/fact-check

description: Claim-level verification of any report or data set — extract every figure and factual claim, check each against a primary source (SEC filing, company IR, FRED, exchange data, or the user's own document), recompute derived numbers, mark ✅ verified / ⚠️ mismatch / ❓ unverifiable, and re-issue the report with inline citations and a References section; Verification Score 0–10

Fact Check — Verify the Numbers, Cite the Sources

⚠️ Data Verification — Do This Before Any Analysis

Before verifying anything, establish what data each claim actually needs:

1. **Fetch current price only where a claim needs it** — market claims (price, market cap, yield, multiples "today") are checked against a live quote for the report's as-of date and today's; use web search or ask the user. Reported figures, historical comparisons, and `--recompute` runs need **no** live quote — do not request one, and do not let its absence block them. Never assume a price from training data. 2. **Confirm key figures** — for this skill the "key figures" are the *sources themselves*: the filing (10-K / 10-Q / 8-K / proxy / Form 4 / 13F), the IR release, the FRED series, the exchange or issuer data, or the document the user pasted. Retrieve or request them before judging a single claim. A claim checked against memory is not checked. 3. **State your data source** — fill in the `Data & Sources` header (next section) so the origin, as-of date, retrieval path, and confidence of every figure are explicit at the top of the output. 4. **Flag stale data explicitly** — if a live quote is needed for a market claim and unavailable, display this warning before proceeding, and mark those claims `❓` or `🕒` rather than guessing:

> ⚠️ **Live data unavailable.** The following analysis uses training-data estimates which may be significantly out of date. Verify all prices and metrics before making any decisions.

The warning describes the *inputs actually used*: when every claim was checked against user-supplied documents, say so instead of implying training-data estimates were involved. Never silently substitute training-data estimates for current prices. When in doubt, ask the user to paste the latest quote.

---

📋 Data & Sources Header — Open Every Output With It

The first thing in the output is this provenance block, filled in — never left as placeholders. It is the standard documented on the [Data & Accuracy](https://yennanliu.github.io/InvestSkill/data-and-accuracy.html) page and the first thing `result-validator` looks for:

Data & Sources
  As of:      <date the figures represent, e.g. 2026-06-30>
  Source:     <primary docs — SEC EDGAR 10-K/10-Q, company IR, FRED, exchange data …>
  Retrieval:  <pasted by user | web/tool retrieval | model memory>
  Confidence: <HIGH | MEDIUM | LOW>

• `Retrieval: model memory` must be paired with `Confidence: LOW` — memory is a placeholder until confirmed against a primary source.
• Mixed sources: list each with its own as-of date rather than blending them.
• Sources disagree on the same figure (a closing price, a deal term): prefer the primary source (SEC filing, company release, exchange data) over an aggregator, then the most recent dated one. Show both values with their sources and dates, say which you used and why, and never average them.
• Data the user pasted is reported as `pasted by user`; do not upgrade its confidence beyond what the user's own source supports.

For this skill the header describes the **sources used to verify**, not the report being verified — the report's own header is quoted and judged in Phase 5.

---

Overview

Every InvestSkill analysis opens with a `Data & Sources` header, and `result-validator` scores whether sources are "cited or identifiable". Neither answers the question a careful reader actually has: **is this specific number true, and where exactly does it come from?** A report can carry an honest header, a tidy methodology, and a consistent signal — and still be built on a revenue figure from the wrong fiscal year, a margin computed on non-GAAP earnings, or a share count that was stale when the model wrote it.

This skill closes that gap at the level of the individual claim. It takes any report — an InvestSkill output, a broker note, a spreadsheet, a paragraph from a forum post — pulls out **every figure and factual assertion**, checks each one against a **primary source**, **recomputes** every derived number from verified inputs, and hands back three things: a claim ledger with a verdict per claim (✅ verified · ⚠️ mismatch · ❓ unverifiable · 🔁 recomputed · 🕒 stale), a **corrected copy of the report with inline citations and a References section**, and a **Verification Score 0–10** that `result-validator` can consume as the evidence behind its Data Quality dimension.

**Division of labour.** `result-validator` asks *is this analysis well built?* (methodology, consistency, risk coverage, transparency). `fact-check` asks *are the inputs true?* Run this one first when the stakes are real; the validator's score means little if the numbers underneath it are wrong. `learning-coach` then explains the result to a beginner.

**What it is not.** It does not re-run the analysis, form its own view on the stock, or change the report's signal. It never invents a source to make a claim look verified: an unverifiable claim stays `❓`, and a claim that cannot be traced is more honestly left uncited than given a plausible-looking reference. Fabricating a citation is the one failure mode that makes this skill worse than useless — it is a hard-cap-to-zero offence in the score.

---

1. Modes

| Mode | Invocation | What it does | |------|------------|--------------| | **Verify** (default) | paste the report (+ sources, or permission to retrieve them) | Full ledger, recomputation, consistency check, corrected report with citations, Verification Score | | **Cite** | `--cite-only` + the report + its sources | Attach inline citations and a Ref