The Deep Dive, Day 3

Phase 8 · Week 68 · ~20 hours

Where it sits. The fourth module of the signature-skill phase and the second half of the deep dive. M8.01 taught you to skim a filing in the right order and assemble the source stack; M8.02 compressed a whole business into an eight-line teardown; M8.03 ran Days 1–2 of the deep dive: you read a decade of reports oldest-to-newest, built the spreadsheet history, computed incremental ROIC, mapped the industry with Mauboussin's added-value lens, benchmarked two or three peers, built the 3-statement-lite model into base/bull/bear, and ran the reverse DCF to read what the price is implying. You now hold a business you understand numerically. What you do not yet have is the thing the market is paid to have and you are trying to beat: a qualitative edge, a defensible reason to believe your version of the future over the consensus version priced into the stock. That reason almost never comes from the spreadsheet, because everyone with a terminal has the spreadsheet. It comes from the work of Day 3: scuttlebutt (talking to and reading customers, ex-employees, suppliers, competitors, which is Fisher's method, taught here as a 15-point research checklist plus the modern toolkit), a hard read of the incentives (what management is actually paid to do, read off the US DEF 14A and India's Board's Report, related-party notes, and promoter-pledge disclosures), a disciplined listen across four to eight earnings calls spanning good and bad quarters, and the single highest-return hour in the whole method: steelmanning the bear. Then you run the written checklist, write the one-page thesis in a fixed six-part structure, and finish with a pre-mortem that converts your risks into a live monitoring dashboard. M8.05 takes all of this under a stopwatch and certifies it on companies you have never seen.

The company we carry. Day 3 continues the M8.03 deep dive on Titan Company Ltd (NSE: TITAN, the Tata-group jewellery, watches and eyewear company). A one-paragraph recap of the Day-1/2 findings opens §4.1; the full worked model thesis in §4.8 is built on Titan, and the US worked examples (the DEF 14A read, a second one-page thesis) are built on Copart Inc (Nasdaq: CPRT, the online salvage-auction operator). Every company figure here is illustrative, modeled on ~FY24/FY25 scale. Pull the actual filing before you rely on a number.

Suggested hour budget. Core content: 8h · worked examples: 2h · practice set: 4h · the mini-project (a full Day-3 pass with written thesis): 4h · flashcards + mastery check: 2h.

App link. This module powers the Rapid-Analysis System branch's judgment layer: the proxy-and-incentive case deck (comp-table and RPT excerpts → predict the behavior), the earnings-call tell drill (transcript excerpts → straight answer or dodge, promise kept or broken), the steelman generator (given a bull thesis, the engine grades whether your bear case is stronger than the strawman), and the one-page-thesis rubric grader that scores your write-up against the model answer and feeds it to your permanent company file.

Learning objectives

By the end you can:

1. Run a scuttlebutt program from Fisher's 15 points, treating each point as a research question with a named source, knowing exactly what a customer, an ex-employee, a supplier, and a competitor each reveals, and which of the 15 the numbers can already answer versus which require primary legwork.
2. Deploy the modern qualitative toolkit: expert networks (Tegus, AlphaSense), channel checks, Glassdoor/AmbitionBox, app reviews, and web-traffic/alt-data. Know precisely what each reveals and where each lies, and reconstruct all of it as a solo learner with only free sources.
3. Read the incentives cold from a US DEF 14A and an Indian Board's Report + related-party annexure + promoter-pledge disclosure, and predict the capital-allocation and disclosure behavior each structure buys (applying and extending the M3.07 §4.8 framework as a mandatory Day-3 step).
4. Review four to eight earnings calls across good and bad quarters and score management on the four tells, learning most from how they handle adversity: do they answer straight, own mistakes specifically, keep prior promises (a quarter-by-quarter promise ledger), and hold a consistent story?
5. Steelman the bear: locate the short/bear thesis, rebuild it stronger than its authors made it, and convert it into a short list of falsifiable break-variables, each with a monitoring metric; and classify it as a valuation-bear (wait) or a thesis-bear (fatal).
6. Run the written checklist and write a one-page thesis in the exact six-part structure (business · why it's a good business · the variant perception · the three things that must go right · valuation & expected return · sell criteria) so that all five spine questions are answerable from one page.
7. Conduct a pre-mortem ("it is three years later and this was a disaster — what happened?") and turn each disaster path into an early-warning trigger and a pre-committed action.

Prerequisites & connections

Builds on. M8.03 most directly: the history spreadsheet, the incremental-ROIC exhibit, the peer benchmark, the base/bull/bear model, and above all the reverse DCF are the raw material Day 3 pressure-tests, and the variant perception is literally your view minus the reverse-DCF's implied view. From M3.07 §4.8–4.9, the incentive-reading framework (comp-metric → allocation bias) and the decade allocation scorecard are consumed here as finished tools and made a permanent Day-3 step. The written investment checklist you assembled in M6.05 is the gate you run in §4.7. M6.02 gave you Fisher's scuttlebutt and the growth school; M6.03, inversion and pre-mortems, and expected-value/base-rate thinking; M6.04, the confirmation bias that steelmanning exists to defeat. M4.04–M4.06 hold the moat lenses that scuttlebutt confirms or refutes in the field (a claimed switching cost that ex-employees and customers deny is not a moat). M2.06–M2.08 built the forensic reflex, now pointed at proxies, RPT notes, and management tone instead of the P&L. M5 is the sector playbook that tells you which scuttlebutt questions and KPIs matter for this business.

Feeds forward. M8.05 (reps & certification): the one-page thesis and pre-mortem are half of what the certification grades, now under time pressure on never-studied companies. Phase 9: the thesis's valuation-&-return section feeds position sizing (M9.01), and the sell criteria and pre-mortem triggers become the decision-journal entry (M9.03–M9.04) written before the outcome is known. Phase 10: the institutional capstone deep dives (M10.01–M10.02) are this method run unaided. Competency map: this module is a primary builder of C7 (evaluate management & capital allocation: the incentive read and call audit), C3 (understand a business model cold, because scuttlebutt closes the gaps the filings leave), and C10 (the signature skill, where the write-up is the deliverable), and it carries the ethics & governance thread (§8.1) into live judgment.

The one-sentence version of this module. The numbers are shared by everyone; your edge is what you learn about the customers, the culture, the incentives, and the bear case that the market has priced lazily, and the discipline to write it down as a falsifiable, one-page thesis you can be held to.

Core teaching content

4.1 Day 3 in context: the qualitative edge is the last mile

Roughly half to two-thirds of India's jewellery demand still goes to the neighbourhood goldsmith rather than to any brand at all. Hold that while you recall where M8.03 left you on Titan. The history spread showed a decade of revenue compounding in the mid-teens, driven overwhelmingly by the jewellery segment (Tanishq, Mia, CaratLane, Zoya, at ~85%+ of revenue and an even larger share of profit); watches and eyewear are smaller, steadier, and strategically a brand-and-distribution school for the jewellery machine. ROCE runs high for a retailer because the balance sheet is cleverer than it looks: Titan funds much of its gold inventory through the gold-metal-loan mechanism (it borrows physical gold, not cash, which simultaneously finances inventory and hedges gold-price risk), and takes customer advances, so invested capital per rupee of sales is low. Incremental ROIC on the store roll-out has been comfortably above the ~12% INR hurdle (as of mid-2026, check against your own build). The industry map placed Titan as the scaled, trusted, national brand in a structurally formalizing market: that unorganized half-to-two-thirds is losing share to the organized players, decade over decade. The reverse DCF (your Day-2 output) said the current price implies something like mid-teens revenue growth sustained for a long competitive-advantage period with EBIT margins holding around 12–13% and ROCE staying north of 25% (illustrative; reproduce it from your model). That is a demanding but not absurd set of expectations for a business of this quality. (All figures illustrative, modeled on ~FY24/FY25 scale. Pull the filings.)

Here is the problem Day 3 exists to solve. Everyone with a screener has roughly the same numbers, and any competent analyst can run roughly the same reverse DCF. If your entire case is "great business, long runway," you are not disagreeing with the market — you are agreeing with it and paying its price. An edge is a place where your view of the future is better-founded than the consensus view embedded in the stock. There are only three honest kinds of edge:

• Informational. You know something material that most of the market does not (rare, often fleeting, and for a solo analyst usually a mosaic of small legal facts, not one big secret).
• Analytical. You interpret the same public facts more correctly (the reverse DCF is an analytical-edge machine; so is a better read of the incentives or the moat's durability).
• Behavioral / time-horizon. You are willing to hold a view through discomfort that the market's quarterly clock cannot (the deepest and most durable edge for an individual).

Day 3 is where all three are manufactured. Scuttlebutt builds a legal informational mosaic: no single fact is material, but the assembled picture of customer loyalty, employee morale, and channel health is something the drive-by analyst does not have. The incentive read and the call audit are analytical edge on question 3 (are these good, aligned people?). Steelmanning the bear is analytical edge on question 5 (is the price beatable?). It tells you whether your variant perception survives the strongest counter-argument. And writing the falsifiable thesis with pre-committed sell criteria is what lets you hold through the behavioral discount later without fooling yourself. The five-questions spine gets its final, field-tested answers here, and the write-up is where they become a decision you can defend.

One framing to carry through the whole day: you are not trying to confirm your thesis; you are trying to find out whether it is true. M6.04 named the enemy: confirmation bias, the near-irresistible pull to weight the evidence that flatters the position you have already half-taken after two days of loving work. Every technique below is an engineered corrective. Scuttlebutt sends you to people with no reason to flatter you. The incentive read replaces your impression of management with what they are contractually paid to do. The call audit checks their promises against their results with a ledger, not a feeling. And the steelman forces you to build the case against yourself better than your opponents can.

Skip these and Day 3 becomes a two-day-old crush writing itself a love letter.

4.4 Reading the incentives: a worked pass (US DEF 14A and India Board's Report)

Fisher's point 15 (integrity) and point 12 (long-range horizon) are not judged from vibes. They are read from documents, using the M3.07 §4.8 framework: show me the incentive and I will show you the outcome. Day 3 makes this a mandatory pass, worked here end-to-end on both jurisdictions. The M3.07 mechanics are assumed; here you see them run live on the carried companies.

The United States: Copart's DEF 14A (SEC EDGAR, filed annually before the AGM). Go to the Compensation Discussion & Analysis, the summary-compensation and outstanding-equity tables, the Pay-versus-Performance table (Item 402(v)), and the beneficial-ownership table. Copart is a positive specimen and a superb teaching contrast to M3.07's IBM. What the read surfaces (illustrative, modeled on the filings):

• Ownership is the incentive. Founder Willis Johnson and long-time CEO/Executive Chairman Jay Adair hold large personal stakes (founder-family ownership in the high-single-digit to low-double-digit percent range historically). When insiders own that much stock, the dominant incentive is per-share value over decades, and it needs no clever metric. It is the M3.07 "operative incentive is ownership" case, the benign version of concentrated control.
• Front-loaded equity with price/time hurdles. Copart has repeatedly used large multi-year equity grants in place of cash salary (executives taking a nominal salary and a big option/PSU package that vests over five-plus years, often with stock-price hurdles). Read as a prediction: this buys long-horizon, per-share-oriented behavior and discourages quarter-management. That is exactly what Copart's actual record shows (relentless reinvestment into land and yards, opportunistic buybacks, no dividend, and decades of >20–30% ROIC). The metric loads the gun in the long-term direction; the ownership keeps it pointed there.
• The say-on-pay tell. Those mega-grants have at times drawn lower say-on-pay support and proxy-advisor pushback (dilution and "pay-for-performance" concerns). This is a real and instructive tension: an incentive structure can be long-term aligned yet dilutive/opaque enough to annoy governance screens. The analyst's job is to judge substance (alignment) over form (the score), while noting the dilution honestly.

Predicted behavior from the Copart proxy: disciplined, per-share-value, long-horizon capital allocation with a bias to reinvest at high returns and buy back rather than pay dividends, and the record confirms it. That is what a good incentive read looks like: the prediction and the decade agree (M3.07 §4.9).

India: Titan's Board's Report, related-party annexure, and promoter file. India scatters the same information across the annual report. Run the four-part Day-3 read:

1. The promoter and the pledge. Titan's "promoters" are Tata Sons and TIDCO (the Tamil Nadu government's industrial investment arm), together holding roughly half the equity (illustrative; pull the latest shareholding pattern). Two teaching points. First, check the pledge line and find zero: a pledged-promoter-share percentage of nil is a quiet positive (M2's shareholding-pattern reflex; a pledged Tata company would be a screaming anomaly). The absence of a red flag is itself information, so you confirm it, you don't assume it. Second, a state-government co-promoter (TIDCO) introduces a specific, checkable channel: like a PSU, a government owner can favour dividends to meet its own fiscal needs, which is worth watching in the payout policy, though at Titan the Tata partnership has kept allocation disciplined.
2. Related-party transactions: the Tata brand fee. Titan, like Tata operating companies, pays Tata Sons a brand-subscription/business-promotion fee for use of the Tata name (the Tata Brand Equity & Business Promotion scheme). This is a legitimate RPT, but the analyst's discipline (M3.07 §4.8) is to track it as a percentage of sales over time: a stable, small percentage is fine; a rising skim on flat performance is leakage minorities don't share. Read the AOC-2 / RPT note for its size and trend.
3. RPT as capital allocation: the CaratLane buyout. In 2023 Titan bought out co-founder Mithun Sacheti's remaining stake in CaratLane for roughly ₹4,600 crore (illustrative). This is a live capital-allocation and governance case: was the price fair to Titan's minorities, and was the process arm's-length? The fairness-opinion and valuation basis are exactly what M3.07 §4.8 point 5 ("merger/buyout of promoter-adjacent entities at friendly valuations") tells you to inspect. Here the verdict is broadly favourable (bringing a fast-growing, high-return digital brand fully in-house), but you check, you don't assume.
4. The remuneration smell test. The Section 197(12) disclosures in the Board's Report give each director's pay ÷ median-employee pay and pay-growth vs profit-growth. The five-second test: is remuneration compounding faster than profits toward the statutory caps? At a Tata company the answer is typically no, because this is professional, non-promoter management at market pay, which itself answers a piece of Fisher point 15.

Predicted behavior from the Titan file: professionally managed, minority-aligned, long-horizon allocation, with the two things to keep monitoring being the brand-fee trend and any TIDCO-driven payout pressure. As with Copart, prediction and record agree, which is precisely what lets you trust the qualitative story you are about to write.

The general lesson of the paired read: the US and India hide the same truth in different filings. The DEF 14A concentrates it in one document with rich comp tables; the Indian AR scatters it across the Board's Report, the RPT annexure, the corporate-governance report, and the shareholding pattern. The questions are identical: what are they paid to do, who really owns this, and where does cash leak sideways. Both are answered before you read a single investor deck.

4.7 Running the written checklist

Before you write a word of the thesis, run your written investment checklist, the one you assembled in M6.05 from the program's master template plus your own logged mistakes. The checklist is not a scoring rubric to pass; it is a pre-flight against your own known failure modes, the way a pilot runs the checklist regardless of experience (Gawande, The Checklist Manifesto). Its whole power is that it forces you to consider items your excitement wants to skip, and the items most likely to be skipped are the ones that have burned you before.

Run it as five blocks, each a gate that can stop the thesis:

1. Circle of competence. Can I actually explain how this business makes money and what kills it, without hand-waving? If not, stop: an incomplete Day 1–2 is not a research gap to paper over in the write-up; it is a pass.
2. Business quality and moat (Q2). Is ROIC durably above the cost of capital, proven in the numbers (M8.03) and confirmed (not contradicted) by scuttlebutt (§4.2–4.3)? Which of the moat sources, and is it widening or eroding?
3. Management and incentives (Q3). Does the incentive read (§4.4) and the call audit (§4.5) say aligned-and-candid, or extractive-and-promotional? Any single integrity red flag (Fisher 15) is a veto, not a deduction.
4. Risks and the bear (Q4). Have I steelmanned the bear (§4.6)? Do I have the break-variables written? Is the worst realistic case survivable (no permanent capital impairment I can't monitor for)?
5. Valuation and margin of safety (Q5). What must be true to justify the price (reverse DCF)? Is that plausible against base rates (M8.03, Mauboussin)? Where is my variant perception, and is it an edge or just agreement at a high price?

A checklist item that trips is not a nuisance to rationalize away. It is the checklist doing its job. The most valuable outcome of §4.7 is sometimes the decision not to write a thesis at all, because a gate stopped you. That is a win: the cheapest mistakes are the ones you decline to make. When all five blocks clear, you have earned the right to write the page.

4.9 The pre-mortem: inverting to a monitoring dashboard

The final step, and the cheapest insurance you will ever buy. Gary Klein's pre-mortem (via M6.03's inversion) is prospective hindsight: instead of asking "what are the risks?", which invites vague, easily-dismissed answers, you assume the failure has already happened and explain it. The prompt is exact: "It is three years later. This investment was a disaster — I lost most of the money. Write the story of what happened." The psychological trick is real: certainty of failure defeats the optimism that makes ordinary risk assessment toothless, and people generate far more, and far more specific, failure paths when the failure is stipulated than when it is hypothetical.

The method (three steps).

1. Write the disaster stories: specific, company-level, not "the market fell." Each is a concrete causal path to permanent loss.
2. Rank by probability × severity: which paths are both plausible and fatal (thesis-breakers), versus plausible-but-survivable or fatal-but-remote.
3. Derive an early-warning trigger for each: the specific, monitorable metric and threshold that would tell you this path is happening, early enough to act. This is what converts the pre-mortem from a worry list into a live monitoring dashboard wired to your sell criteria (§4.8, part 6).

Common mistakes & how experts think differently

Mistake 1: Treating scuttlebutt as confirmation instead of falsification. The amateur reads three bull articles and one glowing Glassdoor review and calls it "primary research." The expert goes looking for the disconfirming voice: the angry ex-employee, the one-star review that recurs, the competitor who explains why they're winning. The point of scuttlebutt is to find out whether the thesis is true, not to feel better about it. If your field work only ever agrees with your model, you are doing it wrong.

Mistake 2: Weighting one loud source instead of triangulating. A single Glassdoor rating, one channel-check store, or one web-traffic chart is noise, and each carries a known bias (angry leavers over-post, one store is unrepresentative, traffic ≠ revenue). The expert treats every tile as weak and trusts only the mosaic: the picture that survives when four independent sources point the same way. Novices get whipsawed by whichever source is loudest; professionals ask "what do all the sources jointly imply?"

Mistake 3: Judging management from the investor deck and the good-quarter call. Anyone sounds impressive when results are good and the slides are polished. Character shows under stress. The expert deliberately reads the bad-quarter transcript and builds the promise ledger across 4–8 quarters, because Fisher's point 14 (do they clam up when trouble hits?) is the whole game, and it is invisible in a single good call.

Mistake 4: Skipping the steelman because the thesis "already handles the risks." After two days of loving work, the risks in your own model are strawmen you have pre-rebutted. The expert knows this is exactly when the mind is least trustworthy and spends the highest-value hour rebuilding the bear case stronger than its authors. A thesis that has not survived a genuine steelman is a mood, not a conviction, and it will not survive the first drawdown.

Mistake 5: Confusing a valuation bear with a thesis bear. "The stock could fall" and "the business is broken" demand opposite actions, and conflating them is how people either panic-sell great businesses at lows or ride broken ones to zero. The expert classifies every bear point: a valuation bear means wait for a better price (the thesis stands); a thesis bear means there is no price at which the story is intact. The LGD risk to Titan is a thesis bear on the margin engine; "the multiple is high" is a valuation bear. Different variables, different responses.

Mistake 6: Writing a thesis with no variant perception. The most common "thesis" is a beautifully argued case that a great company is a great company, which is agreeing with the market at its price, not beating it. The expert insists on articulating, in one sentence, where and why my view of the future differs from the price's implied view. No variant perception, no edge, no reason to act (versus simply admiring). This is the discipline that separates an analyst from a fan.

Mistake 7: Leaving risks vague instead of converting them to triggers. "Competition could intensify" cannot be monitored, so it will be ignored until it's too late. The expert runs the pre-mortem to turn every risk into a falsifiable break-variable with a threshold ("studded margin down two years running") wired directly to a pre-committed sell criterion. Risks you can't measure are risks you won't act on.

Mistake 8: Writing sell criteria after you own it (or not at all). Decided in the grip of a position, "when to sell" becomes infinitely rationalizable: every fall is "a buying opportunity," every break-variable "temporary." The expert writes the sell criteria into the one-page thesis before buying, when judgment is coldest, precisely so the future, emotionally-invested self is bound by the present, clear-eyed one. This is the whole reason the thesis is written down and dated.

The through-line. Every expert habit here is a pre-commitment device against your own future bias. The numbers were the easy part and everyone has them. The hard part, and the edge, is the honesty infrastructure: sources that can contradict you, a bear case you built to hurt, a variant perception you can defend, and sell rules written before the money is on the table. Day 3 is less about learning new facts than about installing the discipline that lets facts overrule feelings.

Worked examples

Worked example 1: A full free-source scuttlebutt sprint (Copart, US)

The task: reconstruct an expert-network-grade qualitative read on Copart using only free sources, in about 90 minutes, and score the relevant Fisher points. (Illustrative, modeled on the public record; do your own pull.)

Step 1: Map the two-sided business first (so you know who to "call"). Copart runs online salvage auctions: sellers are mostly insurers disposing of total-loss vehicles; buyers are dismantlers, rebuilders, dealers, and exporters. So scuttlebutt targets are: insurers (sellers), the buyer community, ex-employees, and the competitor (IAA/RB Global).

Step 2: The competitor's mouth (Fisher 4, 11). Read IAA/RB Global's earnings calls and investor decks. When the weaker competitor keeps explaining its plans to "close the yard-density and cycle-time gap," it is telling you for free that Copart's yard density and processing speed are the industry benchmark, a competitor-confirmed moat (efficient scale + cost advantage). This single free read replaces an expensive expert call on "competitive positioning."

Step 3: The buyer community (Fisher 1, 4, 6). Read the Copart buyer app reviews and salvage-buyer forums/subreddits and YouTube "how to buy from Copart" channels. Recurring themes (deep buyer liquidity, national inventory, but complaints about fees) tell you the marketplace has strong network effects (buyers show up because the cars are there) and pricing power (fees rise and buyers still come). Pricing power that customers grumble about but pay is the strongest kind (Fisher 6).

Step 4: Culture and depth (Fisher 7–9). Glassdoor: read the trend and the text, not the average. Long-tenured management, promote-from-within culture, and the founder-family stewardship (Johnson/Adair, §4.4) show up as stability. Cross-check on LinkedIn for executive tenure: the absence of a senior-exit wave is a positive.

Step 5: Alt-data on the cheap (Fisher 1). Google Trends for "Copart" and salvage-buying interest as a demand proxy; LinkedIn company headcount trend and job postings (where are they hiring: new yards? tech? that's where they're investing, Fisher 2). No Sensor Tower needed to see the direction.

Step 6: Score and write the line. Fisher scorecard: runway (1–2) strong (rising total-loss frequency, global expansion); sales org / network (4) benchmark; margins & defense (5–6) elite and defended by fees customers pay grudgingly; culture/depth (7–9) stable, founder-stewarded; industry tell (11) yard density + insurer relationships = the moat; horizon/dilution/candor/integrity (12–15) long-horizon, low dilution, owner-aligned (§4.4). Output line for the file: "Free-source scuttlebutt confirms the numbers' moat — competitor-acknowledged efficient scale, customer-grudged pricing power, founder-stewarded culture; no field evidence contradicts the ROIC-persistence story." Total cost: zero rupees, ~90 minutes, and a read most drive-by analysts never do.

Worked example 2: The earnings-call promise ledger (side-by-side)

The task: show how the same management-quality question is answered by a ledger, not an impression. Two synthetic companies, same sector, over four quarters. (Labeled synthetic; clean numbers to teach the mechanic.)

"CandidCo" ledger:

"PromoCo" ledger:

Reading it. CandidCo scores strong on all four tells: straight answers, mistakes owned in the first person ("we over-ordered"), promises kept, story consistent. PromoCo fails all four: the same recovery reissued (Tell 3), "transitory headwinds" laundering the miss (Tell 2), and the pivot from margin to "pipeline" the instant margin turned ugly (Tell 4). Neither company's investor deck would tell you this. Only the ledger, built across quarters, does. When CandidCo promises "double-digit next year," it has earned belief; when PromoCo touts its "record pipeline," you now know to discount it. This is Fisher 14–15, made quantitative.

Worked example 3: A second one-page thesis (Copart, US), abbreviated

So you see the six-part structure applied to a US company too. (Illustrative; pull the 10-K/DEF 14A.)

COPART INC (CPRT) · one-page thesis · [date] · price [$x] · sources: 10-Ks, DEF 14A, competitor calls, buyer-community scuttlebutt (WE1) 1. Business. Online salvage-auction marketplace: insurers sell total-loss vehicles; a global buyer base bids. Makes money on fees from both sides, not on the cars. Real engine: the two-sided network + a hard-to-replicate land/yard bank near population centers. 2. Why good. Moat = network effects + efficient scale + cost advantage (yard density, zoning barriers), proven in persistent >20–30% ROIC and confirmed by a competitor who openly chases its benchmarks (WE1). Management founder-owned, long-horizon, per-share-oriented (§4.4 DEF 14A read). 3. Variant perception. Price implies durable high growth + margins; the variant view is that total-loss frequency structurally rises (car complexity → more cars totaled) and international + EV salvage extend the runway longer than the implied fade: the network moat compounds rather than fades. 4. Three things that must go right. (i) Total-loss frequency keeps rising; (ii) the network/yard moat holds vs IAA-RB Global; (iii) EV salvage economics don't structurally impair per-unit fees. 5. Valuation & return. Base [$range] on continued mid-teens growth; premium multiple = thin margin of safety. Interested below [$X]. 6. Sell criteria. Exit if: ROIC persistence breaks; the competitor closes the yard/cycle-time gap (moat erosion); EV salvage structurally cuts fee-per-unit; or price runs far past value.

The point of the pair: the structure is jurisdiction-agnostic. The Indian and US theses differ in content (promoter/RPT read vs DEF 14A read; gold/LGD vs total-loss frequency) but are identical in shape: six parts, all five spine questions answerable, a falsifiable variant perception, and sell criteria written before ownership.

Worked example 4: Titan (India, ₹): putting numbers on the variant perception, reverse-DCF implied growth, the thesis's expected value, and the bear-vs-base gap

The task: §4.8's write-up states the variant perception, the expected return, and the sell criteria, but a page of adjectives hides the arithmetic underneath. Here we make the Titan thesis (§4.8) numeric. Every input is illustrative and assumed for the exercise (not pulled figures; rebuild them from your own Day-2 model per M8.03); the point is the mechanic, not the specific rupee. Teaching specimen, not investment advice.

Assume Titan trades at ₹3,400/share, your normalized next-year free cash flow is ₹42/share, and the INR hurdle is 12%.

Step 1: the single-stage reverse DCF (a Gordon inversion). The simplest possible reverse DCF treats the price as a growing perpetuity: price = FCF₁ ÷ (r − g), so the implied growth is g = r − FCF₁/price. FCF yield = 42 ÷ 3,400 = 1.24%, so g = 12% − 1.24% = 10.76%. The price implies FCF compounding at ~10.8% in perpetuity, nearly the whole discount rate. This is a sanity check, not the model: value = FCF/(r−g) explodes as g → r, which is exactly why your M8.03 reverse DCF used an explicit multi-year fade to a terminal value. But the one-line read already earns its keep: the market is paying for aggressive, durable growth, so your variant perception has to be about whether that growth is achievable at a high return, not about whether Titan is a nice business (everyone agrees it is).

Step 2: the thesis's expected value (probability-weight base/bull/bear). From your model, intrinsic value is ₹3,800 base (probability 50%), ₹4,900 bull (25%), ₹2,300 bear (25%). Expected value = 0.50×3,800 + 0.25×4,900 + 0.25×2,300 = ₹3,700. Against the ₹3,400 price, that is (3,700 − 3,400)/3,400 = +8.8% of upside to fair value (a convergence move, not an annual return) — thin.

Step 3: the bear-vs-base gap and the margin of safety. The downside if the bear plays out is (3,800 − 2,300)/3,800 = 39.5% from base value (or (3,400 − 2,300)/3,400 = 32.4% from today's price). The margin of safety at ₹3,400 against the ₹3,800 base is only (3,800 − 3,400)/3,800 = 10.5%. The price that would carry a 25% margin of safety against the same base is 3,800 × (1 − 0.25) = ₹2,850, a property of these assumed inputs rather than a level to act at.

Interpretation. Positive expected value (+8.8% to fair value) alongside a ~40% bear gap and a 10.5% margin of safety is the arithmetic behind §4.8's valuation part. On these assumed inputs the price already respects the quality, so the research that matters is the runway-and-reinvestment variant perception: whether it holds is what would move the base case, and it is the next thing to test.

P, fcf, r = 3400, 42, 0.12
g = r - fcf/P                          # 0.1076  -> ~10.8% implied perpetual growth
EV = 0.50*3800 + 0.25*4900 + 0.25*2300 # 3700.0
upside = (EV - P)/P                    # +0.0882 -> +8.8% to fair value
bear_gap = (3800 - 2300)/3800          # 0.3947  -> 39.5%
mos = (3800 - P)/3800                  # 0.1053  -> 10.5% margin of safety
buy_below = 3800*(1 - 0.25)            # 2850    -> price for a 25% margin of safety

Worked example 5: Copart (US, $): from thesis to position size, the edge, and the Kelly ceiling

The task: §4.8's valuation-&-return section and Phase 9's sizing (M9.01) meet here. Once the write-up gives an expected value and a downside, how large should the position be? Inputs are your deep-dive judgment, assumed for the exercise (not filing data); Copart is a teaching specimen, not advice.

From the deep dive and the steelman (§4.6) you judge: a 60% probability the thesis is right and the stock returns +50% as it converges to value; 40% it is wrong (a thesis-bear or a de-rate) and you lose −50%.

Step 1: is there an edge at all? Expected value per $1 = p×gain − q×loss = 0.60×0.50 − 0.40×0.50 = +0.10, i.e. +10% expected. Positive, so a position is justified. You never size a bet before confirming the edge is positive (a negative-EV "cheap" stock is a value trap, not a bargain).

Step 2: full Kelly. For a two-outcome bet with gain fraction a (prob p) and loss fraction b (prob q), the growth-maximizing fraction is f* = p/b − q/a = 0.60/0.50 − 0.40/0.50 = 1.20 − 0.80 = 0.40: full Kelly says 40% of the portfolio in one stock. (Equivalently f* = (p·a − q·b)/(a·b) = 0.10/0.25 = 0.40.)

Step 3: why you never bet full Kelly on one name. Full Kelly assumes your p and payoffs are exactly right and the bets are independent and repeated. Neither holds for a single, correlated equity whose odds you estimated. Overbetting is punished asymmetrically (Kelly's downside is brutal), so practitioners use fractional Kelly (half-Kelly = 20%, quarter-Kelly = 10%) and a hard single-name cap (say 8%). The binding size here is min(quarter-Kelly 10%, cap 8%) = 8%.

Step 4: feel the difference. If the thesis-bear hits (−50%), an 8% position costs the portfolio −0.50×0.08 = −4%; the 40% full-Kelly position costs −0.50×0.40 = −20%, the kind of drawdown that ends careers if your "60%" was really 50%. The Kelly math informs the size; the cap and fractional scaling govern it, because the probability is the input you are least sure of. Phase 9 (M9.01) formalizes exactly this.

p, a, b = 0.60, 0.50, 0.50; q = 1 - p
edge   = p*a - q*b            # +0.10  -> +10% expected value per $1
f_full = p/b - q/a            # 0.40   -> full Kelly (also (p*a-q*b)/(a*b)=0.10/0.25)
f_qtr  = f_full/4             # 0.10   -> quarter-Kelly
size   = min(f_qtr, 0.08)     # 0.08   -> the 8% cap binds
hit_8  = -0.50*0.08           # -0.04  -> -4% of the portfolio
hit_40 = -0.50*0.40           # -0.20  -> -20% at full Kelly

Practice set

Ramped guided → independent → timed. Write your full answer before reading the solution.

P1 (guided: map Fisher to free sources). For a listed quick-service-restaurant chain you are deep-diving, name a free source that helps answer each of Fisher points 4 (sales/distribution org), 7 (labor relations), and 11 (industry-peculiar tell).

Solution. (4) Franchisee forums, Glassdoor reviews from store/area managers, and a channel check (visit outlets: service speed, cleanliness, staffing). (7) Glassdoor/AmbitionBox trend and text for crew and manager sentiment; attrition/turnover disclosures; news of labor disputes. (11) The QSR-specific tell is same-store sales growth (SSSG) split into traffic vs ticket and unit-economics (new-store payback, average unit volume), found in the company's and peers' investor decks and calls; a chain growing only on price/ticket with falling traffic is a red flag the generic reader misses.

P2 (guided: which points do the numbers already answer?). Of Fisher's 15 points, which three can your M8.03 spreadsheet already answer, so you should not spend field-scuttlebutt time re-deriving them? State what each is read from.

Solution. Point 5 (worthwhile margin) is read directly from the P&L margin stack. Point 6 (margin trend/defense) is read from the multi-year margin trend and mix disclosures (the mechanism still benefits from scuttlebutt, but the fact of it is in the numbers). Point 13 (will growth require dilutive equity) is read from the share-count trend over the decade. Field scuttlebutt's unique value is on 4, 7, 8, 9, 11, 12, 14, 15: culture, candor, integrity, and the industry tell, none of which appear in a filing.

P3 (Glassdoor interpretation: signal vs noise). A company's Glassdoor rating is 3.4/5. Ten recent reviews: six are generic ("long hours," "typical corporate"), but four independently say "sales are booked aggressively at quarter-end to hit targets; a lot reverses next quarter." How do you weight this, and what does it make you check in the financials?

Solution. Ignore the six generic gripes (present at almost every company; noise). The four specific, independent, convergent complaints about quarter-end channel-stuffing are exactly the signal scuttlebutt exists to catch. Weight them heavily despite the mediocre average rating. They are a testable hypothesis, so check the numbers: rising DSO/receivables vs sales, a pattern of strong quarter-ends followed by weak quarter-starts, revenue growth outrunning cash collection (CFO/NI divergence, Phase 2), and any lumpy quarter-end revenue. Scuttlebutt pointed the forensic tools at the right footnote. (This is Fisher 10 and 15 meeting M2.06.)

P4 (channel-check design). You want to sanity-check demand at a jewellery retailer before its results. Design a free channel check and state its single biggest limit.

Solution. Visit 3–4 stores across at least two cities and price tiers (e.g., a flagship and a smaller-town store), ideally on a weekend and a weekday: note footfall, billing-counter activity, discount/scheme signage, what staff push (studded vs plain), and inventory freshness; visit a direct competitor (Kalyan/Malabar) and a local goldsmith the same day for contrast; check the brand's app and recent reviews. Biggest limit: sample bias / non-representativeness. A few stores on a few days in a few cities is a tiny, possibly unrepresentative sample of a national chain, and you'll unconsciously see what you came to find. Correct for it: pre-write what you expect, visit competitor stores as a control, and treat the check as one tile in the mosaic, never a verdict.

P5 (earnings-call tell: score the excerpt). An analyst asks: "Your EBIT margin fell 180bp — was that the new-plant ramp or pricing pressure?" Management A: "As we flagged last quarter, ~150bp is the plant ramp, unwinding by Q4; ~30bp is a genuine price give-back in the export segment we're addressing by [X]." Management B: "We remain confident in our structural margin trajectory and the significant long-term opportunity ahead." Score each on Tells 1 and 2.

Solution. Management A. Tell 1 (answers the question): strong; they decompose the exact number asked about (150bp ramp + 30bp pricing) and tie it to prior guidance. Tell 2 (owns issues): strong; names the price give-back specifically and states the fix. This is candor. Management B. Tell 1: fails; it answers a question no one asked ("structural trajectory") and never addresses the 180bp. Tell 2: fails; no acknowledgment, passive optimism, zero specifics: a textbook dodge. On this exchange alone, A earns belief and B earns a discount; across 4–8 such exchanges the pattern becomes your management-candor verdict.

P6 (India incentives: promoter and RPT read). In an Indian company's disclosures you find: promoter holding 62%, of which 40% of the promoter stake is pledged; a royalty to a promoter-owned entity rising from 1.5% to 3.5% of sales over four years while EBIT margin was flat; and an inter-corporate deposit (ICD) to a promoter group company. State the incentive/allocation concern for each and the one-line verdict.

Solution. Pledged 40% of a 62% stake. The promoter has borrowed against control; a price fall can trigger forced sales and a control scramble, and it signals promoter-level cash stress that can drive value extraction from the listco (M2 shareholding reflex; the IL&FS/DHFL-adjacent warning). Royalty 1.5%→3.5% on flat margins. A rising pre-dividend skim to the promoter that minorities don't share; ~2% of sales is being routed sideways with no performance justification (M3.07 §4.8 point 1: leakage). ICD to a group company. Company cash lent to a related party rather than deployed on the §4.7 menu or returned; classic tunneling channel (Satyam reflex). Verdict: promoter-first allocation with active cash-stress signals: every rupee is at risk of being routed to the promoter's needs, not the minority's; incentive read fails, integrity (Fisher 15) in serious doubt, so likely a pass regardless of business quality.

P7 (steelman: build the bear). You are bullish on a fast-growing Indian SaaS company. Write the three strongest bear points and, for each, the falsifiable break-variable you would monitor.

Solution (a model; yours may differ if equally rigorous). (1) Net revenue retention is quietly falling: growth is masking churn; if NRR drops below ~100%, the compounding engine reverses. Break-variable: disclosed NRR/NDR trend; a print below 100–105%. (2) CAC payback is lengthening as it moves upmarket/saturates: growth is being bought, not earned; unit economics decay. Break-variable: CAC payback months and magic number trend; S&M % of revenue rising without commensurate net-new ARR. (3) A platform/hyperscaler or incumbent bundles the feature for free: the product becomes a commodity. Break-variable: gross margin and pricing trend; logo churn among mid-market customers; competitor bundling announcements. Then classify: (1) and (2) are thesis bears (the compounding model breaks); (3) is a thesis bear if it materializes but currently a watch. A bull thesis that hasn't written these down hasn't been stress-tested.

P8 (state the variant perception). Your reverse DCF says the price of a consumer-staples leader implies ~9% revenue growth and stable ~22% EBIT margins for 12 years. Your own work says rural recovery plus premiumization supports ~11% growth with 100–150bp of margin expansion. Write the variant perception in one sentence, and say whether it is an edge or mere agreement.

Solution. "The market prices a good-but-maturing staples business at ~9% growth with flat margins; the variant view is that the rural-demand recovery and premiumization give it ~11% growth and modest margin expansion, so the implied fade underrates both the top line and the mix — the disagreement is specifically about rural volume and premium-mix, both monitorable." This is an edge (analytical): it states a specific, falsifiable way your view of the future differs from the price's implied view (growth 11 vs 9, margins expanding vs flat), and names the variables that settle it. Contrast with "it's a great company with a long runway," which is agreement at the market's price: no edge, no reason to act.

P9 (the three things that must go right). For a hospital chain you are deep-diving (thesis: high-return bed expansion + case-mix improvement), write the three falsifiable conditions the thesis depends on.

Solution. (i) New beds reach target occupancy and ARPOB on schedule: incremental ROIC on the expansion clears the hurdle rather than diluting returns for years (the M3.07 §4.6 Trap-1 risk for capital-heavy growth). Monitor: new-hospital occupancy ramp and mature-vs-new EBITDA/bed. (ii) Case/payer mix keeps improving: higher-acuity, better-paying work grows faster than low-margin volume. Monitor: ARPOB trend and specialty mix. (iii) No margin-destroying price regulation or payer squeeze: the pricing environment holds. Monitor: regulatory news, payer-negotiation commentary, realization vs cost inflation. Three, falsifiable, each with a metric — not a ten-item risk-factor recital.

P10 (pre-mortem → triggers). Run a five-line pre-mortem on the P9 hospital thesis: "it's three years later, the investment was a disaster." Give five disaster paths, each with a monitoring trigger.

Solution. (1) New beds sat empty: occupancy ramp stalled, incremental ROIC below hurdle. Trigger: new-unit occupancy <60% at 24 months. (2) A pricing/regulatory shock cut realizations. Trigger: adverse regulation; realization growth turning negative. (3) Doctor attrition: key specialists left, taking patient volumes. Trigger: senior-clinician exits; specialty-revenue concentration + LinkedIn departures. (4) Debt-funded expansion into a demand air-pocket strained the balance sheet. Trigger: net debt/EBITDA rising past ~3× with occupancy soft. (5) Clinical/quality incident damaged the brand and drove patients away. Trigger: accreditation issues, litigation, review-sentiment drop. Each path is specific, causal, and wired to a number you can watch: a live dashboard, not a worry list.

P11 (independent: critique a weak thesis). A peer's "thesis" reads: "XYZ is a fantastic compounder with a wide moat, best-in-class management, and a huge TAM. It's a long-term buy." List everything wrong with it against the six-part structure.

Solution. It fails almost every part. No variant perception: it's pure agreement ("fantastic compounder") with no statement of how the author's view differs from the price; there is no edge, only admiration. No valuation: "long-term buy" at any price ignores that a great business above intrinsic value is a poor investment (there's no reverse-DCF, no range, no "what must be true"). No falsifiable "three things that must go right": "wide moat, huge TAM" are adjectives, not monitorable conditions. No sell criteria: nothing pre-committed, so the position is un-fireable. Weak on Q3: "best-in-class management" asserted, not evidenced (no incentive read, no call audit). It reads like a fan, not an analyst. A passing rewrite would add a specific variant perception, a valuation range with margin of safety, three falsifiable conditions, incentive/candor evidence, and written sell criteria.

P12 (timed: 30 min mini-thesis). Synthetic company "Meridian Logistics" (labeled synthetic): asset-light freight broker; take-rate on shipments matched between shippers and carriers; ROIC ~28%, growing ~20%, share count flat; management owns 15% and is paid on FCF/share + relative TSR; reverse DCF implies ~15% growth for 10 years at stable take-rate; bears say a freight recession + carrier-app disintermediation will crush take-rate. Write a six-part one-page thesis.

Solution (model: a passing answer).

1. Business. Asset-light digital freight broker; earns a take-rate matching shippers to carriers; no trucks, so capital-light and high-ROIC. Profit engine = the two-sided network's liquidity (density of shippers ↔ carriers). 2. Why good. Network-effect moat proven in 28% ROIC + flat share count (self-funded); management 15%-owned, paid on FCF/share + relative TSR, an incentive structure (M3.07 §4.8) predicting per-share discipline, not empire-building. 3. Variant perception. Price implies 15% growth at stable take-rate; the variant view is that digitization is still early and network density is widening the take-rate moat, so the market underrates both durability of take-rate and the runway: the disagreement is specifically about disintermediation risk, which network liquidity offsets. 4. Three things that must go right. (i) Take-rate holds (network density > app disintermediation); (ii) it keeps gaining share through a freight-cycle trough; (iii) incremental ROIC stays >hurdle as it scales. 5. Valuation & return. Base [range] on 15% growth/stable take-rate; bear = freight recession + take-rate compression; thin margin of safety at the current premium; interested below [$X]. 6. Sell criteria. Exit if take-rate declines two+ quarters (disintermediation realized); share stalls in a trough (moat weaker than believed); incremental ROIC breaks hurdle; or price runs far past value.

Self-score against the mini-project rubric: full marks require a falsifiable variant perception (not "great network"), a valuation range (not a target), and sell criteria tied to the bear's break-variables. If your answer named the disintermediation risk but didn't turn it into a monitorable take-rate trigger, that's the gap to close.

P13 (guided: reverse-DCF implied growth). A US compounder (Ridgeline Corp, synthetic) trades at $520; your normalized next-year FCF is $10.40/share; the USD hurdle is 9%. (a) Using a single-stage (Gordon) reverse DCF, what perpetual FCF growth does the price imply? (b) What does that tell you about where your variant perception must live?

Solution. (a) FCF yield = 10.40 ÷ 520 = 2.0%; the single-stage reverse DCF sets price = FCF₁/(r − g), so g = r − yield = 9% − 2% = 7%. The price implies ~7% FCF growth in perpetuity. (b) Since ~7% forever is already well above long-run GDP, the market is not underpricing quality — it is pricing durable high growth. Your edge, if any, must be an analytical claim that the runway supports more than 7% at a high reinvestment return (or lasts longer than the fade assumes), not the observation that "it's a great business," which is agreement at the price. Trap: inverting the wrong way (r + yield) or using the earnings yield instead of FCF.

P14 (guided: expected value and margin of safety). For an Indian FMCG name (Sahyadri Consumer, synthetic) your model gives intrinsic values ₹1,500 base (55%), ₹2,000 bull (20%), ₹900 bear (25%); the price is ₹1,450. (a) Expected value per share? (b) Expected upside to fair value? (c) Margin of safety vs base? (d) Buy or watch?

Solution. (a) EV = 0.55×1,500 + 0.20×2,000 + 0.25×900 = 825 + 400 + 225 = ₹1,450. (b) (1,450 − 1,450)/1,450 = 0%: the market price equals your probability-weighted value. (c) (1,500 − 1,450)/1,500 = 3.3%, negligible. (d) Watch, don't buy: zero expected-value edge and no margin of safety; you would need a lower price or a better-founded variant view before committing. Trap: quoting only the ₹2,000 bull ("it could be worth 2,000!") and ignoring the 25%-weighted ₹900 bear, the tail the steelman exists to size.

P15 (independent: bear-vs-base gap and classification). For a US company you value at $90 base and $54 bear, two bear stories survive your steelman: (A) the multiple de-rates because rates rose; (B) a new entrant structurally resets the moat and margins. (a) Compute the bear-vs-base downside. (b) Classify each bear. (c) Which one belongs in the sell criteria as a break-variable?

Solution. (a) (90 − 54)/90 = 40% downside if the bear is realized. (b) (A) is a valuation bear: the business is intact, only the price/multiple is contested → wait for a better entry. (B) is a thesis bear: the business itself is impaired → no price makes the story true. (c) The thesis bear (B) drives the monitored break-variable (a moat/margin metric with a threshold); the valuation bear (A) is an entry / margin-of-safety signal, not a thesis-killer. Trap: seeing one 40% gap and reacting to both stories the same way: panic-selling on (A) or buying the dip into (B).

P16 (independent: position sizing with Kelly). From a deep dive on Nayara Components (synthetic, India) you judge a 55% chance the thesis is right (+60%) and 45% wrong (−40%). (a) Expected value per ₹1? (b) Full-Kelly fraction? (c) Quarter-Kelly, and with an 8% single-name cap, the position? (d) One sentence: why not full Kelly?

Solution. (a) 0.55×0.60 − 0.45×0.40 = 0.33 − 0.18 = +0.15 → +15% expected. (b) f* = p/b − q/a = 0.55/0.40 − 0.45/0.60 = 1.375 − 0.75 = 0.625 → 62.5%. (c) Quarter-Kelly = 15.6%; with the cap, min(15.6%, 8%) = 8%. (d) Full Kelly assumes your 55% and the payoffs are exactly right and the bets independent. For one correlated equity with estimated odds, overbetting risks ruin, so you scale to a fraction and cap it. Trap: reporting the +15% edge as the position size, or betting the 62.5% Kelly literally.

P17 (independent: the variant perception, in numbers). Your reverse DCF says a consumer-durables leader's price implies ~9% revenue growth for 10 years at a flat 18% EBIT margin. Your own work supports ~12% growth with the margin rising 18%→21% over five years. (a) State the variant perception in one sentence. (b) Edge or agreement? (c) Name the two monitorable variables that settle it.

Solution. (a) "The price implies a maturing durables business at ~9% growth and flat 18% margins; the variant view is that distribution expansion and premiumization deliver ~12% growth and ~300bp of margin expansion, so the implied fade underrates both the top line and the mix." (b) An edge (analytical): it states a specific, falsifiable difference from the implied view (12 vs 9; 21 vs 18), not "great company, long runway." (c) Realized revenue growth vs the implied ~9%, and the EBIT-margin trajectory vs flat 18%. Trap: writing a variant perception with no number in it ("the market underestimates them"): unfalsifiable, hence not an edge.

P18 (timed/stretch: 20 min, integrated). For Vantage Analytics (synthetic, US): price $200; normalized FCF $6/share; USD hurdle 9%; your intrinsic values $230 base (50%), $320 bull (20%), $140 bear (30%). From the steelman you judge a 55% chance the thesis is right (+55%) vs 45% wrong (−35%). Compute, in order: (i) the single-stage reverse-DCF implied growth; (ii) the thesis's expected value and the upside to fair value; (iii) the bear-vs-base downside; (iv) full Kelly, quarter-Kelly, and, with a 7% cap, the position size; (v) a one-line verdict.

Solution. (i) FCF yield = 6/200 = 3% → g = 9% − 3% = 6% implied perpetual growth. (ii) EV = 0.50×230 + 0.20×320 + 0.30×140 = 115 + 64 + 42 = $221; upside = (221 − 200)/200 = +10.5% to fair value. (iii) (230 − 140)/230 = 39.1% downside if the bear hits. (iv) f* = 0.55/0.35 − 0.45/0.55 = 1.571 − 0.818 = 0.753 → 75.3% full Kelly; quarter-Kelly = 18.8%; capped = min(18.8%, 7%) = 7%. (v) "Positive edge (+10.5% to fair value) and a positive Kelly, but a ~39% bear gap and an absurd 75% raw Kelly mean a small, capped position (7%) — and given the thin single-digit upside, wait for a dip that widens the margin of safety before committing even that." Scoring: 5/5 with the arithmetic right in time = fluent; if you reported the raw 75% Kelly as the size, re-read Worked example 5. Trap: forgetting the cap binds below quarter-Kelly here, and treating +10.5% (a to-fair-value move, not an annual return) as large.

Applied mini-project

Run Day 3 end-to-end on a real company and produce the written thesis. Continue the company from your M8.03 Day-1/2 work (or, if working standalone, pick one from the starter bank; India: Asian Paints, Pidilite, DMart, TCS, Titan, Bajaj Finance; US: Costco, Visa, Moody's, Copart, Chipotle). Alternate jurisdiction from your last teardown. Budget ~4 hours.

Exactly what to pull and do:

1. Scuttlebutt (Fisher 15, free sources). Run the §4.2–4.3 stack: read one competitor's latest call/deck; mine Glassdoor/AmbitionBox (trend + text); read recent app and product reviews; do at least one real channel check if it's a consumer business (visit a store/branch, or use the product); pull Google Trends and the LinkedIn headcount trend. Write a Fisher-15 scorecard, marking which points the numbers already answered.
2. Incentives read. US: pull the DEF 14A from SEC EDGAR (CD&A metrics, Pay-versus-Performance, beneficial ownership). India: pull the Board's Report (Section 197(12) ratios), the RPT/AOC-2 annexure, the corporate-governance report, and the shareholding pattern (promoter holding + pledge %) from the AR / BSE-NSE. Predict the allocation and disclosure behavior; note anything to monitor.
3. Earnings-call audit. Read/listen to 4–8 quarters spanning at least one bad quarter (screener.in concalls for India; company IR / Seeking Alpha / AlphaStreet for US). Build the promise ledger and score the four tells; write the one-line candor verdict.
4. Steelman the bear. Find the bear case (short reports, VIC/SumZero/Seeking Alpha bears, risk factors, competitor decks); rebuild it stronger; write the break-variables; classify each as valuation-bear or thesis-bear.
5. Checklist + write-up. Run your M6.05 checklist; write the one-page thesis in the six-part structure; finish with the pre-mortem table (disaster paths → triggers). File it in your permanent company notes.

Deliverable: the one-page thesis + the pre-mortem/monitoring table + the supporting Fisher scorecard, promise ledger, and incentive-read notes (2–3 pages of support behind the one-page front).

Scoring rubric (0/1/2 each; 20 points; pass ≥16 = 80%)

Log your total and your single biggest miss in the decision journal. A thesis that scores <16 is not ready to be sized (Phase 9). Restudy the weak block and rewrite.

Case Lab: run case CL-8.4 (The Two-Day-Old Crush, a synthetic mid-cap whose Day-1/2 numbers look wonderful) in the app's Cases tab for this module: you read the incentives from a scattered Indian annual report, build an 8-quarter promise ledger spanning one ugly quarter, steelman the bear into falsifiable break-variables, put numbers on the variant perception (reverse-DCF implied growth vs your view) and the expected value, and decide whether the one-page thesis clears the checklist or a gate stops you.

Reading & resources

Core books (the Day-3 canon).

• Philip Fisher, Common Stocks and Uncommon Profits, Ch. 2 (the 15 points) and Ch. 3 ("What Scuttlebutt Can Do for You"). [Paid][Beginner–Intermediate] The founding text; read the 15 points as a research checklist, exactly as taught here.
• Alfred Rappaport & Michael Mauboussin, Expectations Investing (rev. ed.). [Paid][Intermediate] The discipline behind the variant perception: reading the price's implied expectations and finding where they're beatable. Pairs with M3.05 and M8.03's reverse DCF.
• Aswath Damodaran, Narrative and Numbers. [Paid][Intermediate] Turning the qualitative story into disciplined numbers and back, the connective tissue of the whole write-up.
• Atul Gawande, The Checklist Manifesto. [Paid][Beginner] Why the pre-flight checklist beats expertise; the philosophy behind §4.7.
• Gary Klein, "Performing a Project Premortem" (Harvard Business Review, 2007). [Free][Beginner] The pre-mortem method in six pages.
• Thorndike, The Outsiders and Cunningham (ed.), The Essays of Warren Buffett, for the management/incentive judgment feeding §4.4 (also M3.07).

Incentive & governance filings (learn to read the primary documents).

• US: SEC EDGAR DEF 14A (full-text search at efts.sec.gov): CD&A, Pay-versus-Performance (Item 402(v)), beneficial ownership. [Free][Intermediate]
• India: the annual report's Board's Report + AOC-2 (RPT) + corporate-governance report + shareholding pattern (company IR, BSE/NSE, MCA21); screener.in for quick shareholding/pledge history. [Free][Intermediate]

Scuttlebutt & alt-data sources (the free solo stack).

• Glassdoor (global), AmbitionBox and MouthShut (India), Blind (tech), Indeed reviews: culture/attrition. [Free]
• App Store / Google Play reviews, Trustpilot, Amazon/Flipkart product reviews, category subreddits, YouTube reviews: customer voice. [Free]
• Google Trends (demand proxy), SimilarWeb (free tier, web traffic), LinkedIn company page (headcount trend, exec tenure, job postings). [Free/Freemium]
• Expert networks: Tegus, AlphaSense, GLG, Third Bridge. [Paid][Advanced] Know they exist and what they do; you approximate them free per §4.3.

Earnings-call transcripts.

• India: screener.in "concall" tab, Trendlyne, company IR, AlphaStreet India. [Free/Freemium]
• US: company IR, Seeking Alpha, AlphaStreet, Motley Fool transcripts, The Transcript (newsletter). [Free/Freemium]

Finding the bear case.

• Short-seller archives: Hindenburg, Muddy Waters, Kerrisdale, Grizzly, Viceroy, Bear Cave, Culper. [Free] Read for method, not verdict.
• Value Investors Club (free, delayed), SumZero, Seeking Alpha bear pieces, FinTwit/X, category subreddits. [Free/Freemium]
• The company's own 10-K/AR Risk Factors: the bear case its lawyers wrote for you. [Free]

Podcasts (recorded expert interviews, free scuttlebutt). Acquired, Business Breakdowns, Invest Like the Best (US); The Long View and India investor podcasts for local names. [Free][Beginner–Intermediate]

Going deeper.

• Going deeper: Michael Mauboussin, Measuring the Moat and What Does a Price-to-Earnings Multiple Mean? (Morgan Stanley / Counterpoint Global, free PDFs) [Free][Intermediate] The reverse-DCF's implied-growth and competitive-advantage-period math behind the variant perception (Worked example 4) and §4.8's valuation section; pairs with M8.03.
• Going deeper: William Poundstone, Fortune's Formula [Paid][Beginner–Intermediate] The story and intuition of the Kelly criterion (sizing a bet by its edge) that Worked example 5 previews and Phase 9 (M9.01) formalizes; the cautionary half of the book (why full Kelly ruins people) is exactly Step 3's point.

The Modern Analyst's Addendum

Everything above teaches this skill from first principles, by hand — that is how you learn it, and the mastery check still tests it that way. This addendum shows how a working analyst amplifies the same skill today. It adds; it never replaces. (R1/R10)

AI-Augment this skill

{
  "skill": "Day 3's qualitative edge: the 15-point scuttlebutt checklist, the incentive read from a DEF 14A or a Board's Report, four to eight earnings calls read as a time series, the steelmanned bear, the one-page thesis, and the pre-mortem turned into a monitoring dashboard",
  "use": "Day 3 is the day with the most documents and the least arithmetic, which makes RETRIEVAL OVER YOUR OWN DOCUMENT SET the one AI pattern that is genuinely defensible here — and it is defensible for exactly one reason: every claim can be pinned to a call date and a page, so nothing has to be taken on trust. Three legitimate jobs, all of them locating rather than judging. (1) THE PROMISE LEDGER of section 4.5 and Worked example 2: eight transcripts is a hundred thousand words, the ledger needs the guidance sentence and the later outcome sentence side by side, and 'find every forward-looking quantitative statement in these eight calls, with the call date and the paragraph' is a search problem with a checkable answer. You still grade whether the promise was kept. (2) LOCATING the incentive text: which pay metrics, what vesting, what the peer group is, where the related-party schedule sits — the DEF 14A is long and the Board's Report is buried, and finding a table is not the same as reading it. (3) FINDING bear material that already exists in the world — a short report, a competitor's call describing your target's weakness, a ValuePickr thread — which section 4.3 already licenses as the free approximation of an expert network. THE BANS, and the middle one is the module's whole argument: never the incentive VERDICT (section 4.4 is a judgement about what a pay plan rewards, and a plan that pays on adjusted EBITDA while the cash flow rots is a sentence you have to write yourself); never THE STEELMAN, because section 4.6 calls writing the bear case the highest-value hour in the method and its value is entirely in the discomfort of having to hold the other side yourself — a bear case someone else produced is a bear case you never had to face, which is the precise opposite of the exercise; and never THE ONE-PAGE THESIS, which is the day's deliverable and the artifact the certification grades. A thesis a model wrote is fluent, correctly structured, and evidence of nothing.",
  "tools": ["Document/PDF AI with retrieval over your own downloaded set — eight transcripts, the proxy or Board's Report, the prior year's thesis — with a call date or page locator required on every claim", "EDGAR: the DEF 14A for pay metrics and Form 4 for insider transactions, both free and both structured; data.sec.gov/submissions for the filing history that dates them", "screener.in's concalls tab, BSE/NSE announcement archives, and SEBI LODR filings — the Indian equivalents, in PDF rather than in XML", "The free scuttlebutt stack section 4.3 already names — Glassdoor and AmbitionBox, app-store reviews, Google Trends, LinkedIn headcount by function — read for trend and verbatim text, never for a star average", "A notebook for the Worked-example-4 and -5 arithmetic: the Gordon inversion, the probability-weighted expected value, and the Kelly fraction with its sensitivity"],
  "prompt": "Attached are <COMPANY>'s last eight earnings-call transcripts, oldest to newest. Build a table with one row per forward-looking quantitative statement management made — margin, growth, capacity, store count, mix — with these columns and nothing else: call date, speaker, the statement quoted verbatim, and the paragraph or page locator. Then, as a SEPARATE table, list every later statement that refers back to any of those topics, with the same four columns. Do not pair them up, do not say whether a promise was met, do not characterise tone, and do not include any sentence you cannot locate. I am building the ledger; you are finding the sentences.",
  "verify": "Make the citation structural rather than advisory, because on this day it is the only thing standing between a mosaic and a fabrication. The rule: every retrieved sentence carries a call date and a locator; you open a random sample of them before you believe any of them; and a locator that does not resolve invalidates the WHOLE extraction rather than that one row — a partially fabricated ledger is more dangerous than no ledger, since the true rows lend the false ones their credibility. Then verify the arithmetic against this module's own worked figures: Worked example 4 must give a 1.24% FCF yield and 10.76% implied perpetual growth at Rs 3,400 and Rs 42, an expected value of Rs 3,700, +8.8% to fair value, a 39.5% bear-vs-base gap, a 10.5% margin of safety and a Rs 2,850 entry for a 25% margin; Worked example 5 must give a +10% edge, a 40% full-Kelly fraction, 10% at quarter-Kelly, an 8% sized position and a -4% portfolio hit against -20% at full Kelly. And verify the one thing retrieval cannot: that the mosaic stayed on the legal side of section 4.3's line. Assembling non-material public facts into a material picture is the method; being handed one material non-public fact is a crime, and no tool in the stack knows which side of that line a sentence sits on.",
  "diy": "The gate is unaided: read a proxy or a Board's Report and say in one sentence what the pay plan actually rewards; read four calls and produce the promise ledger from your own notes; write the steelmanned bear case in your own words until you can argue it convincingly enough to unsettle yourself; write the one-page thesis in its six parts; and invert it into a monitoring dashboard with named tripwires. Nothing in this callout is admissible in that room, and M8.05 certifies the whole deliverable set."
}

Modern Data Analysis

By hand first. You built a promise ledger from calls read side by side (WE2), read a pay plan for what it rewards rather than what it pays (§4.4), wrote a bear case strong enough to unsettle your own thesis (§4.6), and turned a pre-mortem into a dashboard with named tripwires (§4.9). Then you put numbers on the variant perception (WE4) and on the position (WE5). Keep all of it. What the modern layer adds is not judgement; it is that eight transcripts stop being a reading marathon and become a searchable corpus, with one condition attached, and the condition is the whole section.

Today's workflow, and the condition. Retrieval over a document set you assembled yourself is the one place in Phase 8 where an AI tool sits comfortably inside the method, and the reason is narrow: every claim it returns points at a sentence you can go and read. That property is what makes it evidence rather than assertion, so make it structural, not advisory. Advisory looks like "ask it to cite its sources" and fails silently the first time it doesn't. Structural looks like three rules with teeth: the output format has a locator column and rows without one are dropped before you read them; you open a random sample of locators every session, not just when something looks odd; and a locator that does not resolve invalidates the entire extraction rather than its own row. That last rule is the one people resist and the one that matters, because a ledger that is 90% real is more dangerous than no ledger: the true rows lend their credibility to the false ones, and you have no way to tell which is which without redoing the work you were trying to avoid.

The structured half of Day 3, and where the two markets diverge. §4.3 already maps the unstructured scuttlebutt stack: Glassdoor and AmbitionBox for trend and text, app-store reviews for recurring specific failures, Google Trends and LinkedIn headcount as free alt-data. What it does not name is the part of the incentive read that arrives as data: in the US, Form 4 is a structured XML filing made within two business days of an insider transaction, so "did the people who signed this guidance buy any of it" is a query with a dated answer, and the DEF 14A gives you the pay metrics beside it. India's equivalent, SEBI's insider-trading disclosure regime filed to the exchanges, exists and is public, and it arrives as an announcement PDF on a portal rather than as a feed. So the same question costs a query on one side of the world and a scrape on the other. The compensating Indian asymmetry runs the other way: SEBI's LODR regime requires disclosures that have no US analog at all: the quarterly shareholding pattern with its promoter-pledge percentage under Reg 31, and the half-yearly related-party schedule under Reg 23. That means the two governance questions §4.4 cares most about are, in India, answered by a mandatory standing disclosure and, in the US, reconstructed from a proxy and a footnote.

Tools & sources (IN + US). US: the DEF 14A for the pay-metrics table, the peer group and the vesting terms; Form 4 for dated insider transactions and Form 3/5 around them; data.sec.gov/submissions/CIK##########.json for the filing history that dates every one of those; EDGAR full-text search to run a §4.5 tell across a decade of calls and filings at once; the 8-K trail for the events a promise ledger has to line up against; competitors' transcripts and decks, which §4.3 correctly calls the best free expert network there is. India: the Board's Report and the managerial-remuneration note (including the median-employee ratio) for the incentive read; SEBI Reg 31 shareholding-and-pledge and Reg 23 related-party disclosures; SEBI's insider-trading disclosures filed to BSE and NSE; screener.in's concalls tab for the transcript corpus itself; ValuePickr threads and broker bear notes as §4.6 hypothesis generators you verify against filings; MCA21 for the unlisted promoter entities a related-party schedule names but does not explain. Both: Damodaran's data pages for the cost-of-capital frame WE4's inversion needs, and the free scuttlebutt stack of §4.3. Run any code in your notebook.

# R3 — recomputed in-session on this module's own worked examples.
P, F, r = 3400.0, 42.0, 0.12                       # WE4: price, normalized FCF/share, INR hurdle
g = lambda p: r - F / p
print(f"FCF yield {F/P:.2%}, implied perpetual g {g(P):.2%}")            # 1.24% ; 10.76%
# Which input is the implied growth actually about?  dg/dr = 1 exactly.  dg/dP = F/P^2.
for mv in (0.10, 0.20, 0.30):
    print(f"  price -{mv:.0%} -> g {g(P*(1-mv)):.2%} ({100*(g(P*(1-mv))-g(P)):+.3f} pp)")
    #      -10% -> 10.63% (-0.137)   -20% -> 10.46% (-0.309)   -30% -> 10.24% (-0.529)
print(f"  price needed to move g by 1 pp on its own: {F/(F/P+0.01):.0f} "
      f"({(F/(F/P+0.01)-P)/P:+.1%})")                                    # 1879  (-44.7%)
print(f"  WE4's 25%-margin-of-safety price (Rs 2,850) implies g {g(2850):.2%} "
      f"— a {(2850-P)/P:+.1%} price cut buys {100*(g(2850)-g(P)):+.2f} pp")   # 10.53% ; -16.2% ; -0.24

# WE5: Kelly, and the sensitivity of the answer to the input you are least sure of.
a = b = 0.50
f_star = lambda p: p / b - (1 - p) / a
print(f"df*/dp = 1/a + 1/b = {1/a + 1/b:.0f}  -> 1 pt of p moves full Kelly {100*(1/a+1/b)/100:.0f} pts")
for p in (0.50, 0.55, 0.58, 0.60, 0.65):
    ev, fs = p*a - (1-p)*b, f_star(p)
    print(f"  p={p:.0%}: EV {ev:+.1%} | full Kelly {fs:.0%} | quarter {fs/4:.1%} "
          f"| sized min(qtr, 8% cap) {min(fs/4, 0.08):.1%}")
    #  50%: 0.0%, 0%, 0.0%, 0.0% | 55%: +5.0%, 20%, 5.0%, 5.0% | 58%: +8.0%, 32%, 8.0%, 8.0%
    #  60%: +10.0%, 40%, 10.0%, 8.0% | 65%: +15.0%, 60%, 15.0%, 8.0%

Verify. Prove the engine against this module's own worked figures before trusting it anywhere: WE4 must reproduce a 1.24% FCF yield and 10.76% implied perpetual growth, an expected value of ₹3,700, +8.8% to fair value, a bear-vs-base gap of 39.5% (32.4% from today's price), a margin of safety of 10.5% and a 25%-margin price of ₹2,850; WE5 must reproduce a +10% edge, 40% full Kelly, 10% at quarter-Kelly, an 8% sized position, and portfolio hits of −4% against −20%. All of them reproduce exactly. Then verify what the arithmetic quietly assumes. WE4's expected value uses probabilities of 50/25/25 that you assigned and cannot check; WE5's 60% is a number you produced from a steelman you wrote. The Quant lens below shows how hard that second number is working.

Quantitative lens

Both of this module's numeric worked examples end in a single figure (an implied growth rate and a position size), and both figures rest on one input each that the analyst supplies rather than observes. Quantifying how hard those inputs work is the useful thing to do with them. Every figure below comes from the block above, on the module's own assumed inputs; they are properties of that arithmetic, not measurements of any company or anyone's skill.

At a low free-cash-flow yield, the single-stage inversion is a statement about your hurdle, not about the price. WE4's Titan trades at ₹3,400 on ₹42 of normalized free cash flow, a 1.24% yield, so g = r − F/P gives 10.76%. Now move each input in turn. A one-point change in the required return moves implied growth by exactly one point, since ∂g/∂r = 1. A price move does almost nothing: a 10% fall shifts implied growth by 0.137 points, a 20% fall by 0.309, a 30% fall by 0.529. To move implied growth a single point on price alone the stock has to fall 44.7%, to ₹1,879. Read it against WE4's own conclusion and the point sharpens: the module's 25%-margin-of-safety entry at ₹2,850 is a 16.2% price cut that changes the implied growth by 0.24 points, from 10.76% to 10.53%. You would be buying a materially better price and a barely different implied story.

This is not a defect in the tool, it is the tool telling you something. At a 1.24% yield the inversion is dominated by r, so the sentence it supports is "at my 12% hurdle the price needs about 10.8% forever," and the same price at an 11% hurdle needs 9.8%, and at 13% needs 11.8%. Three analysts can look at one screen and honestly report three different "market expectations" that differ by exactly their hurdles. Which is why M8.01's WE6 spends four steps recovering a sell-side note's undisclosed r from its "3% priced in," why §4.8's write-up must state the required return it used, and why M8.03 §4.13 refuses the single-stage read for the real work and runs an explicit multi-year fade instead. The one-line inversion is a sanity check whose resolving power is proportional to the yield: sharp on a 6% free-cash-flow yield, nearly blind on a 1.2% one.

And the Kelly fraction is four times more sensitive than the probability it is built on. With WE5's symmetric ±50% payoffs, f* = p/b − q/a reduces to 4p − 2, so df*/dp = 4: a one-point error in your probability moves the full-Kelly recommendation four points. Walk the range your steelman could plausibly have produced. At p = 60% the edge is +10% and full Kelly is 40%. At 55% (a five-point overestimate, well inside what a judgement about a business can be wrong by) the edge halves to +5% and full Kelly halves to 20%. At 50% the edge is exactly zero and Kelly says bet nothing. So a ten-point overestimate of a probability you produced by thinking hard about a company converts a "+10% expected value" into a coin flip, while the sizing formula, read naively, would still have had you at 40% of the portfolio.

That is the quantitative case for §4.9's structure, and it is stronger than "be humble." Quarter-Kelly equals the 8% single-name cap at p = 58%, which means that across the entire plausible range below 58% the cap is what sizes the position, not the Kelly arithmetic. The formula's job is to confirm the edge exists and to tell you the sign; the cap does the sizing, precisely because the input the formula is most sensitive to is the input you are least able to defend. WE5's Step 4 makes the consequence concrete: an 8% position costs 4% of the portfolio if the bear hits, a 40% position costs 20%. The sensitivity above says how easily you could have been at 40% while believing you were being rigorous.

Honest limits. Every number here is arithmetic on assumed inputs: WE4's ₹42 of free cash flow, its 50/25/25 scenario probabilities and its ₹3,800/₹4,900/₹2,300 values are stated as illustrative in the module and are not pulled figures, and WE5's 60% is explicitly "your deep-dive judgment, assumed for the exercise." Nothing above measures Titan, Copart, or the accuracy of anybody's variant perceptions. The Kelly sensitivity is derived for the symmetric ±50% case; with asymmetric payoffs df*/dp = 1/a + 1/b, so a bet with a small downside is less sensitive to p and one with a large downside is more: the general lesson holds, the number 4 does not travel. The single-stage inversion is a growing perpetuity and therefore wrong in the way §4.8 and M8.03 both say it is: it has no fade, no reinvestment tie, and it explodes as g → r, which is why it is a check and not a model. And the deepest limit is the one the whole day is about: none of this arithmetic knows whether your probability is any good. Calibration (whether a stated 60% comes true 60% of the time) is measured over years of resolved journal entries, and M0.04 and M6.05 have already done that sample-size arithmetic; the honest summary is that a gross miscalibration is cheap to detect and a ten-point one is not. Go and read those sections rather than trusting a number produced here.

Do it in code: log the required return beside every implied-growth figure you ever write down; report the inversion's sensitivity (how much the price would have to move to shift the implied story a point) next to the story itself; and when you size a position, print the Kelly fraction at your probability and at your probability minus five points, before you print the position. See G1/QM1.01, G5/AI0.04 and G3/CN1.02.

Where this goes next: galaxy cross-links

• G5 · AI0.04 Retrieval & RAG — Grounding AI in Filings & Documents: the machinery that makes the promise ledger defensible: chunking eight transcripts, retrieving over them, and returning a locator with every claim so the citation rule is a property of the pipeline rather than a request in a prompt.
• G1 · QM1.01 Probability, Distributions, Sampling & Estimation: the expected value, the Kelly fraction and their sensitivities: what a probability-weighted outcome is, how error in an estimated probability propagates into a decision, and why the derivative is the number to look at rather than the point estimate.
• G3 · CN1.02 Hypothesis-Driven Problem Solving: the steelman as a discipline rather than a mood: stating the opposing claim precisely, naming the evidence that would settle it, and going to get that evidence, which is §4.6's hour in its general form.
• G5 · AI0.06 Failure Modes, Verification & the Primary-Source Guardrail: why a partially fabricated extraction is more dangerous than none, and what a verification rule looks like when it is structural rather than advisory.
• G5 · AI0.03 Prompt Engineering & Structured Prompting: forcing a locator column, refusing gap-filling, and separating extraction from judgement in the prompt itself, which is how the ledger prompt above is built.

Teach it back & journal

Feynman prompt. Explain to a smart 15-year-old why, after spending two days becoming convinced a business is great, the smartest thing you can do next is spend an hour trying as hard as you can to prove yourself wrong, and why you should go read what your company's enemies (short-sellers, angry ex-employees, rival companies) say about it. Use this analogy to make it stick: "It's like checking your homework by pretending you're the strictest teacher trying to fail you, instead of the proud kid who wants an A." Then write the one-page explainer: "An edge isn't knowing a company is good — everyone with a spreadsheet knows that, and the price already reflects it. An edge is one specific, testable thing you believe about its future that the price doesn't — and the honesty to write down, in advance, exactly what would prove you wrong." If the 15-year-old can't tell you the difference between "this is a great company" and "here's my one specific bet the market is getting wrong," the gap is yours to close.

Journal reflection. Take the company from this week's deep dive. Before re-reading your work, answer in writing: which of the four confirmation-bias correctives did I actually do, and which did I fake? Did I read the bad-quarter call or just the recent good one? Did I build the promise ledger or just form an impression? Did I steelman the bear or pre-rebut a strawman? Did I write the sell criteria before deciding I liked the stock, or after? Be brutal — the faked steps are exactly where a future loss is already hiding. Then record the single break-variable from your pre-mortem you are most tempted to dismiss, and commit to monitoring it. Phase 9 will size the position; this entry is what your future self reads when the price moves and you need to know whether you were disciplined or just lucky. File it, dated, never to be edited.

This module's flashcards and mastery quiz are wired into the app: see the node's Quiz and Reviews.