment) defined in skills/earnings-review.md【/cache/repos/github.com/xbtlin/ai-berkshire/main/skills/earnings-review.md#L24-L31】. A-rated tickers have accessible original filings permitting full analysis; B-rated tickers have partial primary sources requiring adjusted weighting; C-rated tickers rely only on news and trigger restricted output or data-availability warnings.

What happens when data discrepancies exceed 5%?

When tools/financial_rigor.py detects discrepancies greater than 5% between primary and secondary sources, the workflow forces an immediate return to the original filing for reconciliation【/cache/repos/github.com/xbtlin/ai-berkshire/main/skills/financial-data.md#L1-L4】. This prevents publication of potentially corrupted metrics and ensures the final report reflects source-verified data rather than consensus errors.

Which primary sources does earnings-review prioritize?

The skill prioritizes original filings in descending order: SEC EDGAR 10-K/10-Q documents for US equities, HKEX 披露易 PDFs for Hong Kong listings, A-share annual report PDFs (年报) for mainland China exchanges, and verbatim earnings-call transcripts. Secondary sources such as financial news or analyst summaries are explicitly deprioritized unless no primary alternative exists, in which case the C-rating restriction applies.

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