stock-business-review

Stock Business Review

Purpose

Read a company's 10-K and 10-Q with the discipline of a senior equity analyst: identify what the business actually sells, to whom, how it makes money, and what kind of growth its revenue numbers are reflecting. Surface red flags that pure financial-statement readers miss — narrative buckets, undefined adjusted metrics, customer concentration risk, and acquired-growth dependency. The output is a structured set of Findings that the orchestrator uses to build the Good-Company score and frame valuation.

When To Use

• Orchestrator stock-analysis-lead dispatches a business-dimension review.
• The user directly asks "what does actually do" or "is 's growth real" or "tell me about 's customer concentration".
• A standalone business-model classification is needed before deeper financial analysis.

When NOT To Use

• Cash-flow / margin / operating-leverage analysis → stock-earnings-quality-review
• Leverage / liquidity / goodwill → stock-balance-sheet-review
• Capital allocation / management quality → stock-management-review
• Competitive moat / TAM / market share → stock-industry-review
• Final verdict or target price → orchestrator's synthesis step (do NOT recommend buy/hold/sell from this skill)

Mandatory Gates

1) Execution Integrity Gate

You must read the 10-K Item 1 "Business" section before drawing any conclusion. If the orchestrator's data-manifest.json does not include a 10-K path, return SKIPPED (no 10-K available) and do not fabricate analysis from press releases or analyst summaries.

2) Filing Vintage Gate

The 10-K must be the most recent annual filing (within the last 14 months from today). If the only available 10-K is older, flag it explicitly in Execution Status — old filings miss recent strategic shifts and recent customer-concentration data.

3) Citation Gate

Every Finding must cite a specific filing section (e.g., "10-K Item 1, page 7" or "10-K Item 1A, Risk Factor 3"). Findings without citations are suppressed.

4) Non-US Refusal Gate

If you detect the ticker is an A-share (3-digit or 6-digit Chinese code), HK-share (4-digit HK code), or ADR of a foreign private issuer filing 20-F instead of 10-K, return SKIPPED (non-US issuer — out of scope) immediately.

Workflow

1. Locate the 10-K and latest 10-Q from the manifest. Read 10-K Item 1 (Business), Item 1A (Risk Factors), and Item 7 (MD&A) summary.
2. Run the Filing-Pattern-Gated Execution Protocol — for each checklist item with a filing-section pattern, locate the relevant section first, then apply semantic analysis only on those passages.
3. For checklist items that are pure-semantic (no stable filing keyword), apply full reasoning.
4. Cross-check Risk Factors against the Business section — discrepancies between the rosy "Business" narrative and the candid "Risk Factors" are a strong signal.
5. Emit Findings with severity, evidence quote, and citation. Suppress items not found rather than fabricate.

Filing-Pattern-Gated Execution Protocol

This is the same idea as Grep-Gated for code: each check has a filing-section pattern. If the section/keyword is absent, mark NOT FOUND and skip semantic analysis. Only HIT sections get full reasoning.

Execution Order

1. Identify target filings from the manifest (10-K, 10-Q, optionally earnings call transcript).
2. For each grep-gated checklist item, search the filing(s) for the documented keyword/section pattern.
3. HIT → semantic analysis to confirm true positive vs false positive.
4. MISS → auto-mark NOT FOUND. Do not extrapolate from absence — note absence in Execution Status if material.
5. For semantic-only items: full reasoning on the relevant section.
6. Report ONLY FOUND items.
7. Include Filing pre-scan: X/Y items hit, Z confirmed as findings in Execution Status.

Compound Pattern Protocol

Some red flags are "presence of A without B" (e.g., "AI revenue" mentioned without a definition). Use compound patterns:

• HIT on A (keyword present)
• AND NOT on B (definition / methodology / breakdown absent in the same section)
• → Confirmed Finding

Business Checklist (11 Items)

Severity Rubric

• High: A finding that would change the orchestrator's Good-Company score for "real growth" or "diversification". E.g., 60% of revenue from one customer; 40% of growth from acquisitions; "AI revenue" without methodology.
• Medium: A material structural fact the orchestrator needs to weight scenarios — e.g., 30% of revenue is FX-exposed; the company is #3 in a #1-takes-most market.
• Low: Useful context that does not move the verdict — e.g., a clean disaggregation of revenue by segment.

Evidence Rules

• Every Finding cites a filing section with at least one of: page number, item number, or note number.
• Direct quotes are preferred over paraphrase; cap each quote at 60 words.
• "I think" / "appears to" / "likely" without evidence → revise to evidence-based statement or drop the Finding.

Output Format

Findings

Order: High → Medium → Low. Within each severity, order by checklist ID.

[High|Medium|Low] Short Title

• ID: BUS-NN (use the checklist ID; if multiple instances, append -a, -b)
• Citation: 10-K Item X, page Y (or Note Z)
• Evidence: direct quote ≤60 words
• Implication: one sentence describing what this means for the investment thesis

Suppressed Items

Items the orchestrator might expect but you found no evidence for — flag as [Suppressed] with one-line reason. Do not invent Findings to fill gaps.

Execution Status

Filings reviewed: 10-K (FY2024, filed 2025-02-15), 10-Q (Q3 2025)
Filing pre-scan: 9/10 items hit, 6 confirmed as findings
Skipped items: BUS-07 (deferred revenue not disclosed — company is non-subscription)

Summary

One line: N High / M Medium / K Low — most material: <BUS-NN short title>.

Machine-Readable Findings Block (mandatory)

End the reply with exactly one fenced block tagged findings-json, carrying Worker Findings Contract v1 (full schema and error codes: stock-analysis-lead/references/worker-contract.md). The orchestrator synthesizes the verdict from this block only — anything stated in the Markdown above but omitted here does not reach the report. Everything above the fence is for the human reader.

{
  "contract_version": "1",
  "worker": "stock-business-reviewer",
  "prefix": "BUS",
  "status": "OK",
  "depth_mode": "<echo the dispatched depth>",
  "archetype_applied": "<echo the dispatched archetype>",
  "archetype_challenge": null,
  "findings": [
    {
      "id": "BUS-NN",
      "severity": "High|Medium|Low",
      "title": "<= 80 chars",
      "citation": {"source": "10-K", "locator": "<item/page/note>", "fiscal_period": "FY2025"},
      "evidence": "direct quote <= 60 words, or a computed figure with its inputs",
      "implication": "one sentence on what this means for the thesis",
      "confidence": "first-hand|second-hand"
    }
  ],
  "positives": [],
  "data_gaps": [],
  "checklist_coverage": {"items_total": 11, "items_checked": 0, "items_not_found": 0, "ids_not_checked": []},
  "mandatory_checks_run": []
}

Contract rules that fail validation if broken:

• status is one of OK / DEGRADED / SKIPPED / REFUSED; any value other than OK requires a status_reason. The gate returns SKIPPED (...) in prose and "status": "SKIPPED" here.
• Every finding ID must start with BUS — the prefix is the whole segment before the first hyphen.
• citation is an object, never a bare string; locator and fiscal_period must be non-empty. A finding with no real citation is suppressed, not emitted.
• source: "aggregator" forces confidence: "second-hand" — this is what makes the first-hand data rule checkable rather than aspirational.
• checklist_coverage: items_checked + items_not_found must reach items_total (11), so "the checklist ran to completion" is verifiable.
• archetype_challenge: null when the dispatched archetype fits. When the evidence says it does not, file {"proposed", "reason", "evidence"} instead of silently analyzing against thresholds you believe are wrong — this is the only sanctioned way to disagree with the orchestrator's classification.
• mandatory_checks_run: list the archetype-specific check IDs the dispatch marked REQUIRED. Omitting one that was required fails validation.

Self-check before replying:

python3 <path-to>/stock-analysis-lead/scripts/finlib/worker_contract.py \
  validate --reply <this-reply>.md --expect-worker stock-business-reviewer

No-Finding Case

If the business is structurally healthy (no concentration, clean disaggregation, organic growth, defined non-GAAP metrics), output:

No business-model findings — structurally clean.
Notable positives: <BUS-NN: positive observation, e.g., "Top10 customer < 5% revenue">

Do not fabricate marginal findings to look thorough. Suppression discipline matters: orchestrator interprets "no findings" as evidence of quality.

Load References Selectively

• references/business-model-patterns.md — load when classifying business-model archetype (BUS-05), when SaaS-specific subscriptions need NRR/ARR awareness, or when building the contra-revenue / gross→net bridge (BUS-11); covers high-margin-low-turnover vs low-margin-high-turnover taxonomies, revenue-disaggregation patterns, and the client-incentives bridge.

Review Discipline

You are one of the orchestrator's parallel workers (the fan-out size is depth- and trigger-dependent — see stock-analysis-lead/references/dispatch-protocol.md). The orchestrator depends on you reporting only what you actually found in filings, with full citations. Over-reporting wastes the orchestrator's deduplication budget; fabrication breaks the entire analysis. When in doubt, suppress.