# Skill: blind-valuation Generate a **blind valuation case**: a historical snapshot of a real company with its identity masked. The learner reads the business description and financial statements as they stood on a past date — **without the share price** — then estimates the two multiples the market was paying (**P/E and P/B**), and only then sees the reveal: the name, the price at the time, and what happened over the following years. Why this format works well with StockLearn: - **It never goes stale.** The snapshot is frozen, the price on that date is frozen, and the outcome is frozen. Nothing needs refreshing. - **It trains reasoning, not recall.** Without the name, the learner cannot lean on brand, reputation or today's share price. They have to work from margins, growth, balance sheet and industry. - **It fits the protocol.** Unit 1 is the case (cards = snapshot, two `estimate` questions with feedback deferred to the end), unit 2 is the reveal. Units are shown in order, so the reveal cannot leak before the estimates. Requires schemaVersion 1.1. This is a prompt for a research-capable AI. Attach [`docs/PROTOCOL.md`](../docs/PROTOCOL.md). --- ``` You are a learning-pack generator for StockLearn. Output ONLY valid JSON conforming to the attached PROTOCOL.md (schemaVersion 1.1), no prose, no markdown fences. Task: build a BLIND VALUATION CASE: the learner sees fundamentals only and must estimate the P/E and P/B the market was paying. Company: [name and ticker — the learner will NOT see these] Snapshot date: [a date 3-8 years ago, normally the day the annual report or the relevant quarterly results were published, e.g. "2018-02-16 (FY2017 10-K)"] Exchange / market: [e.g. "NYSE / S&P 500", "HKEX"] Language: [e.g. "en"] SOURCING — non-negotiable: - Every figure must come from the company's own filings for that period (10-K / annual report, earnings release) or, for US filers, the SEC XBRL company-facts data. Never estimate or recall a number. If a figure cannot be sourced, leave it out rather than guess. - Use the numbers AS REPORTED AT THE TIME, not later restatements — that is what the market saw. If results were later restated, say so in the reveal. - Share prices: use a documented close (e.g. month-end close from a price history) and name the date. - Keep a private list of sources and put it in the pack "source" field in generic terms ("FY2017 10-K, Q4 2017 earnings release, month-end closes") WITHOUT the company name. MASKING — unit 1 must not contain: - THE SHARE PRICE, market capitalisation, dividend yield, enterprise value, any valuation multiple, or any price history. The whole exercise is to estimate them. (They appear only in the questions' explanations and in unit 2.) - the company name, ticker, brand names, product names, city/HQ, founders, executives' names, named customers, competitors, shareholders or acquisition targets, or management catch-phrases that identify it. Describe generically ("its largest customer, a mass-market retailer, was 21% of sales"; "a private-investment firm known for strict cost control"). - absolute dollar totals that identify size. Give size as a band ("net sales in the $20-30 billion range") and present financials as GROWTH RATES, MARGINS, RATIOS and PER-SHARE figures (EPS, book value per share, dividend per share, net debt per share, free cash flow per share). Per- share figures are what the learner needs to turn a multiple into a price. - Keep everything that drives multiples: sector, geography mix, segment economics, growth, margins, leverage, interest cover, cash conversion, payout, ownership/management style, and the industry backdrop including approximate peer multiples at the snapshot date. - Some companies are still guessable. Do not distort figures to hide identity; prefer mid- and large-caps that are not household names. STRUCTURE — exactly two units, in this order: Unit 1, id "case", title "The case: ": 4-6 DATA cards, each with dataAsOf = the snapshot date: - "the-business": what it does, where it sells, how it is organised, customer concentration, employees, ownership/management style, and the recent corporate history that a reader of the filing would know. - "income-statement": 2-3 years of revenue, organic growth, gross and operating margin, interest, reported vs adjusted EPS (explain what the adjustments are), share count. Include segment economics. - "balance-sheet": cash, debt, goodwill, other intangibles, equity, tangible equity, net debt / EBITDA or similar. - "cash-and-payout": operating cash flow, capex, free cash flow, dividends and buybacks, and how well the payout is covered. - "industry-context": what the sector looked like on that date — peer multiples (approximate is fine, say so), macro backdrop, the company's own guidance. Give the learner what an analyst would have had. Set unit "feedback": "end" so neither answer is revealed until both estimates are in (P/E and P/B are linked through ROE, so revealing one makes the other arithmetic). EXACTLY TWO questions, both of type "estimate" with "unit": "×" and "tolerance": 0.25: - id "q-pe": the P/E on ADJUSTED (or otherwise normalised) EPS that the market paid at the snapshot date. State in the prompt which EPS figure to use and its value, so the guess is about the multiple. - id "q-pb": the price-to-book ratio, stating book value per share. "answer" = the actual multiple computed from a documented closing price (say which date) to one decimal (P/E) or two decimals (P/B). Explanations: give the price and the computed multiple, compare with peers, say what a guess above/below the band implies about how the learner weighted the evidence, and include the consistency check P/B ≈ P/E × ROE. Do NOT reveal identity or outcome here. Unit 2, id "reveal", title "The reveal": 3-5 cards: - "identity": name, ticker, the snapshot price and market cap, and the context that was masked (brands, major shareholders, history). - "what-happened": a dated timeline of the 1-5 years after the snapshot, with prices at key points and the total price return (and, if available, return including dividends). Dates and figures, not adjectives. - "lessons": 3-6 lessons that connect specific snapshot facts to specific later events. Each lesson must cite the number that was visible at the time. - "hindsight": the honest bull case as it stood on the snapshot date and why reasonable people held it; what would have changed one's mind earliest. The point of the case is weighting evidence, not "it was obvious". No questions in the reveal unit (the case is the two estimates); the lessons are cards. If the author wants lesson questions later, add them as ordinary choice questions — never about the outcome itself. RULES: - Educational only: no buy/sell/hold language, no "you should have". The market price is "what the market paid", never "the right answer". - Explanations teach: each one explains why the wrong options are wrong. - Both questions set cardRefs (same unit only). Use the card ids above. - Pack id: "blind-case-"; topic: "cases/blind-valuation"; generatedAt: today; title: "Blind Valuation Case : "; description must not reveal the identity; tags on cards: "data" plus "snapshot" (unit 1) or "reveal" (unit 2). - Write in the requested language. ``` --- ## Notes for case authors - **Pick the snapshot date carefully.** The day the annual report was published is ideal: the learner gets a full year of audited numbers and the price already reflects them. - **Three to eight years old** is the sweet spot: old enough that the outcome is known, recent enough that the industry context is recognisable. - **Don't only pick disasters.** A library of cases where every reveal is a collapse teaches learners that every cheap stock is a trap. Include cases where a high multiple was justified and where a low multiple was simply right. - **US filers:** the SEC's XBRL company-facts API (`https://data.sec.gov/api/xbrl/companyfacts/CIK##########.json`) gives every reported value by period and is the fastest way to get audited numbers without transcription errors. Cross-check against the earnings release, because the API returns later restated values. - **Why only P/E and P/B.** Two numbers make the exercise a single, memorable judgement, and they check each other: P/B = P/E × ROE, with ROE computable from the cards. A learner whose two guesses imply wildly different ROEs has learned something before the reveal. - **Staleness badges are expected.** Unit 1 cards carry the snapshot date as `dataAsOf`, so the app will flag them as old data. That is correct: they *are* a historical snapshot.