Reps & Certification

Phase 8 · Week 69 · ~20 hours · CROWN MODULE — THE M9 MILESTONE (SIGNATURE SKILL CERTIFIED)

Where it sits. Two timed sixty-minute teardowns and one full two-to-three-day deep dive, on companies you have never studied, scored against printable rubrics and cleared at the bar: that is the week. It is the last module of the signature-skill phase and the crown gate of the whole program's first three-quarters. M8.01 gave you the motor skills (skim a filing in the right order, run the Ctrl-F list, assemble the source stack); M8.02 assembled them into the twelve-question teardown and the eight-line output; M8.03–M8.04 took the same twelve questions and spent two to three days on them: the ten-year history and incremental ROIC, the industry map, the three-statement-lite model, the reverse DCF, the scuttlebutt, the incentive read, the steelman, the one-page thesis, and the pre-mortem. Those four modules taught you the method. What follows is about the difference between knowing the method and being the skill: performance under time pressure with the scaffolding removed. It does three things. First, it teaches you how to run a rep so that it actually improves you, through the deliberate-practice loop, self-grading against a rubric, and tracking your own systematic errors. Second, it hands you the scaffold-fading ladder and a large bank of practice targets so the reps have somewhere to go. Third, it certifies the skill against those rubrics. Passing certifies competency C10, the signature skill, and it is the milestone the entire program was backward-designed to deliver: understand any business in one to two hours, and know it deeply after a two-to-three-day deep dive. It unlocks Phase 9 (running money with the skill you have just proven you own).

Suggested hour budget. Core content (the deliberate-practice loop, the fading ladder, the failure-mode catalog, the time budget, the two rubrics, the target bank): 3.5h · practice set (rep-design, rubric-scoring, timed reverse-DCF and tripwire drills): 3h · Certification Part 1, two timed 60-minute teardowns on never-studied companies, self-graded: 2.5h · Certification Part 2, one full 2–3-day deep dive on a never-studied company (spread across the week; this is the bulk): ~10h · mastery check + flashcards + the calibration journal entry: 1h. The deep dive is where most of the twenty hours live; budget it across three sittings, not one.

App link. The certification tier of the Rapid-Analysis System branch. The app's timed 1-hour teardown mode (built in M8.02) becomes the certification instrument: it serves a never-served company's source stack, starts the 60-minute clock, auto-grades the objective parts (profit-engine ID, the ROIC-vs-hurdle call, the reverse-multiple arithmetic to a tolerance band, moat-source classification) and rubric-scores the eight-line output against a model answer. The "analyze this business" case engine hosts the deep-dive certification, rubric-scoring the one-page thesis and pre-mortem. Your finish times, per-criterion scores, catch/false-positive rates, and calibration all feed the mastery heatmap. After certification, the unified spaced-review scheduler keeps resurfacing skill reviews ("you haven't torn down a bank in 45 days") so the certified skill never quietly decays.

Learning objectives

By the end of the week you can:

1. Run a deliberate-practice rep that improves you rather than merely entertains you: produce the full output before peeking, self-grade it line-by-line against the rubric, diff it against a model answer, and log the systematic error the miss reveals.
2. Track your own error signature: maintain a miss taxonomy (coverage, calibration, speed, false-precision, sourcing, tripwire misses), read it for the one or two errors you make repeatedly, and design your next reps to attack exactly those.
3. Climb the scaffold-fading ladder deliberately: move a teardown from guided → faded → independent → timed, and a deep dive from guided → faded → independent, knowing precisely which support each rung removes and how you know you are ready to climb.
4. Name and defuse the failure modes that appear only under time pressure: rabbit-holing, anchoring on the story, skipping the reverse DCF, missing the related-party/pledge tripwires, and false precision. Each comes with the mechanical fix you run in the moment, not the resolution you make afterward.
5. Budget sixty minutes across the twelve questions so that a partial answer to all twelve beats a perfect answer to six, timeboxing each block and parking every rabbit-hole.
6. Score any teardown or deep dive against its printable rubric to the same standard a grader would, and read the total honestly (effort is not a criterion; the output clearing the bar is).
7. Select the right practice target for the skill you are building, from a graded bank of thirty-plus India and US companies: an easy clean-compounder to learn the shape of "good," a sector specimen to drill a lens, a cyclical/financial/holdco/turnaround to practise where the method is hardest.
8. Pass the certification: two timed 60-minute teardowns (one India, one US) and one full 2–3-day deep dive, all on never-studied companies, each scored at or above the rubric bar. Then run the lifelong weekly rep cadence that keeps the skill sharp for the decades after week 69.

Prerequisites & connections

Builds on. Everything, and M8.01–M8.04 most immediately. No new analytical technique appears here. The entire content is how to convert technique you already have into an automatic, fast, unscaffolded skill, and how to prove you have done so. From Phase 8: the source stack and 45–60-minute skim (M8.01); the twelve questions, the moat quick-test, and the eight-line output (M8.02); the ten-year history and incremental ROIC, the industry map, the three-statement-lite model, and the reverse DCF (M8.03); the scuttlebutt, the DEF 14A / Board's-report incentive read, the steelman, the one-page thesis, and the pre-mortem (M8.04). Underneath those: the three-statement linkage (M1.05), NOPAT and invested capital (M2.03), cash conversion (M2.04), the quality-of-earnings Ctrl-F reflexes and the red-flags checklist (M2.06–2.08), the reverse DCF and "what's priced in" (M3.05), value drivers and the g = reinvestment rate × ROIC discipline (M3.07), the four-lens moat protocol (M4.04–4.05), the sector KPI panels (Phase 5), the mental-model and personal checklists (M6.05), and the macro overlay (Phase 7). The two reusable playbooks, artifacts/playbook-1-2-hour-teardown.md and artifacts/playbook-2-3-day-deep-dive.md, are the templates you fade away from here.

Feeds forward. Phase 9 (Portfolio, Risk & Process) presumes a certified analyst: M9.01–M9.04 take the theses your certified skill produces and teach you to size them, hold them, sell them, and journal them. From week 70 the company-of-the-week becomes teardown → watchlist decision → simulator action, journaled (Section 5.2 of the master map). Phase 10's two institutional-grade capstone deep dives (M10.01–M10.02) are this certification's deep dive, repeated unaided at higher stakes and graded against an institutional rubric. Competency C10 is certified here; C3 (understand a business model cold) and C11 (communicate like a consultant) are exercised on every rep and every thesis.

The one-sentence version of this module. You have the method; this module makes it a reflex you can run cold, under a stopwatch, on a company you have never seen. Then it makes you prove it, twice on the clock and once in depth, against a rubric that does not care how hard you tried.

Core teaching content

4.1 From method to skill: the automaticity thesis

There is a chasm between knowing how to do a teardown and being able to do one: cold, fast, on an unfamiliar company, with a decision at the end. You have crossed chasms like it before without noticing. You know the rules of chess in an afternoon; playing a decent blitz game takes years, because blitz demands that the right move occur to you in seconds, not that you be able to reconstruct why it is right given an hour. A pilot can recite the engine-failure checklist on the ground; flying the aircraft through an actual failure, in weather, is a different competency built by simulator hours. The gap in both cases is the same one you are closing now: the gap between declarative knowledge (you can state the method) and procedural automaticity (you execute it under load without conscious effort, freeing your scarce attention for the judgment that actually matters).

Here is why the distinction is not academic for an analyst. A teardown has a fixed budget of attention, sixty to ninety minutes, and the twelve questions will consume all of it if each one still requires you to think about how to answer it. The learner who has to stop and recall "how do I get to a clean ROIC again?" spends the minute they needed for the actual judgment ("is this ROIC durable given the new entrant?") on plumbing. Automaticity is not about speed for its own sake; it is about spending your sixty minutes on the two or three judgments that decide the verdict, because everything else runs itself. The plumbing (walk revenue to FCF, build NOPAT over invested capital, reverse the multiple, scan for the pledge and the related-party note) must be so grooved that it happens while you think about something else.

That is the state being trained and certified.

Automaticity is built by one thing: deliberate reps with feedback. Not reading about teardowns (you are done reading), not watching someone else do one, not doing more teardowns the comfortable way you already do them well. What builds it is reps designed at the edge of your ability, produced at full effort, graded immediately against a standard, with the errors fed back into the next rep. You have already run roughly sixty scaffolded reps: the company-of-the-week since week 5, progressively overloaded (Section 5.2). Those built the base. This week tightens the loop, removes the last scaffolds, and puts the skill under certification stress so that when you leave Phase 8 the skill is yours, portable to any company, any market, with or without any of these materials, for the rest of your working life. That is the whole promise of the program, and it comes due this week.

The five questions are the invariant. Under all the machinery, every teardown and every deep dive answers the same five questions (the program's spine, Section 2.1 of the master map): (1) What does it do and how does it make money? (2) Is it a good business — ROIC durably above the cost of capital? (3) Is it run by honest, aligned people who allocate capital well? (4) What can kill it? (5) What is the price implying, and is that beatable? The twelve teardown questions and the deep-dive workbook are just the five questions with enough structure to be run fast and completely. When time pressure scrambles the structure (and it will), fall back to the five. They are the thing you are certifying automaticity on; the twelve questions are the training wheels for the five.

4.3 The scaffold-fading ladder

Scaffolding is the set of supports that let you produce a competent output before you could produce one unaided: the worked example you follow, the template that prompts each step, the open filing you can re-read, the untimed afternoon. Supports are how you start; the skill is not yours until the supports are gone. The error most self-learners make is to remove all scaffolds at once (jump from following a worked example straight to a cold timed test, fail, and conclude they are bad at it) or to remove none (do comfortable open-book untimed reads forever and never build speed). The fix is a ladder: remove one support at a time, in order, climbing only when the current rung is comfortable. You have been on this ladder since week 5.

Here it is made explicit for the final ascent.

The teardown ladder (four rungs).

Each rung removes exactly one class of support: worked answers (0→1), structural prompts (1→2), and spare time plus backtracking (2→3). Rung 3 is the certification standard for the teardown. The reason you climb rather than jump is that a support masks a weakness until it is removed: an open template hides that you cannot recall the twelve questions; unlimited time hides that your plumbing is slow. Fade one support and the weakness it masked surfaces in isolation, where you can drill it. That is exactly the "target at the edge" of deliberate practice (§4.2).

The deep-dive ladder (three rungs). The deep dive is longer, so its ladder is shorter and its rungs are days, not minutes.

Rung 2 is the certification standard for the deep dive. Note what "independent" does not mean: it does not mean working without the checklists and playbooks as reference. A professional keeps the red-flags checklist and the sector KPI cheat-sheets on the desk forever. It means working without the step-by-step prompting: nobody tells you "now compute incremental ROIC"; you know the deliverable set and you produce it.

The Phase-8 fading calendar. Across weeks 64–69 the ladder is walked once, deliberately. M8.01 (week 64) drilled the sub-skills in isolation; M8.02 (week 65) ran the guided and faded teardown rungs (Titan and Costco worked, then your own); M8.06 (week 66) widened the source stack beyond the annual report; M8.03–M8.04 (weeks 67–68) ran the guided and faded deep-dive rungs. Week 69 runs independent-then-timed teardowns and the independent deep dive, the top rungs, and certifies them. If a rung is not yet comfortable when certification week arrives, the slipping rule applies (Section 4.4 of the master map): do not climb on schedule; the calendar bends, the gates do not.

4.4 The failure modes under time pressure: and the fix for each

Untimed and open-book, you make few of these errors; the method is sound and you have the leisure to catch yourself. Add a sixty-minute clock and a never-seen company and a specific set of failure modes appears: predictable, nameable, and each with a mechanical fix you run in the moment. Learn the fix as a reflex, because under time pressure you will not have the spare attention to invent it. These are the errors the certification is built to expose, so drill them out before you sit it.

Failure mode 1 — Rabbit-holing. Symptom: twenty-five minutes gone and you are still reconstructing the exact revenue split of a minor segment, or chasing one footnote's tax detail, while eight questions remain untouched. Why under pressure: the interesting detail is more comfortable than the hard verdict; depth on one point feels like rigour while it is actually avoidance. Cost: coverage misses. You produce a beautiful answer to question 3 and no answer to questions 8–12, and the verdict was going to turn on question 11. The fix: the timebox and the parking lot. Every block gets a hard minute allocation (§4.5); when the timer for a block rings, you stop mid-sentence and move on, writing the unfinished thread into a "parking lot" line at the bottom of the page ("segment margin split — resolve in deep dive"). A teardown's job is breadth-to-a-verdict, not depth on any one point; depth is what the deep dive is for. The parking lot also does double duty: it becomes the deep dive's opening agenda.

Failure mode 2 — Anchoring on the story. Symptom: you read the investor deck or the bull note first, absorb management's narrative, and then every number you see afterwards is interpreted to fit it: the moat is "obviously" durable, the growth "clearly" continues. Why under pressure: the narrative is pre-packaged and cognitively cheap; building your own read from the numbers is expensive, and under a clock you reach for the cheap input. Cost: calibration misses. You certify management's story instead of testing it, and you miss the disconfirming trend sitting in the ten-year panel. The fix: the numbers-first order. Read the ten-year panel (revenue, margins, ROIC, cash conversion, share count, net debt) and form your provisional verdict before you open the deck or the bull note. Let the company's own history, not its narrative, anchor you. Then read the deck to see what management wants you to believe, and note the gap between the numbers and the narrative as a finding in itself. Soros's discipline in a sentence: the story is data about management, not evidence about the business.

Failure mode 3 — Skipping the reverse DCF. Symptom: you produce a rich qualitative read (great business, wide moat, honest managers) and then judge the price with a lazy "trades at 40x, looks expensive" or, worse, skip the price question because you ran out of time. Why under pressure: the reverse DCF requires arithmetic and the qualitative read is more fun; when time is short, the numeric question is the first casualty. Cost: you have answered four of the five spine questions and skipped the one where edge actually lives: question 5, what the price implies. A wonderful business at a price that already assumes a wonderful future is not an opportunity; a mediocre business priced for catastrophe can be. The fix: the reverse DCF is not optional and it goes in the time budget with protected minutes (§4.5 reserves ~8–10 minutes for it). Run the fast inversion every time: the Gordon read (implied perpetual growth g ≈ required return − FCF yield) as the floor, and the "how many years of high growth is this multiple paying for" cross-check. A teardown that does not end with "the price is implying X; I judge that beatable/not beatable because Y" has not done the job, however elegant the rest.

Failure mode 4 — Missing the related-party / pledge tripwires. Symptom: you run the "is it a good business" questions with care and never scan the shareholding pattern, the related-party note, or the ALM ladder, and you miss that the promoter has pledged 70% of a stake beside a leveraged group, or that "revenue" is a receivable from an associate that never collects. Why under pressure: the tripwires live in the boring back-pages (the shareholding pattern, note 30-something), and under a clock you spend your minutes on the exciting front-half. Cost: the worst kind of miss. You can be exactly right that it is a good business and lose all your capital to the governance landmine you never scanned for (the entire lesson of M2.08's India ten). The fix: the fixed tripwire scan, run on every company regardless of how clean it looks, as a non-negotiable reflex costing ~3 minutes: (a) the shareholding pattern, for promoter holding and pledge % (India); (b) the related-party note, asking whether material revenue or large advances are flowing to promoter-linked entities; (c) for lenders/NBFCs, the ALM ladder and CP reliance; (d) the auditor's report, for any qualification, KAM on a scary line, or mid-tenure resignation; (e) contingent liabilities and guarantees. You are not doing a forensic audit in three minutes; you are running the smoke detector. If it beeps, the price verdict becomes "un-ownable until resolved" no matter how good the business, and that becomes the variant question for the deep dive.

Failure mode 5 — False precision. Symptom: your teardown output says "fair value ₹2,847 per share" or "the stock is worth exactly 18.3x." Why under pressure: a precise number feels like authority and rigour, and under stress the feeling of rigour is comforting. Cost: precision you have not earned is worse than an honest range, because it invites you (and anyone reading you) to act on a false confidence: the "precise garbage" that a clean-looking model produces from shaky inputs. The fix: ranges and verdicts, never point targets, from a teardown. The teardown's price output is a verdict ("demanding," "generous," "roughly fair") supported by the reverse-DCF read, not a target price. Save the (still-a-range) numeric valuation for the deep dive, and even there present a base/bull/bear band, not a single number. The discipline is Damodaran's: valuation is a range, and the width of the range is information. A wide range is an honest signal that the business is hard to value, and pretending otherwise with a decimal point is a tell that you do not understand it.

A sixth, quieter failure — recency and the latest quarter. Symptom: you over-weight the most recent quarter's beat or miss and let it swing the whole verdict. Fix: the ten-year panel exists precisely so that one quarter is one data point among forty; read the through-cycle series, and treat the latest print as news to be contextualized, not as the trend.

The unifying lesson of the failure-mode catalog: under time pressure, structure protects you from yourself. The time budget defeats rabbit-holing; the numbers-first order defeats anchoring; the protected reverse-DCF minutes defeat skipping it; the fixed tripwire scan defeats the governance miss; the ranges-not-targets rule defeats false precision.

You do not out-discipline these failures with willpower in the moment — you out-structure them, in advance, so that the reflex is already built when the clock is running.

4.5 The sixty-minute time budget: minute by minute

The single most important tool for timed performance is a pre-committed allocation of the sixty minutes across the twelve questions, so that you never have to decide where your time goes while the clock runs. The decision is already made. The governing principle: a partial answer to all twelve questions beats a perfect answer to six, because the verdict is a function of the whole spine, and a spine missing question 11 (what's priced in) or question 9's tripwire scan is not a smaller teardown — it is a broken one. Timebox every block; when its timer rings, stop and move on, parking the unfinished thread. Here is the default budget for a normal industrial/consumer/services company; the sector variant follows.

Two features of this budget do most of the work. First, the numbers come before the narrative (Block A reads the panel, not the deck), a structural defence against anchoring. Second, the reverse DCF and the tripwire scan have protected minutes in Blocks E and D: the two things most likely to be skipped under pressure are the two things guaranteed a slot. Print this budget on the back of your one-page teardown sheet; run three or four reps against it with a visible countdown timer until the block transitions happen without your thinking about them. Speed is a graded outcome at certification, and it is built the same way as everything else here, by reps against a standard.

Sector variants (the budget bends to the business model). For a bank or NBFC, Block B shrinks and a new asset-quality block expands: you spend the freed minutes on GNPA/NNPA/PCR and slippage, the ALM ladder and CP reliance (NBFC), and Block E's reverse DCF becomes a P/B-vs-ROE read (justified P/B ≈ (ROE − g)/(r − g)), not an FCF-yield inversion, because banks have no clean FCF. For a cyclical or commodity, Block B normalizes to mid-cycle earnings and Block E refuses the trailing P/E entirely (lowest P/E at the peak is a sell signal), reading EV/EBITDA against replacement cost and the cost-curve position. For a hyper-growth money-loser, there is no ROIC yet, so Block C's moat evidence leans on unit-economics and cohort data, and Block E's reverse DCF is the whole analysis: the price is a bet on a future the current numbers do not yet show. Knowing which variant a company needs, and reaching for it in the first two minutes, is itself a graded skill; practice-set problem P6 drills it.

4.6 The two rubrics as practice instruments

The teardown rubric and the deep-dive rubric are printed in full in §8 (the certification), because they are the certification instrument. But their first and more frequent use is as a practice instrument on every rep you run, and using them that way is a skill worth stating explicitly.

Read the rubric before the rep. The rubric is a description of what "good" looks like, criterion by criterion. Reading it before you start a rep is not cheating. It is calibration: it tells you that "moat" means confirmed in the ROIC series, not asserted; that "what's priced in" means the reverse DCF actually run, not "looks expensive." Elite performers in every graded domain study the scoring standard before they perform to it. You are not memorizing answers; you are internalizing the standard.

Score after the rep, honestly, line by line. The self-grading protocol (the same one the whole program uses for written deliverables, Section 0.1 rule 1): produce the complete output → open the rubric → score each criterion 0, 1, or 2 against its printed anchor → total → compare to the bar → log the category of every criterion that scored below 2 into your miss taxonomy (§4.2). The honesty is the entire value. A generous self-grade ("I basically covered capital allocation") is a loan you take from your future self and repay at certification with interest. The rubric anchors exist to make the grade objective: a "2" on the reverse-DCF criterion has a printed definition, and either your output meets it or it does not, regardless of how insightful the rest was. Effort is not a criterion. The output clears the bar or it does not.

The rubric total is a signal, not a verdict on you. A 19/28 teardown is not a judgment that you are a poor analyst; it is a precise map of which criteria cost you the points, and therefore exactly what to drill next. Over a dozen reps the pattern of where you lose points is your error signature, and the signature is more useful than any single total. This is why you keep the scores: not to feel graded, but to aim your next reps.

4.7 The bank of practice targets

Reps need targets, and the target you choose determines the skill you build. The bank below is graded by difficulty and tagged by sector, with a note on what each company teaches, because a practice target is not a random company, it is a chosen exercise for a specific sub-skill. The organizing logic of difficulty:

• Difficulty 1 (Easy) — learn the shape of "good." Clean compounders with an obvious, durable moat: ROIC visibly and persistently above the cost of capital, simple accounting, light balance sheet, self-funding. You tear these down not because the verdict is in doubt but because you must know what a great business looks like in the numbers (the stable-high ROIC, the negative working capital, the FCF conversion) so that later you recognize deviations from it instantly. Start here; the shape you learn becomes your reference pattern.
• Difficulty 2 (Medium) — apply a lens or resolve a real tension. A sector-specific KPI set you must reach for (retail's SSSG and turns; SaaS's NRR and Rule of 40; a cyclical-lite industrial), or a genuine judgment call (a rich multiple on a great business; a moat under mild question; growth-vs-value tension). Here the method stops being automatic-easy and starts requiring the judgment the whole program built.
• Difficulty 3 (Hard) — where the method is hardest. Banks and NBFCs (different accounting; provisioning and ALM judgment); insurers (embedded value); commodities and deep cyclicals (mid-cycle earnings; the trailing-P/E trap); holding companies (sum-of-the-parts and the holdco discount); promoter-pledged or governance-complex groups (the tripwire scan is the whole game); and hyper-growth money-losers (no ROIC yet, so the reverse DCF is the analysis). These are the reps that expose whether you truly own the skill or only own it on easy companies.

Treat every company here as a teaching specimen, not a recommendation. Several are included precisely because their accounting, governance, or cyclicality is hard, which is the opposite of an endorsement. Statuses and structures change; confirm the current picture before you rely on any framing below (as of mid-2026). Figures you will pull are the reps; the bank just tells you where to point.

Difficulty 1 — Easy (learn the shape of "good"; India / US alternating).

Difficulty 2 — Medium (apply a lens or resolve a tension; India / US alternating).

Difficulty 3 — Hard (where the method is hardest; India / US alternating).

That is forty targets across two markets, three difficulty tiers, and roughly fifteen sectors, more than enough to run the certification and a year of weekly reps without repeating. Sequencing rule: do not tear down five easy compounders in a row (comfort reps that groove nothing) and do not start on a bank (you will drown). Interleave instead: an easy one, then a medium one, then a hard one; India, then US; a consumer name, then a financial, then a cyclical. Interleaving builds the retrieval strength that a novel certification company demands (Section 4.1 of the master map's learning science). And hold out a private sub-list of names you deliberately do not touch, so you always have genuinely never-studied companies for certification and for the periodic self-recertification of §4.8.

4.8 The ongoing weekly rep cadence: after certification, for life

Certification is a driver's license, not a trophy. The skill you prove in week 69 will decay if you stop using it, exactly like a language; the antidote is a permanent, low-friction rep cadence that most working analysts never formalize and that will compound into an enormous edge over the years. Here is the regimen, and it starts the week after you certify.

The weekly teardown continues — forever. The company-of-the-week does not end at Phase 8; it changes shape (Section 5.2 of the master map, weeks 70–80+): from week 70 each teardown ends not in a filed eight-line note but in a watchlist decision (own / watch / pass) and, in the app, a simulator action with a journalled thesis. One timed teardown a week is ~50 a year. The master map is precise about what this buys: the difference between the analyst you are at week 80 and the analyst you are at year four is "roughly 150 more teardowns, 30 more deep dives, and a few full market seasons lived through your own journal." That is not a metaphor. It is a rep count, and it is entirely within your control.

A deep dive a month (or a quarter). When a weekly teardown surfaces a company whose variant question is genuinely alive, where the market seems to be pricing something you think you can judge better, promote it to a full independent deep dive. Roughly one a month while you are building the watchlist, settling to one a quarter as the watchlist matures, is a sustainable rate that produces ~30 deep dives in the first few years, the master map's number again.

The feedback loop lengthens and deepens. During certification you lived on process feedback (the rubric). After it, the decision journal (installed in M0.04, and the spine of Phase 9) begins harvesting outcome feedback: you wrote a variant view and a set of "things that must go right" with dates; months later, reality grades them. The discipline (working-agreement eight, and the calibration dashboard) is to score the reasoning, not the P&L: a variant view that was well-reasoned and lost teaches more than one that was sloppy and won. Re-read the thesis you wrote, ask "was the process sound, independent of the outcome," and update your base rates. This is how a mechanically-competent analyst becomes a calibrated one: the difference between week 80 and year four in a single word.

Keep the error log for life; interleave so no lens rusts. The miss taxonomy (§4.2) is not a Phase-8 exercise. It is a permanent instrument. Re-read it before each teardown so your known failure modes are top-of-mind (the pre-flight checklist a pilot never outgrows). And interleave markets and sectors deliberately: if you have not valued a cyclical in forty days, value one before your instinct for mid-cycle normalization rusts. The app's spaced-skill scheduler will prompt exactly this, and away from the app you prompt yourself with the held-out bank.

Periodic self-recertification. Once or twice a year, sit a genuinely cold timed teardown on a held-out name and score it against the §8 rubric. If the score has slipped, you have found decay before it cost you money, and you know which criteria to re-drill. The certification is not a one-time event; it is a standard you return to, which is the only kind of standard that stays real.

Common mistakes & how experts think differently

The failure-mode catalog in §4.4 is about the timed rep itself; these are the deeper, mindset-level errors about practice and certification as a whole: the ones that keep an analyst plateaued for years despite plenty of activity.

1. Practicing what you are already good at. The novice reaches, again, for a clean FMCG compounder they read beautifully. It feels productive and produces a lovely output. The expert reaches for the bank, the cyclical, the holdco: the reps that feel bad, because those are the ones at the edge of ability, and the edge is the only place practice moves the needle (§4.2). If your last five reps were comfortable, you were maintaining, not improving. The rule: let your miss taxonomy pick your next target, not your mood.

2. Confusing fast reading with analysis. The novice measures progress by how quickly they can get through a 10-K. But reading is the input; the output is the skill: the twelve answered questions and the verdict. An expert who reads slowly but produces a sharp, decision-shaped eight lines beats a speed-reader who produces mush. Speed matters (it is graded), but it is speed to a good output, never speed for its own sake. Grade the output, not the reading time.

3. Peeking, and grading generously. The two halves of the same sin. Glancing at the model answer before finishing converts retrieval (which builds skill) into recognition (which builds only comfort); scoring your own rep a charitable "close enough" converts a diagnostic instrument into a self-esteem instrument. The expert produces the complete output cold, then grades it like a hostile examiner, because the only person a generous self-grade fools is the one who has to sit the certification. Write before you peek; grade against the anchor, not the effort.

4. Treating the teardown as an essay, not a decision. The novice writes a rich description of the company and stops, satisfied. The expert remembers that a teardown ends in a verdict and a variant question: good/average/poor business, demanding/generous price, and the one thing to resolve next. A teardown that does not commit to a verdict has not been done; description is not analysis, and the discipline of forcing an eight-line decision out of ninety minutes of reading is the entire point of the format.

5. Chasing completeness over the variant question. The novice tries to learn everything about the company and runs out of time at 40% coverage. The expert knows the goal is not a Wikipedia entry but the one thing the market appears to have wrong (the variant perception). You do not need to know everything; you need to know the two or three things that decide the verdict better than the market does. Completeness is the enemy of the variant question, and the variant question is where edge lives.

6. Mistaking the deep dive's length for its value. A three-day deep dive that produces a beautiful history spreadsheet, a detailed model, and no falsifiable variant view and no pre-mortem is a book report, not a thesis. The expert judges a deep dive by whether it produced (a) a differentiated, testable claim about the future and (b) an honest inversion of it in the pre-mortem, not by the page count. The rubric (§8) weights the thesis, reverse DCF, steelman, and pre-mortem precisely because those are where a long analysis either earns its length or reveals that it was just activity.

7. Certification anxiety — treating the timer as the enemy. The novice dreads the clock and the never-seen company as an unfair exam. The expert knows the timer is the skill being certified: a real analyst gets a name at 8am and an opinion is wanted by the close; the constraint is the job, not an artificial cruelty. The fix is not calming down. It is reps under the constraint until the constraint feels normal, exactly as a musician stops fearing the recital by performing, not by hoping. You have run ~60 scaffolded reps; the timer is the last scaffold to remove, not a new obstacle.

8. Cheating "never-studied." The temptation, in self-certification, to pick a company you secretly already half-know, or to peek at a bull note first. The expert understands that self-certification's only value is its honesty. The certification exists to tell you whether the skill is real, and a rigged pass tells you nothing except at the moment, later, when real money is on the line and the skill is not there. Genuine novelty (the held-out list, or a fresh screener name) is not a formality; it is the whole assay.

9. Thinking certification means done. The novice frames week 69 as a finish line. The expert frames it as a license to begin: the point at which the real education (thousands of reps against reality, market seasons lived through a journal) starts. The master map is explicit that mastery continues far beyond week 80; certification proves the skill is real, not that it is finished. Treating the license as a trophy is how skills quietly decay (§4.8).

10. Anchoring the self-grade on effort. "I worked so hard on that deep dive" is not a rubric criterion. The expert scores the output against the bar and is unmoved by how much sweat produced it — because the market is unmoved too. Effort that does not clear the bar is a signal to change how you practise (the target, the feedback, the error analysis), not a currency you can spend against the standard.

Worked examples

Three examples, done the way the method demands: a full model teardown you can grade yourself against, a full model one-page thesis you can grade yourself against, and a diagnosis of a failing rep so you can see what below-the-bar looks like and how to fix it. Company figures are round approximations labeled illustrative, modeled on ~FY24/FY25 scale: pull the actual filing and recompute as your drill; the companies are teaching specimens, not recommendations, and every valuation is a range.

WE1: A model 60-minute timed teardown: Copart (US)

This is a filled model answer for a Rung-3 (timed) teardown, the certification standard. Read it as "what a passing output looks like," then grade it against the teardown rubric (§8) yourself. Source stack assembled in the first minutes: Copart's 10-K Item 1 + latest investor materials; the last earnings-call transcript; the ten-year panel from stockanalysis.com; one competitor (IAA/RB Global) reference and one bear note.

The twelve questions, answered under the clock (terse, as they would actually be written):

1. What it does & the real profit engine. Online auctions of salvage/total-loss and specialty vehicles. Insurers consign wrecked cars; Copart stores, titles, and auctions them to a global buyer base (dismantlers, rebuilders, exporters, dealers). Revenue ~$4.4bn: mostly service fees, meaning seller (consignment/processing) and buyer fees plus transport/storage, with a smaller purchased-vehicle (principal) slice. The real profit engine is the fee-per-lot consignment model (Copart does not own most of the cars), run over owned land. The land is the capital and the moat, not a revenue line.
2. Customers & why they buy. Sellers are concentrated (a handful of large auto insurers = most volume) and deeply integrated: Copart plugs into their claims workflow and handles catastrophe surges. Buyers are a fragmented global base of ~thousands. Sellers stay because switching disrupts claims operations; buyers come for the largest inventory. Must-have for the seller's process, recurring by nature (cars crash every day).
3. $1 of revenue to FCF. ~38–40% operating margin; ~32% net margin (~$1.4bn NI). FCF < NI because Copart continually buys land (growth capex), leaving FCF ~$1.0–1.2bn. The "leak" is deliberate: land purchased ahead of need is the moat being widened, not money lost.
4. Unit economics. The repeatable unit is a lot sold: fees per lot × volume. Incremental lots on owned, already-capitalized land carry very high contribution margins (operating leverage on land + platform). No CAC in the SaaS sense; the two-sided marketplace is the acquisition engine.
5. Competition & right to win. Effectively a US duopoly with IAA (RB Global). Copart has led on share for years via owned land (vs IAA's leased) and a superior tech platform. Share stable-to-rising, the moat's numbers test.
6. Moat. Three reinforcing sources: network effect (more sellers → more buyers → better clearing prices → more sellers), efficient scale / cost advantage (owned salvage yards near cities are near-impossible to replicate: zoning, NIMBY, decades of assembly), and switching costs (integration into insurer claims). Confirmed by operating ROIC ~25% sustained across cycles: a real moat, not a hot product.
7. The economics. Growth = volume (rising total-loss frequency + international) + fee/mix. Margins very high and stable. Operating ROIC ~25% vs ~8.5% WACC, a ~16-point spread, the definition of a good business. Balance sheet: net cash (~$3–4bn cash, negligible debt). Cash conversion strong (CFO/EBITDA high; FCF/NI ~0.8, the gap being growth land-capex).
8. Capital allocation. Cash has gone into land and international expansion (reinvested at high incremental returns), plus periodic buybacks; no dividend, little debt. The reinvestment-at-high-ROIC record is the compounder signature, exactly what you want the profit engine's cash doing.
9. Management & incentives + tripwire scan. Founder-influenced, long-tenured, heavy insider ownership (aligned). Tripwire scan (US): proxy pay tied to performance; no pledge concept; related-party clean; auditor unqualified, no mid-tenure change; contingencies routine. Smoke detector silent.
10. The three risks that matter. (1) A structural fall in accident/total-loss frequency (safer cars, autonomy; long-dated); (2) regulatory/title-law or environmental change raising the cost of salvage yards; (3) the growth runway maturing faster than the multiple assumes.
11. What's priced in (reverse the multiple). Market cap ~$50bn; FCF ~$1.1bn → FCF yield ~2.2%. Gordon inversion at r = 8.5% USD: implied perpetual FCF growth g ≈ r − FCF yield = 8.5% − 2.2% = 6.3%. Cross-check: to justify ~33× earnings you need roughly a decade of low-double-digit growth fading to ~4%. So the price implies sustained above-nominal-GDP compounding (~6% forever, or ~a decade of low-teens then fade) — demanding, but not euphoric for a duopoly compounder.
12. Cheap or expensive vs quality. A truly excellent business at a demanding price. Not a bargain; not obviously mispriced. Edge here is not "is it a good business" (obviously yes) but whether the growth duration beats the ~6%-perpetual the price embeds.

The eight-line output:

1. Business: the dominant online auction house for salvage/total-loss vehicles, connecting insurers (sellers) to a global buyer base over owned salvage yards.
2. Revenue engine: fee-per-lot consignment (seller + buyer fees + transport/storage); volume × fee, on land it owns — asset-light on the cars, asset-heavy on the ground beneath them.
3. Unit economics: incremental lots on already-owned land carry very high contribution margins; operating leverage is structural.
4. Moat: network effect + efficient-scale/cost advantage (irreplaceable urban land) + switching costs: strong, confirmed by ~25% ROIC through cycles.
5. Economics verdict: good. ROIC ~25% vs ~8.5% WACC, net cash, strong cash conversion; a genuine compounder.
6. The three things that matter: total-loss frequency (secular up as cars get more electronic), international expansion runway, and fee/mix.
7. The two things that could break it: a structural decline in accidents/total-losses (autonomy, long-dated), or a regulatory/environmental shock to salvage-yard economics.
8. Price verdict + the one question: demanding (FCF yield ~2.2% implies ~6% perpetual growth / ~a decade of low-teens); the one question: what is the realistic ceiling on US total-loss frequency and how much international runway remains, because that decides whether ~6% perpetual is too low (upside) or the low-teens decade is already fully paid for.

Self-grade against the teardown rubric (§8), to show what a pass looks like:

Total: 28/28. It clears the bar (≥22/28). Note why it clears: it is not that every line is deep (a teardown is not deep), but that every question is genuinely addressed, the reverse DCF is actually run, the tripwire scan happened, and the output ends in a verdict and a single sharp variant question: the four things the rubric refuses to let you skip. Note also what a teardown deliberately leaves for the deep dive: the actual ceiling on total-loss frequency (Q11's variant question) is not resolved here. It is named, and named precisely, which is the teardown's entire job.

WE2: A model one-page thesis: Asian Paints (India)

This is the deep-dive output: the one-page thesis, the artifact the deep-dive rubric (§8) scores hardest. It presumes the two-to-three days of work behind it (the ten-year history, incremental ROIC, industry map, three-statement-lite model, reverse DCF, scuttlebutt, incentive read, and steelman) and distils them to one page. Asian Paints is chosen because it is, as of mid-2026, a genuine live moat-durability debate, the ideal shape for a thesis, because a thesis needs a variant perception, and a variant perception needs a real disagreement. Presented evenhandedly and illustratively; not advice; figures modeled on ~FY24/FY25 scale, verify against the filings.

Business. India's dominant decorative-paints company (~40%+ share of the organized decorative market), selling through an unmatched network of ~150,000+ dealers equipped with in-store tinting machines, backed by a best-in-class supply chain (near-two-day replenishment). Smaller industrial/auto (JV), international, and home-décor adjacencies. It makes money on paint volume × realization, with premiumization lifting mix.

Why it is a good business. The moat has four sources: brand (Asian Paints ≈ "paint" to the Indian consumer, for an infrequent, high-stakes purchase where trust dominates), distribution depth (dealer reach rivals cannot quickly match), a supply-chain/cost advantage (scale + logistics), and dealer switching costs (installed tinting machines, credit, service). The numbers prove it: ROCE historically ~25–30%, asset-light-ish with tight working capital, net cash, strong FCF conversion. Incremental ROIC on a decade of reinvestment has stayed well above the ~12% INR hurdle: a compounder that reinvested at high returns, which is the truest test.

The variant perception. The market, after Grasim/Aditya Birla's Birla Opus entered decorative paints in 2024 with ~₹10,000 crore of capacity and aggressive dealer incentives (plus JSW Paints and Pidilite's forays), has derated Asian Paints from ~55–60× toward ~45× (illustrative) and compressed near-term margins on the dealer-incentive war, pricing in effect a structural impairment of the moat. The variant to test is whether the derating over-extrapolates the entrant's early, subsidy-bought share gains: the distribution moat (150k dealers, installed tinting, two-day replenishment, brand trust) erodes far more slowly than a P&L incentive war suggests; a greenfield entrant needs years of utilization to earn its own cost of capital and a decade of dealer investment to match Asian Paints' service; and Indian paint volume (low per-capita consumption, housing tailwind) is a rising tide that lets the incumbent cede some share and still grow. If the moat's competitive-advantage period is intact, the incumbent's ROCE stays well above WACC through the disruption and the derated multiple under-prices that durability; if it is not, the derating is right. The evidence that decides between them is dealer counts, tinting-machine placements and the entrant's utilization, read quarter by quarter.

Valuation & expected return. Market cap ~₹2.8 lakh crore; FCF ~₹5,000 crore → FCF yield ~1.8%. Reverse DCF at r = 12% INR: implied perpetual growth g ≈ 12% − 1.8% = 10.2%, roughly India's nominal-GDP rate. So even after the derating, the price implies continued category leadership growing ~in line with nominal GDP forever at a high ROCE: no longer the old premium hyper-growth, but not a broken business either. Base/bull/bear from the three-statement-lite model, each stated as its assumptions: base, volume recovers to mid-single-digit, EBITDA margin stabilizes ~18–20%, ~low-teens earnings growth, flat multiple; bull, moat visibly holds, multiple re-rates toward ~55×; bear, permanent margin/volume impairment, multiple to FMCG-average ~30×. The expected return is the probability-weighted IRR across the three, computed in the model on the figures' own date rather than asserted on the page.

The three things that must go right. (1) Decorative volume growth re-accelerates to mid-single-digit+ as the category grows and the entrant's disruption normalizes; (2) EBITDA margins stabilize ~18–20% (the dealer-incentive war proves cyclical, not permanent); (3) the multiple holds ~45×+, i.e., the market does not conclude the moat is structurally broken.

Sell criteria (the thesis is wrong if…). (a) Asian Paints' decorative volume growth trails the category for more than four consecutive quarters (share loss faster than "cede-and-grow" allows); (b) gross/EBITDA margin steps down another leg and stays there (structural, not input-cost); (c) dealer count or exclusivity erodes materially (the moat's physical measure weakening); (d) capital allocation deteriorates: the paint cash cow funds value-destructive diversification.

Pre-mortem (it is 2029 and this was a mistake — what happened?). Birla Opus and JSW sustained losses far longer than any incumbent expected; they built "good-enough" distribution and tinting at ~20% lower price; paint commoditized the way it has in some other markets; Asian Paints defended share by cutting price and A&P, so volume and margin fell together; ROCE drifted from ~25% toward ~15%; and the market re-rated it to FMCG-average ~30×. Flat earnings times a compressed multiple = a lost five years. The pre-mortem's payoff: every one of those failure paths has a measurable early-warning sign (dealer exclusivity, incremental margins, the entrant's plant utilization and cash burn), and those signs are precisely the sell tripwires above. The inversion is what converts a hopeful story into a falsifiable, monitorable thesis.

Deep-dive rubric criteria this artifact exercises (see §8): incremental-ROIC and history (the ROCE series and reinvestment record), industry map (the entrant's threat placed in a five-forces/capital-cycle frame), three-statement-lite (the base/bull/bear), reverse DCF (the 1.8%-yield inversion), steelman (the bear case stated as the market's derating logic, then answered), one-page thesis (this whole artifact), and pre-mortem (the 2029 inversion). Scuttlebutt and the incentive read stand behind the "moat holds" claim (dealer channel checks; management's capital-allocation history and ownership). It is evenhanded by construction: the bear case is stated at full strength before the variant view answers it, and the valuation is a range, widest in the bear leg, because the honest width of the range is the information.

WE3: Diagnosing a failing rep (so you know what below-the-bar looks like)

The most useful thing you can study before certification is not another passing answer but a failing one, graded, because your job all week is to catch your own reps failing and fix them. Here is a learner's timed eight-line output on Meridian Logistics (synthetic; an Indian third-party logistics roll-up, promoter-controlled), followed by the grade.

The learner's output. 1. Business: a fast-growing logistics company. 2. Revenue engine: freight and warehousing, growing 30% a year. 3. Unit economics: good, scaling well. 4. Moat: strong — great management and market leader in its region. 5. Economics verdict: good — growing fast with expanding margins. 6. Three things that matter: growth, growth, and execution. 7. Two things that could break it: competition and a slowdown. 8. Price verdict: cheap — fair value ₹512 vs ₹430 today; buy.

Grade against the rubric: Q1 (1: "logistics company" is not the profit engine; which service actually earns the returns?); Q2 customers (0: absent); Q3 P&L→FCF (0: absent; a 30%-growing roll-up's cash is the whole question); Q4 unit economics (0: "good, scaling well" is an adjective, not a unit); Q5 competition (1: mentioned, no share direction); Q6 moat (0: "great management and market leader" are two of the four false moats M8.02 forbids; no ROIC evidence); Q7 economics vs hurdle (0: no ROIC, no cash conversion, no balance sheet, all critical for a roll-up); Q8 capital allocation (0: absent, and this is a roll-up, so acquisitions-vs-returns is central); Q9 management + tripwire scan (0: no pledge/RPT scan on a promoter-controlled roll-up, the exact landmine); Q10 risks (1: generic); Q11 reverse the multiple (0: not run); Q12 cheap/expensive (0: a point target from a teardown, false precision, and unsupported); output+variant question (0: no variant question; ends in "buy," not a question); time/process (0: anchored on the growth story, skipped the reverse DCF and the tripwire scan, and produced false precision).

Total: 3/28. It fails badly. And it fails in a diagnosable pattern: every failure mode from §4.4 at once:

• Anchoring on the story (failure mode 2): "great management," "market leader," "growing 30%". The narrative was absorbed and never tested against a ten-year cash series. Fix: numbers first, because a 30%-growing roll-up's ROIC and FCF are the first thing to compute, not the last.
• Missed tripwires (failure mode 4): a promoter-controlled roll-up and no pledge scan, no related-party scan: the single most important three minutes on this company, skipped. Roll-ups fund acquisitions with debt and equity and hide returns in goodwill; the governance scan is not optional here, it is the thesis.
• Skipped reverse DCF (failure mode 3): Q11 blank, so the one place edge lives was never touched, and then Q12 asserted "cheap" with no basis.
• False precision (failure mode 5): "₹512 vs ₹430": a decimal-point target from a ninety-minute read, the exact tell that the learner does not understand the business well enough to know how wide the range is.
• The false moats (a calibration miss): "great management" and "market leader" are two of the four advantages M8.02 explicitly rejects as moats, because they erode; a moat must be confirmed in a persistent-ROIC series, which was never computed.

The fix is not "try harder"; it is structural. Run the time budget (§4.5) so Q3, Q7, Q8, Q11 get their protected minutes; run the fixed tripwire scan (§4.4) on the promoter and related parties before believing any headline; read the ten-year cash series before the growth story; and end in a verdict and a question, not a target and a "buy." Re-run cold with those structures in place and the same learner produces a passing output. The difference is not talent or effort, it is the scaffolding-into-reflex this module exists to build. Log the miss as: coverage (Q2–4, 7–9, 11), calibration (false moats), tripwire (pledge/RPT), false precision (Q12); root cause: anchored on the growth deck, no numbers-first pass. That one log line, attacked over the next three reps, is worth more than a dozen comfortable teardowns.

WE4: The certification scoreboard: a full teardown numeric output and the compound gate (India)

This is certification day. You have finished all three performances; now you score them and apply the gate. Two things are worked here: what a passing timed-teardown numeric output looks like when the arithmetic is shown in full, and how the four component scores combine, because the certification is not one number but four independent gates, and a strong average cannot buy back a single failed one. Company figures are synthetic (clean numbers, labelled); recompute your own from the filing on the day.

Teardown A — Aarogya Diagnostics (synthetic), an Indian diagnostic-lab chain, torn down on a hard 60-minute clock. Ten-year panel (synthetic): revenue ₹3,000 crore (₹810 crore a decade ago, a ~14% CAGR), ROCE ~28% against a ~12% INR hurdle, PAT ₹520 crore, net cash, FCF ₹480 crore. Market cap ₹24,000 crore. The eight-line output:

1. Business: a branded diagnostic-lab network: patients and referring doctors pay per test across owned and franchised collection centres that feed regional reference labs.
2. Revenue engine: test volume × realization; the profit engine is asset-light franchise collection feeding high-throughput core labs, where incremental tests on an already-built lab carry very high contribution margin.
3. Unit economics: the repeatable unit is a test processed; once a reference lab clears its fixed-cost hurdle, operating leverage is steep.
4. Moat: brand/trust (patients do not gamble on a cheap unknown lab), a logistics/scale cost advantage, and doctor switching costs, all confirmed in ROCE ~28% sustained, not asserted.
5. Economics verdict: good: ~14% growth, ROCE ~28% vs a ~12% hurdle (a ~16-point spread), net cash, strong cash conversion.
6. The three things that matter: test-volume growth, the franchise-vs-own mix, and price competition from online and hospital-captive labs.
7. The two things that could break it: aggressive discounting by well-funded online-diagnostics entrants compressing realization; a regulatory price cap on common tests.
8. Price verdict + the one question: demanding. FCF yield 480 ÷ 24,000 = 2.0%; the fast Gordon inversion at r = 12% gives implied perpetual growth g ≈ 12% − 2.0% = 10.0% (solved exactly, g = (0.12 − 0.02) ÷ 1.02 = 9.8%), roughly India's nominal GDP, so the price already assumes category-leader compounding in perpetuity. The one question: can realization hold as funded online entrants buy share, because that decides whether ~10% perpetual is achievable or optimistic?

Score that output against the 14-criterion teardown rubric (§8). The per-criterion scores were [2,2,2,1,2,2,2,1,2,2,2,2,2,1] → total 25/28 (three 1s, no zeros). Check the pass conditions: 25 ≥ 22 ✓; no criterion scored 0 ✓; and the four hard gates, criteria 6 (moat-in-ROIC), 9 (tripwire scan), 11 (reverse DCF) and 13 (output + variant question), scored 2, 2, 2, 2, each ≥ 1 ✓. Teardown A clears the bar at 25/28 (89.3%).

The scoreboard — all four components. Teardown A (India) 25/28; teardown B (US, a separate never-studied company) 23/28; the deep dive 28/32 (with hard-gate criteria 2, 7, 11, 12, 13 all ≥ 1 and no zeros); the mastery check 11/12. Now apply the compound gate; every condition must hold simultaneously:

All four clear → C10 certified; the M9 milestone recorded; Phase 9 unlocks. The teaching point lives in the arithmetic of the gate, not any single score: the four are joined by AND, not averaged. Averaged, this candidate scores a comfortable ~87%. But a 23/28 teardown paired with a 30/32 deep dive would also average ~87%, and had that teardown instead been a 21/28 the whole certification would fail on that one component while the average still looked fine. A certification is a chain; it is exactly as strong as its weakest independent link, which is why you score each link against its own bar and never let a strong performance elsewhere console you for a weak one.

WE5: Scoring your own calibration: hit rate, the overconfidence gap, and the Brier score (US)

The reps and the certification live on process feedback (§4.2); the decision journal (§4.8) later harvests the slow, noisy outcome feedback: did the variant views actually come true? This example shows how to score that feedback so it improves you, using the calibration arithmetic Phase 9's dashboard formalizes. All calls and figures are synthetic. The lesson: a good win rate is not the same as good calibration, and the number that exposes the difference is one you compute yourself.

Over two years you logged thirty variant calls on US names, each tagged at the time with a confidence and later graded 1 (came true) or 0 (did not) by reality. One call, to show where a confidence comes from: Sentinel Systems (synthetic, US): market cap $45bn, FCF $900m → FCF yield 900 ÷ 45,000 = 2.0%; at r = 9% the price implied g ≈ 9% − 2.0% = 7.0% (exactly 6.9%), which you judged beatable and logged at 75% confidence. Grouped into three confidence bins:

Step 1 — the overall win rate looks fine. (8 + 6 + 5) ÷ (12 + 10 + 8) = 19 ÷ 30 = 63.3%. Nearly two calls in three came true. It feels like skill.

Step 2 — calibration, not the win rate, is the test. Calibration asks whether reality matched your stated confidence, bin by bin. Average stated confidence = (12×0.60 + 10×0.75 + 8×0.90) ÷ 30 = 21.9 ÷ 30 = 73.0%. You claimed 73% and hit 63.3%, an overconfidence gap of 9.7 percentage points. Worse, the gap widens with confidence: at ~60% you are slightly under-confident (hit 66.7%), at ~75% overconfident (hit 60.0%), and at ~90% badly overconfident (hit 62.5%). Confidence rising while accuracy does not is the single most common analyst pathology, and it is invisible in the 63.3% win rate.

Step 3 — the Brier score, one number for the whole record. The Brier score is the mean squared error of your probability forecasts, mean((p − outcome)²), lower is better:

[8×(0.6−1)² + 4×(0.6−0)²] + [6×(0.75−1)² + 4×(0.75−0)²] + [5×(0.9−1)² + 3×(0.9−0)²] = 2.72 + 2.625 + 2.48 = 7.825, divided by 30 = 0.261.

The reference point that gives it meaning: a forecaster who ignores every call and simply says "50%" every time scores 0.25. Your 0.261 is worse than that coin-flipper: the confident calls you got wrong (each 90%-bin miss costs 0.81) more than erased the credit for the ones you got right. Step 4 — the fix. Not "try to be right more" (unactionable), but discount every stated confidence by the ~10-point overconfidence you have now measured (a 90% call is really a ~65% call until your record earns otherwise) and widen the low-to-high spread only when the evidence justifies it. This is exactly why the journal logs a confidence with every thesis before the outcome is known (§4.8): you cannot compute this correction, or even see that you need it, without the honest record.

Practice set

Twelve problems, ramped guided (P1–P4, ~12 min each, method prompted) → independent (P5–P9, ~20 min, you choose the tools) → timed (P10–P12, exam conditions). The set drills the skill of practising the skill: reverse-DCF under the clock, tripwire scanning, rubric-scoring, rep design, sector budgeting. It includes two full timed teardown targets with model answers (P5, P12). All companies are synthetic unless named (clean numbers, labeled synthetic); work each fully before reading the solution.

P1 (guided — the reverse-DCF reflex, both markets). (a) NorthPeak Software (US): market cap $30bn; annual FCF $600m; required return r = 8.5%. Compute the FCF yield and the Gordon-implied perpetual growth, and judge it against US nominal GDP (~4%). (b) Bharat Staples (India): market cap ₹1,20,000 crore; FCF ₹2,400 crore; r = 12%. Same computation, judged against India nominal GDP (~10–11%).

Solution. (a) FCF yield = 600 ÷ 30,000 = 2.0%. Gordon inversion g ≈ r − FCF yield = 8.5% − 2.0% = 6.5% (exact, solving 0.02 = (0.085−g)/(1+g), gives g ≈ 6.4%). The price implies ~6.4–6.5% FCF growth forever, well above ~4% nominal GDP, so the market is pricing sustained above-GDP compounding: demanding; edge requires believing the growth duration exceeds what a mature-economy fade would give. (b) FCF yield = 2,400 ÷ 1,20,000 = 2.0%; g ≈ 12% − 2.0% = 10.0%, roughly India's nominal GDP, so the price implies category-leader growth in line with the economy in perpetuity: roughly fair for a durable leader, demanding only if the moat is in question. The lesson: identical 2.0% yields imply very different verdicts because the required return differs by market. Never reverse a multiple without the currency-matched r (M3.03).

P2 (guided — the fixed tripwire scan). Sundaram Agro Holdings (India, promoter-controlled). Shareholding pattern: promoter holding 58%, of which 46% pledged (up from 30% a year ago). Notes: "loans and advances to related parties" ₹1,900 crore = 22% of net worth (up from 9%); a corporate guarantee of ₹800 crore to a promoter-owned SPV. Auditor: unmodified opinion, but a new statutory auditor this year after the predecessor resigned mid-tenure citing information access. Run the tripwire scan and state the consequence for the price verdict.

Solution. Three independent tripwires converge: (1) pledge 46% and rising beside a promoter group, a fuse whose length is printed quarterly; (2) ecosystem leak, with RPT advances doubling to 22% of net worth plus an ₹800 crore guarantee to a promoter SPV (M2.08's play 7: the listed company's cash/credit migrating to the promoter's world); (3) mid-tenure auditor resignation, the highest-conviction public signal in India (M2.08). Convergence across three independent categories is a position decision, not a question. Consequence: the price verdict is "un-ownable until resolved," at any multiple, however good the underlying logistics business. The variant question for a deep dive is "where is the cash actually going, and is the listed entity becoming the ecosystem's lender of last resort?" Severity L3–L4. This is the exact scan that costs three minutes and saves the whole position.

P3 (guided — score the rep). A learner's timed eight-line output on CloudLedger Inc (US SaaS): "1. Business: cloud accounting software for SMBs. 2. Revenue engine: subscriptions, ~90% recurring, ~85% gross margin. 3. Unit economics: strong — customers stay for years. 4. Moat: switching costs — accounting data is sticky. 5. Economics verdict: good — 25% revenue growth, FCF-positive. 6. Three things: net retention, SMB churn, upsell. 7. Two risks: competition from incumbents, an SMB recession. 8. Price: trades at 12x sales — rich, but justified by growth." Score it against the rubric and name the decisive misses.

Solution. Strong on Q1–Q2 (2, 2), Q6 moat named and mechanism given (2: though ROIC-proof is weak for a young SaaS, so unit-economics evidence must carry it), Q5 competition (1), Q10 risks (2, specific), output structure mostly present. But the decisive failures: Q4 unit economics asserted, not computed (0: no CAC payback, LTV/CAC, NRR figure; the SaaS unit-economics panel is the whole point, M5.04); Q11 reverse the multiple not run (0: "12x sales, rich but justified" is an assertion, not an inversion; what growth-for-how-long does 12x sales imply?); Q12 verdict floats on that missing inversion (1); tripwire scan (US, so pledge N/A, but SBC dilution and the GAAP-vs-adjusted gap unaddressed: 1). Approximate total ~19–20/28, just under the ≥22 bar. The lesson: an output can be articulate and still fail, because the rubric protects the two things articulateness papers over: the actual unit-economics numbers and the actual reverse DCF. Fix: compute NRR and CAC payback; invert the 12x sales (e.g., what revenue CAGR sustains a ~30% FCF-margin terminal to justify it).

P4 (guided — design your next reps). Over 15 reps your miss taxonomy reads: coverage 2, calibration 1, speed 7, false-precision 0, sourcing 1, tripwire 4. Read your signature and design your next three reps.

Solution. Signature: speed (you get the right answers but need ~85–90 minutes) and tripwire (you skip the governance scan under pressure). Design: Rep 1: a timed 45-minute teardown of an easy company (a clean US compounder from Difficulty 1): compressing the clock on a company where the judgment is easy isolates and attacks the plumbing speed without confounding it with hard judgment. Rep 2: a teardown of a promoter-controlled Indian group with the tripwire scan run first, in the opening three minutes, before anything else, forcing the missed reflex to the front until it grooves. Rep 3: a bank (Difficulty 3), which drills both: banks demand a different budget (speed under a new structure) and their asset-quality/ALM scan is the tripwire reflex in another dress. Standing change for the next month: make the tripwire scan literally minute 0–3 of every rep. Note the meta-lesson: your fixes were chosen by the data, not by mood: the whole purpose of the miss log.

P5 (independent — a full timed teardown target, with model answer). Vindhya Consumer Foods (India, synthetic, packaged foods). Ten-year panel: revenue ₹8,000 crore (₹2,200 crore a decade ago → ~14% CAGR), gross margin 42%, EBITDA margin 16%, PAT ₹720 crore, ROCE 26%, net cash ₹900 crore, CFO/EBITDA 0.9, FCF/PAT 0.8, promoter 52% (pledge 0%), wide dealer distribution, volume growth mid-single-digit plus premiumization. Market cap ₹36,000 crore (P/E ~50×), FCF ~₹575 crore. Run the teardown; produce the eight-line output including the reverse DCF.

Solution (model output). 1. Business: a branded packaged-foods maker selling volume × realization through a wide dealer network, with premiumization lifting mix. 2. Revenue engine: branded staples/snacks; the profit engine is brand-plus-distribution pricing power, not any single SKU. 3. Unit economics: incremental volume on an established distribution base carries high contribution margin; negative-ish working capital typical of the model. 4. Moat: brand + distribution depth; confirmed in ROCE 26% vs ~12% hurdle: a real, if not enormous, moat; watch for modern-trade/D2C fragmentation. 5. Economics verdict: good: 14% growth, 26% ROCE, net cash, 0.8 FCF/PAT. 6. Three things that matter: volume growth durability, premiumization/mix, and distribution defense against D2C. 7. Two things that could break it: category fragmentation to regional/D2C brands eroding pricing power; a commodity-input spike compressing margin faster than price can follow. 8. Price verdict + question: demanding. FCF yield 575 ÷ 36,000 = 1.6%; Gordon g ≈ 12% − 1.6% = 10.4%, roughly nominal GDP, so priced for continued category leadership in perpetuity. The one question: can it hold mid-single-digit volume plus premiumization as modern trade and D2C fragment the category, because that decides whether ~10% perpetual is achievable or optimistic? (Score check: all twelve addressed, tripwire scan clean (pledge 0%, no RPT flags stated), reverse DCF run, output ends in a question. Clears the bar.)

P6 (independent — sector-variant budgeting). For three companies, state how the default 60-minute budget (§4.5) shifts and what reverse-DCF method replaces the FCF-yield inversion: (a) Kaveri Bank; (b) Deccan Steel (integrated steelmaker); (c) QuickCart (loss-making e-commerce).

Solution. (a) Bank: Block B (unit economics) shrinks; a new asset-quality block expands: NIM, GNPA/NNPA/PCR, slippage, CASA, CRAR/CET1, CD ratio. Reverse DCF → justified P/B = (ROE − g)/(r − g) vs actual P/B (banks have no clean FCF). Tripwire scan → RBI divergence disclosures, restructured book. (b) Steel (cyclical): Block B normalizes to mid-cycle earnings (never trailing); Block E refuses trailing P/E (lowest P/E at the peak is a sell signal), reading EV/EBITDA on mid-cycle EBITDA vs replacement cost and the cost-curve quartile; Block D weights net debt into a downturn. (c) Money-loser: no ROIC yet, so Block C's moat evidence leans on cohort/contribution-margin data; the reverse DCF is the thesis: solve for the revenue and terminal margin the price implies and judge that path's plausibility against base rates. The meta-skill: identify the variant in the first two minutes and reach for the right toolkit. Running the default FCF inversion on a bank or a trailing P/E on a steelmaker is itself a failing move.

P7 (independent — difficulty classification). Tag each company Difficulty 1/2/3 and state the sub-skill it trains: HUL (India FMCG), ICICI Bank (India), Copart (US), Vedanta (India metals), Adobe (US), Tata Steel (India).

Solution. HUL (1): clean staples compounder; learn the shape of good (stable-high ROCE, negative WC, FCF conversion). ICICI Bank (3): the bank lens and accounting (spread business, provisioning judgment, P/B-ROE), a different toolkit entirely. Copart (1, edging to 2 on the rich multiple): a textbook confirmable moat; learn to read the moat in the ROIC series (and to reverse a compounder's multiple). Vedanta (3): the hardest composite: commodity cyclicality and holdco/SOTP and high leverage and governance/RPT tripwires at once (teaching specimen). Adobe (2): the SaaS lens (NRR, Rule of 40, RPO) plus a live disruption judgment (generative AI). Tata Steel (3): the cyclical discipline: mid-cycle normalization and the trailing-P/E trap. The point of tagging before you practise: you choose the rep for the sub-skill you are building, not at random.

P8 (independent — spot the failure mode). A learner writes: "Everyone's saying Birla Opus is killing Asian Paints and the moat is dead, so I passed on it without bothering to model anything — why value a broken moat?" Name the failure mode(s) and the fix.

Solution. Two overlapping failures: anchoring on the (bear) story (failure mode 2: absorbing a pre-packaged narrative, here the consensus bear, instead of building a read from the numbers) and skipping the reverse DCF / numbers-first pass (failure mode 3), plus a dose of herding (M6.04). The fix is the numbers-first order and an actual inversion: the market has already derated the stock (from ~55–60× toward ~45×), so the reverse DCF tells you the bear case may be substantially priced in already: the implied ~10% perpetual growth is below the old premium. The disciplined question is not "is the moat under attack" (obviously) but "is the degree of impairment the price now implies too much or too little", which is exactly the variant analysis of WE2. Passing "without modeling anything" forfeits the one place edge could exist. A pass can be the right answer, but only after the inversion, never instead of it.

P9 (independent — reverse-DCF plausibility vs base rates). A reverse DCF on a high-flyer shows the price implies 18% revenue growth sustained for 10 years. How do you judge whether that is plausible, and what discipline applies?

Solution. Judge the implied number against the outside view / base rate (M6.03; Mauboussin's base-rate work): of the large universe of public companies, only a low-single-digit percentage sustain ~15–18%+ revenue growth for a full decade. Growth mean-reverts, and the faster the rate the rarer its persistence. So an implied 18%-for-10-years sits in the extreme right tail of the historical distribution: the market is pricing a tail outcome. That does not make it wrong (some companies are the tail), but it shifts the burden of proof: you must name the specific, durable reason (a widening moat, a vast under-penetrated TAM with a structural lock-in) that would place this company in the tail, or conclude the price is implausibly demanding. The discipline is to always translate a reverse DCF's output into a base-rate question ("how often does the thing this price requires actually happen?") rather than judging the growth number in a vacuum. That is the antidote to being seduced by a great narrative into paying for a tail as if it were the base case.

P10 (timed — 8 minutes; the bank reverse-DCF). Kaveri Bank: sustainable ROE 16%, sustainable growth g 10%, cost of equity r 13%, current P/B 2.6×. (a) Compute the justified P/B. (b) Compute the ROE the current price implies. (c) State the verdict.

Solution. (a) Justified P/B = (ROE − g) ÷ (r − g) = (0.16 − 0.10) ÷ (0.13 − 0.10) = 0.06 ÷ 0.03 = 2.0×. (b) Set 2.6 = (ROE − 0.10) ÷ 0.03 → ROE − 0.10 = 0.078 → implied ROE = 17.8%. (c) The market prices the bank at 2.6× against a fundamentals-justified 2.0×, i.e., it is pricing ROE expansion to ~17.8% vs the sustainable 16% (or a lower r, or higher g). Verdict: demanding. You are paying for ROE improvement the current book has not yet delivered; the variant question is whether NIM, credit cost, or operating leverage can lift ROE ~180bp durably. The lesson: for a bank the reverse DCF is the P/B-ROE identity, and it inverts just as cleanly as an FCF yield.

P11 (timed — 6 minutes; catch the landmine). Zenith Infra Ltd: revenue +28% YoY, EBITDA margin expanding, ROCE 19%: a great-looking growth story. But: promoter 61%, pledge only 4%; interest expense ₹340 crore on reported average debt ₹1,400 crore; "loans and advances to related parties" ₹2,100 crore (up from ₹700 crore); a subsidiary's ₹300 crore fixed deposit is "lien-marked against group facilities." Run the tripwire scan; state the verdict.

Solution. The low pledge (4%) is the distractor: the leak here is elsewhere. (1) Liability completeness: implied borrowing rate = 340 ÷ 1,400 = 24.3%, roughly double any plausible ~10–11% cost → debt is likely understated; implied true debt at 11% = 340 ÷ 0.11 ≈ ₹3,090 crore, i.e., ~₹1,700 crore of borrowings the balance sheet does not show (the CG Power pattern, M2.08). (2) Ecosystem leak: RPT advances tripling to ₹2,100 crore. (3) Asset serving group debt: the lien-marked ₹300 crore FD (the Zee pattern). Verdict: un-ownable pending resolution. The 28% growth and 19% ROCE are irrelevant beside three convergent governance/liability flags. The drill's lesson: pledge is not the only tripwire, because the related-party, liability-completeness, and lien signatures fire even when the pledge screen is clean.

P12 (timed — a full timed teardown target, with model answer; the cyclical trap). Cascade Materials (US, synthetic, aggregates & cement). Facts: revenue $6bn (near a cyclical high), trailing P/E 9×, EV/EBITDA 7×, EBITDA margin currently 26% vs a through-cycle ~19%, net debt/EBITDA 2.5× and rising, heavy sustaining capex, pricing strong on a construction boom. Produce the eight-line output.

Solution (model output). 1. Business: a heavy-materials producer (aggregates, cement) selling into construction: a local-monopoly-ish, freight-limited, deeply cyclical commodity. 2. Revenue engine: volume × price of a bulky low-value-to-weight product; pricing power is local (freight moats) but demand is cyclical. 3. Unit economics: high fixed costs → strong operating leverage both ways; margins peak with volumes and collapse in downturns. 4. Moat: efficient-scale/cost-advantage via local reserves and freight radius: real but cyclical; confirm in through-cycle, not peak, returns. 5. Economics verdict: good business at the wrong point to trust the numbers: 26% margin is peak vs ~19% mid-cycle. 6. Three things that matter: the construction cycle stage, mid-cycle (not trailing) EBITDA, and the net-debt trajectory. 7. Two things that could break it: a construction downturn hitting volumes into rising leverage (net debt/EBITDA 2.5× rising is the danger); a rate/housing shock. 8. Price verdict + question: NOT cheap despite 9× P/E. The earnings are at a peak, and the lowest P/E in a cyclical is often the peak, a sell signal; on mid-cycle EBITDA the real multiple is far higher, and EV/EBITDA must be read against replacement cost. The one question: where are we in the construction cycle and what is normalized mid-cycle EBITDA, because that, not the 9× trailing, is the price you are actually paying. (Score check: recognizes the cyclical trap, refuses trailing P/E, normalizes to mid-cycle, flags leverage into a downturn, ends in the cycle question. Clears the bar.)

P13 (guided — building a coverage universe). From week 70 you run one timed teardown a week; call it T = 50 teardowns a year (allowing two weeks off). An analysis goes stale, so to keep a name live you must re-tear it down on a fixed refresh cadence. (a) If a name must be refreshed every six months, how many names can you keep current? (b) Every quarter? (c) Once a year? (d) State the identity and the lesson.

Solution. The identity is maximum live universe = throughput ÷ refresh-frequency per name, U = T ÷ R. (a) Six-monthly refresh → R = 2/yr → U = 50 ÷ 2 = 25 names. (b) Quarterly → R = 4 → U = 50 ÷ 4 = 12.5 ≈ 12 names. (c) Annually → R = 1 → U = 50 names. (d) Currency and breadth trade off directly on fixed throughput: you can follow 12 names quarter-fresh or 50 names only annually-fresh, but you cannot follow 50 names quarter-fresh (that needs 200 teardowns/yr). Your true coverage universe is bounded not by how many companies you can read once, but by how many you can keep current; everything beyond U is a stale opinion wearing the costume of coverage.

P14 (guided — grade the rep, apply the gate). A timed teardown earns these 14 rubric scores (criteria 1–14, §8): [2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 0, 2, 2, 2]. The candidate says "26 out of 28 — comfortably above the 22 bar, so it passes." (a) Compute the total. (b) State every pass condition and check each. (c) Does it clear the bar? Name the trap.

Solution. (a) Total = thirteen 2s + one 0 = 26/28. (b) The teardown bar has three conditions, all required: (i) total ≥ 22 → 26 ≥ 22 ✓; (ii) no criterion scored 0 → criterion 11 = 0 ✗; (iii) the four hard gates (criteria 6, 9, 11, 13) each ≥ 1 → criterion 11 (reverse DCF) = 0 ✗. (c) It fails, on two of the three conditions, despite the high total. The trap is reading the bar as a single number ("≥ 22") and letting a strong total console you for a hole. The hard gates are independent of the total precisely so that a skipped reverse DCF, the one place edge lives (§4.4), cannot be bought back with points earned elsewhere. A 26/28 with a hard-gate zero is a fail; a 22/28 with all gates ≥ 1 is a pass. The rubric is not an average; it is a chain with four unbreakable links.

P15 (independent — score your calibration). Your decision journal, binned by stated confidence: at ~55%, 20 calls, 12 came true; at ~70%, 15 calls, 9 came true; at ~85%, 10 calls, 6 came true. (a) Compute the observed hit rate in each bin. (b) Compute the overall hit rate, your average stated confidence, and the overconfidence gap. (c) A second pathology hides here beyond overconfidence — name it.

Solution. (a) 12 ÷ 20 = 60.0%; 9 ÷ 15 = 60.0%; 6 ÷ 10 = 60.0%: flat at 60% across all three bins. (b) Overall = 27 ÷ 45 = 60.0%; average stated confidence = (20×0.55 + 15×0.70 + 10×0.85) ÷ 45 = 30 ÷ 45 = 66.7%; overconfidence gap = 66.7 − 60.0 = 6.7 points. (c) The flat hit rate is zero resolution (no discrimination): your "85% confident" calls come true at exactly the same rate as your "55% confident" ones, so your confidence ratings carry no information. They do not separate the calls you should trust from the ones you should not. Calibration (does 70% mean 70%?) and resolution (do your higher-confidence calls actually win more often?) are different virtues; a forecaster can be miscalibrated, un-discriminating, or both, and this record is both. The fix for resolution is not humility but sharper analysis on the calls you rate highly, earning the right to a spread in your confidences.

P16 (independent — a full timed teardown target, with model answer; the cheap side). Ironwood Industrial (US, synthetic; a maker of engineered industrial components). Ten-year panel: revenue $4bn (cyclical, roughly flat-to-up through the cycle), ROIC ~14% vs a ~9% WACC, modest net debt, FCF $540m, no governance flags (US; unqualified audit, aligned insider ownership). Market cap $6bn. Produce the eight-line output including the reverse DCF.

Solution (model output). 1. Business: an engineered-components maker selling into industrial OEMs and the MRO aftermarket: volume × price of specified parts, cyclical with customer capex. 2. Revenue engine: spec-locked components with aftermarket pull-through; the steadier profit engine is the installed-base MRO stream, not new-equipment sales. 3. Unit economics: high fixed-cost plants → real operating leverage both ways; incremental volume drops through at high margin. 4. Moat: narrow, resting on spec-in switching costs and engineering relationships; confirm in a through-cycle ~14% ROIC vs ~9% WACC (a real but modest ~5-point spread). 5. Economics verdict: decent, not wonderful: above its cost of capital across the cycle, sound balance sheet. 6. Three things that matter: the industrial capex cycle, the aftermarket mix, and input-cost pass-through. 7. Two things that could break it: a deep industrial recession into the (modest) leverage; a design-out by a large OEM. 8. Price verdict + question: generous. FCF yield 540 ÷ 6,000 = 9.0%; the Gordon inversion at r = 9% gives implied perpetual growth g ≈ 9% − 9.0% = 0%. The price implies zero growth forever for a business that out-earns its cost of capital, a demanding pessimism. The one question: is the market extrapolating the current cyclical trough as permanent, because if Ironwood merely grows with inflation the price is too low, the mirror image of the expensive-compounder trap. (Score check: reverse DCF run, moat confirmed in through-cycle ROIC, tripwire scan clean, ends in a variant question. Clears the bar, and shows "cheap" is a reverse-DCF verdict too: a low price is an implied story of decline, to be tested, not trusted.)

P17 (independent — the deep-dive promotion funnel and its constraint). You run T = 50 teardowns a year. A teardown earns promotion to a full deep dive when its variant question is genuinely alive; your deep-dive capacity is ~12 a year (one a month). (a) If one teardown in five surfaces a live variant, how many deep-dive candidates per year, and can you pursue them all? (b) If one in three does? (c) You also want to keep a 25-name universe six-monthly-fresh — what does that cost, and what does it imply for (a)/(b)?

Solution. (a) 50 ÷ 5 = 10 candidates/yr ≤ 12 capacity → you can deep-dive every live variant, with slack. (b) 50 ÷ 3 ≈ 16.7 candidates/yr > 12 capacity → you must select, pursuing roughly 12 ÷ 16.7 ≈ 72% of them and parking the rest on the watchlist, and the selection itself becomes a skill (rank by how mispriced × how knowable). (c) Keeping 25 names six-monthly-fresh costs 25 × 2 = 50 teardowns a year, your entire throughput. So the refresh universe and the new-idea funnel compete for the same 50 slots: fully maintaining 25 names leaves nothing for new names, and every name you add is a permanent future refresh liability. The binding constraint is always throughput; a coverage system is a budget, and pretending it is unlimited is how names go stale and live variant questions get missed.

P18 (timed — 12 minutes; the certification-day decision, all parts). You have finished certification. Compute and decide. (i) Teardown A included a US reverse DCF: market cap $30bn, FCF $750m, r = 9%. What is the implied perpetual growth? (ii) Its objective bank cross-check: P/B 2.4×, cost of equity 13%, sustainable growth 9%. What is the implied ROE? (iii) The four component scores: teardown A 24/28, teardown B 22/28, deep dive 27/32, mastery check 10/12. Express each as a percentage. (iv) Apply the compound gate: do you certify? (v) If not, state the minimal remediation.

Solution. (i) FCF yield = 750 ÷ 30,000 = 2.5%; g ≈ 9% − 2.5% = 6.5%, above ~4% US nominal GDP, so demanding. (ii) implied ROE = P/B × (r − g) + g = 2.4 × (0.13 − 0.09) + 0.09 = 0.096 + 0.09 = 18.6%. (iii) 24/28 = 85.7%; 22/28 = 78.6%; 27/32 = 84.4%; 10/12 = 83.3%. (iv) Check each gate: teardown A 24 ≥ 22 ✓; teardown B 22 ≥ 22 ✓; deep dive 27 ≥ 26 ✓; mastery 10/12 → fails (the bar is ≥ 11/12, i.e., ≥ ~85%). Three of four clear, but the gate is joined by AND → NOT certified. (v) The three rubric performances were on never-studied companies and each cleared its bar, so they stand. The only failed link is the mastery check, so the minimal remediation is to wait the mandatory two days and re-sit the mastery check on Form B (working-agreement 1); you do not re-do the teardowns or the deep dive. The stretch lesson: on certification day the arithmetic is trivial; the discipline is refusing to round 10/12 up to "basically passed," because the gate does not round — and neither does the market.

P19 (independent — one spoken rep, scored on the §4.9 rubric). Below is a synthetic transcript of a candidate's two-minute pitch on Lumen Diagnostics (synthetic), timed at 1 minute 52 seconds and delivered from memory. Score it out of 10 on the five criteria and name the single change that raises the score most.

"So, Lumen Diagnostics is an Indian diagnostics chain, about 1,100 collection centres, mostly south and west, and they've been around since the nineties. The industry is consolidating, organised players are taking share from standalone labs, and the demographics are obviously favourable. Margins have been under pressure because of price competition from the online players, but I think that's temporary. Management is good, the founder still runs it, and the balance sheet is clean, no debt. It trades at 38 times earnings, which is not cheap, but quality never is. I think the price competition eases as the loss-making entrants run out of funding, and then margins recover. I'd buy it."

Solution. Claim: 0 of 2. There is a recommendation, "I'd buy it," and it arrives in the last four words with no target, no horizon and no expected return. The first fifteen seconds went to founding date and geography. Variant perception: 1 of 2. There is one, and it is a real one: the market treats price competition as structural, the candidate treats it as funding-driven and temporary. It is stated as an opinion with no number, no funding runway estimate, no count of the loss-making entrants, so it earns half. Support: 1 of 2. Three supports exist by implication, being consolidation, the founder and the clean balance sheet, but only one figure appears in the whole pitch outside the multiple, and "management is good" is a claim rather than evidence. Falsifier: 0 of 2. None offered. The thesis is entirely reversible on a single observable, namely whether the discounting entrants keep raising money, and that observable is never named as a disconfirmer. Delivery: 2 of 2. Inside time, from memory, fluent. Total 4 of 10, against a working bar of 8.

The single highest-value change is the falsifier. It is worth two full points and costs one sentence: "I am wrong if the two funded entrants raise another round at scale, or if realisation per test is still falling four quarters from now; either one and I am out." Note what that sentence does beyond the score. It converts "I think that's temporary" from a hope into a dated, checkable claim, which is the same conversion the eight-line output makes in writing and the same one E11.06's pre-registered sell criteria make in a portfolio. Second-best change is the claim, which is free: move "buy, target ₹X over eighteen months, about Y% annualised" to the front and the pitch loses nothing.

Applied mini-project: THE CERTIFICATION (the M9 milestone)

This is not a mini-project. It is the milestone the entire program was designed to deliver, the certification of competency C10, the signature skill, and the crown gate of Phase 8. Budget the bulk of the module's twenty hours here. It has two parts and a standard, and it is done on companies you have never studied before.

Part 1 — Two timed 60-minute teardowns. One Indian company and one US company, each never torn down before, each on a fresh 60-minute clock, each producing the full twelve-question read and eight-line output, each self-graded against the teardown rubric below. Do them on separate days. Assemble the source stack first (that time is not part of the 60 minutes; the clock starts when you begin question 1, exactly as the app's timed mode runs it).

Part 2 — One full independent deep dive. One never-studied company (either market, ideally the market or sector you are weaker in), run over two to three days at the independent rung (blank workbook; you drive), producing the full deliverable set: ten-year history + incremental ROIC, industry map, three-statement-lite (base/bull/bear), reverse DCF, scuttlebutt, incentive read, earnings-call read, steelman, one-page thesis, and pre-mortem. Self-grade it against the deep-dive rubric below.

Choosing never-studied companies (the integrity rule). Draw from your held-out list (§4.7, the names you deliberately never practised on) or pull a fresh company off a screener you have not opened. Confirm honestly that you have not previously torn it down and have not read a bull/bear note on it. Self-certification's entire value is its honesty (Common mistake 8): a company you secretly half-know certifies nothing. For Part 1, one India + one US is mandatory (the program's both-markets rule, working-agreement 5). Before starting any of the three, spend ten minutes confirming no material development (auditor change, regulatory order, restatement) has intervened. If one has, that company just became a better certification target, not a spoiled one.

The teardown rubric (printable; score each of the two timed teardowns)

Score every criterion 0, 1, or 2 against the anchor. Maximum 28.

Teardown pass bar: ≥ 22/28, no criterion scored 0, and criteria 6, 9, 11, and 13 each ≥ 1. Those are the four things the rubric refuses to let you skip (moat-in-numbers, the tripwire scan, the reverse DCF, and the output-with-variant-question). Both teardowns must clear the bar.

The deep-dive rubric (printable; score the one deep dive)

Score each criterion 0, 1, or 2. Maximum 32.

Deep-dive pass bar: ≥ 26/32, no criterion scored 0, and criteria 2, 7, 11, 12, and 13 each ≥ 1: incremental ROIC, the reverse DCF, the steelman, the one-page thesis, and the pre-mortem, the five that separate a thesis from a book report.

The certification standard, remediation, and what it unlocks

To certify (all four, together): both timed teardowns ≥ 22/28 (one India, one US, never-studied) AND the deep dive ≥ 26/32 (never-studied) AND the Mastery check (§11) passed at ≥85%. Meeting all four certifies C10, the signature skill, records the M9 milestone (SIGNATURE SKILL CERTIFIED), and unlocks Phase 9 (M9.01, Risk & Position Sizing), where you begin running money with the skill you have just proven.

Remediation (the gates do not bend; the calendar does). A teardown below bar is diagnosable: read which criteria cost the points, do three targeted reps on exactly those (§4.4's fixes), wait the mandatory two days (working-agreement 1), and re-sit on a fresh never-studied company (teardowns must always be novel; never re-sit the same name). A deep dive below bar: identify the weak criteria; if the failure is in specific sections (the reverse DCF, the steelman, the pre-mortem), rework those sections on the same company first, because rework builds the skill and a new target only rebuilds the hope; if the whole thing is structurally thin, drop back to the faded deep-dive rung (workbook prompts on) for one rep before re-attempting independent. Two failures of the same type → you removed a scaffold too early; go back one rung and rebuild it (§4.3). None of this is a verdict on your ability; it is a map of what to drill, which is the entire ethos of the module.

A word on what you are certifying. Not that you are a finished analyst. You are certifying that the skill is real: that handed a company you have never seen, on a clock, you can produce a defensible, decision-shaped read, and that given three days you can produce an institutional-grade thesis with a falsifiable variant view and an honest pre-mortem. That is the capability the whole program was backward-designed to build, and clearing this gate is the moment it becomes demonstrably, portably yours, with or without these materials, with or without any AI, for the rest of your working life. Phase 10's two capstones will make you do it twice more at higher stakes; the years after will make you do it hundreds of times against reality. This week you prove the engine runs.

Case Lab: run case CL-8.2 (Sixty Minutes on Trident Components, a synthetic, never-studied company served cold with its full source stack) in the app's Cases tab for this module. It is the certification instrument. The 60-minute clock starts when you open question 1; you produce the twelve-question read and eight-line output from memory; and the app auto-grades the objective parts (profit-engine ID, the ROIC-vs-hurdle call, the reverse-multiple arithmetic to a tolerance band) and rubric-scores the output against the model answer on the total, the no-zero rule, and the four hard gates (§8). It then logs the miss category of every criterion that scored below 2 into your taxonomy.

Reading & resources

• The two program playbooks: artifacts/playbook-1-2-hour-teardown.md and artifacts/playbook-2-3-day-deep-dive.md. The reusable templates you fade away from in this module and keep on your desk forever. Print the teardown sheet (with the §4.5 time budget on the back) and the deep-dive workbook now. [Free] [program artifact]
• Anders Ericsson & Robert Pool, Peak: Secrets from the New Science of Expertise. The science behind §4.2: deliberate practice, why comfortable reps plateau, the role of immediate feedback and error analysis. The single most useful non-finance book for this module. [Paid] [Beginner]
• Alfred Rappaport & Michael Mauboussin, Expectations Investing (revised ed.). The reverse-DCF / expectations-investing spine that questions 11–12 of the teardown and criterion 7 of the deep dive rest on; the Phase 8 reading-spine text. Read Part I now if you have not. [Paid] [Intermediate]
• Aswath Damodaran, Narrative and Numbers. The discipline the deep-dive rubric's criterion 15 certifies: every qualitative claim turned into a number, and every number tied to a story. Pair with his free NYU valuation spreadsheets for the three-statement-lite and reverse DCF. [Paid] [Intermediate]
• Philip Fisher, Common Stocks and Uncommon Profits. The scuttlebutt method and the fifteen points behind deep-dive criterion 8; the Phase 8 spine text alongside Expectations Investing. [Paid] [Intermediate]
• Michael Mauboussin (Counterpoint Global / Morgan Stanley), Measuring the Moat and The Base Rate Book. Free reports: the moat-in-the-ROIC-series logic (teardown criterion 6, deep-dive criterion 4) and the base-rate discipline behind the reverse-DCF plausibility check (practice problem P9). [Free] [Advanced]
• Atul Gawande, The Checklist Manifesto. Why a fixed, boring checklist (the tripwire scan) run every single time beats expert memory under pressure, which is the theory behind §4.4's mechanical fixes. [Paid] [Beginner]
• Daniel Kahneman, Thinking, Fast and Slow. System 1/System 2, anchoring, and the automaticity thesis (§4.1) and anchoring failure (§4.4) in their original form; revisited from Phase 6. [Paid] [Intermediate]
• Annie Duke, Thinking in Bets (and Philip Tetlock, Superforecasting). Process-over-outcome and calibration: the feedback horizons of §4.2 and the lifelong calibration cadence of §4.8; the bridge into Phase 9. [Paid] [Intermediate]
• Primary sources for the certification (all free, and your syllabus for life): SEC EDGAR (10-K Item 1/1A, DEF 14A, transcripts) for US names; BSE/NSE company pages, screener.in, Tijori for Indian annual reports, shareholding patterns (pledge), and ten-year panels; company investor-relations pages for decks and call transcripts. Every certification and every weekly rep is pulled from here: no secondary summary, and no AI, required. [Free] [Beginner–Advanced]
• Business Breakdowns and Acquired (podcasts). Hear full teardowns and deep dives done aloud by professionals; use them for calibration: pause, form your own read, then compare. [Free] [Beginner]
• Going deeper: Good Judgment Open and its calibration/Brier-score primers (goodjudgmentopen.com; the Good Judgment Project's public forecasting platform) [Free] [Intermediate]. Practise probabilistic forecasting on real questions with scored Brier feedback; the live, external version of WE5's calibration loop and the fastest way to shrink your own overconfidence gap. The theory is in Tetlock's Superforecasting (listed above); this is the gym.
• Going deeper: James Valentine, Best Practices for Equity Research Analysts (McGraw-Hill) [Paid] [Intermediate]. The professional's playbook for building and maintaining a coverage universe (the U = throughput ÷ refresh discipline of the practice set in real life), earnings models, channel checks, and idea generation; the sell-side counterpart to this module's independent process, useful precisely because it treats coverage as a throughput budget, not an unlimited list.

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": "The deliberate-practice loop and the miss taxonomy, the scaffold-fading ladder, the five timed failure modes and their mechanical fixes, the sixty-minute budget and its sector variants, the two rubrics as practice instruments, and the four-component certification",
  "use": "This module has a structural problem no other module in the program has, and it has to be named before any tool is discussed: section 4.3's whole method is REMOVING supports one at a time, and a tool is a support that was never on the ladder. Introducing one at rung 3 is not climbing; it is putting the scaffold back under a different name, and the top rung — sixty minutes, a company you have never seen, nothing but the source stack and the clock — is the thing being certified. So the legitimate uses all sit OUTSIDE the rep, never inside it. (1) TARGET SELECTION, which is deliberate practice's first ingredient: 'a rep at the edge of your ability' means a bank, a cyclical at peak, a hyper-growth money-loser, an unfamiliar market — and finding five never-studied companies that fit a specified shape is a screening job that costs nothing analytically and removes the excuse of not knowing what to practise on. (2) SOURCE-STACK ASSEMBLY BEFORE THE CLOCK STARTS, so your sixty minutes are spent on the twelve questions rather than on downloads — this is M8.01's fetch, applied so the timed rep tests the skill rather than your broadband. (3) THE REP LOG AS DATA: section 4.2 asks you to tally every miss by category and re-read the tally, and a tally of forty reps is a table you query, not a notebook you leaf through. (4) A SECOND READ AGAINST THE PRINTED ANCHORS, after you have self-graded — the rubric anchors have printed definitions, so a disagreement is a prompt to re-read the anchor, never a grade that overrides yours. THE BANS: nothing inside a timed rep, ever; no model answer generated on demand before you have produced your own output, because section 4.2's rule 2 is write-before-you-peek and a generated 'here's how I'd approach it' is peeking that does not feel like peeking; and no confidence number for the decision journal, which M6.05 already bans for the same reason — it is the exact judgment the instrument exists to measure.",
  "tools": ["EDGAR full-text search and the data.sec.gov frames endpoint; screener.in and Trendlyne screens for India — used to BUILD the never-seen target bank of section 4.7, not to analyse anything in it", "The XBRL endpoints and aggregator exports — source-stack assembly before the clock starts, so the timed rep tests the method", "A spreadsheet or notebook holding the rep log: date, company, market, sector variant, rung, minutes, the per-criterion rubric scores, and the miss category for every criterion below 2", "Document/PDF AI with retrieval over YOUR OWN past teardowns and theses — the one corpus where retrieval is unambiguously safe, because you wrote it and you are looking for your own pattern", "The decision journal itself, in whatever form you will actually keep — the raw data the calibration work in M0.04 and M6.05 runs on"],
  "prompt": "Attached is my rep log: forty timed teardowns, each with a date, company, market, sector variant, finish time, the fourteen per-criterion rubric scores, and the miss category for every criterion I scored below 2. Produce three tables and nothing else. (a) Miss count by category, and by category crossed with sector variant. (b) For each of the fourteen criteria, my mean score over the first twenty reps and over the last twenty. (c) Finish time by rep number. Do not tell me what the pattern means, do not suggest what to practise, and do not congratulate me on anything. I read my own signature.",
  "verify": "Verify the instrument before you verify yourself. The rep log is self-scored, so it measures your analysis through an instrument you also control, and section 4.6 is blunt that the honesty IS the entire value — a generous self-grade is a loan taken from your future self and repaid at certification with interest. So audit the scoring the way you would audit a filing: re-score two old reps cold against the printed anchors and see whether today's you agrees with that day's you, and treat systematic drift as a finding about the instrument rather than about the skill. Then verify the arithmetic of what a record can and cannot support. This module's Worked example 4 must reproduce 25/28 = 89.3%, 23/28 = 82.1%, 28/32 = 87.5% and 11/12 = 91.7% on a four-way AND gate, and Worked example 5 must reproduce a 63.3% hit rate against 73.0% average stated confidence, a 9.7-point overconfidence gap and a Brier score of 0.261 against the coin-flipper's 0.25. The Quant lens below adds the number those examples imply and do not state: how long a run of reps has to be before a claim about your clearance rate is distinguishable from luck.",
  "diy": "The gate is the whole module: two timed sixty-minute teardowns on never-seen companies in two markets, one independent deep dive, and a mastery check — unaided, on the clock, with no tool in the room. That is the certification, and its entire value is that nothing in this callout was present when you earned it."
}

Modern Data Analysis

By hand first. You ran the rep loop, tallied misses by category, walked the scaffold-fading ladder, drilled the five timed failure modes, scored a full certification scoreboard through a four-way AND gate (WE4), and computed your own hit rate, overconfidence gap and Brier score from a synthetic journal (WE5). Every one of those must stay a hand skill, and one of them, the timed rep itself, must stay a toolless skill, because the certification is defined as the room with nothing in it.

Today's workflow. The modern layer belongs on either side of the rep and never inside it. Before: §4.2's first ingredient is a target at the edge of your ability, and the honest obstacle is usually not willingness but supply. You practise on the companies you can think of, which are the ones you already understand. A screen fixes that in a minute: the frames endpoint returns one concept across every US filer for a period, so "profitable, never-studied, capital-intensive, non-US-listed peer of a name I know" is a filter rather than a memory, and screener.in's screens do the same for India. Then assemble the source stack before the clock starts, so the sixty minutes test the twelve questions and not your download speed. After: the rep log stops being a notebook and becomes a table.

The rep log as data, which is the actual upgrade here. §4.2 asks for a running tally by miss category and says a personal signature emerges over twenty reps that is "almost never what you would have guessed." That is a claim about a dataset, so store it as one: one row per rep with the date, company, market, sector variant, rung, finish time, the fourteen per-criterion scores, and the miss category for every criterion below 2. Four queries then answer questions no amount of leafing through a notebook will: miss category crossed with sector variant (is your coverage problem general, or does it only appear on lenders?); mean score per criterion, first twenty reps versus last twenty (which criteria are actually improving, as against which ones you feel better about); finish time against rep number (whether speed is coming, and from which block); and criteria that co-occur below 2 (a moat miss and a price miss in the same rep is usually one failure, not two). The last one is the miss taxonomy's own logic applied to itself, and it is the sort of thing you can only see once forty reps are rows.

Retrieval over your own past teardowns is the one corpus where the pattern carries no risk at all: you wrote it, the locators are your own files, and you are looking for your own repetitions rather than for facts. "Show me every variant question I have written in the last thirty reps" is a devastating exercise the first time you run it, because most people discover they have been writing the same three questions about very different businesses.

Tools & sources (IN + US). US: data.sec.gov/api/xbrl/frames/us-gaap/<Tag>/USD/<period>.json to build a never-seen target bank by shape rather than by memory; companyfacts and submissions for pre-clock source-stack assembly; EDGAR full-text search to find companies whose filings contain a structure you want to practise on (a going-concern paragraph, a first-time segment re-cut). India: screener.in's screens and its ten-year sheet and concalls tab, Trendlyne and Tijori, for the same target-bank job on the market where the filings are PDFs and the aggregator is doing the structuring work an endpoint does elsewhere; the BSE and NSE filing archives for the raw documents. Both: your own rep log and journal, which is the only dataset in this module that is genuinely about you, and the only one nobody else can build. The calibration side of that journal is not re-derived here: M0.04 §4.4 and M6.05 already computed the sample-size arithmetic for what a confidence record can support, including why a gross miscalibration is cheap to detect and a ten-point one is not, and duplicating it here would be worse than pointing at it. Run any code in your notebook.

# R3 — recomputed in-session. What a certification proves, and what a run of reps proves.
from scipy.stats import binom, beta

# 1. The four components are joined by AND, not averaged (WE4's whole point, as a probability).
for p in (0.80, 0.85, 0.90, 0.95):
    print(f"  per-component pass probability {p:.0%} -> P(certify) = {p**4:.1%}")
    #        80% -> 41.0%      85% -> 52.2%      90% -> 65.6%      95% -> 81.5%
for target in (0.50, 0.80, 0.90):
    print(f"  reliability needed per component for P(certify) = {target:.0%}: {target**0.25:.1%}")
    #        50% -> 84.1%          80% -> 94.6%          90% -> 97.4%

# 2. What ONE certification tells you about your per-component reliability (one-sided 95% floor).
print(f"  passing 4 of 4 components bounds per-component reliability at >= {beta.ppf(.05,4,1):.1%}")  # 47.3%

# 3. How long a rep record must be before a clearance claim beats luck (exact one-sided binomial).
def smallest_n(p0, p1, target=0.80, alpha=0.05):
    n = 2
    while True:
        k = -1
        for kk in range(n + 1):                       # reject 'true rate = p0' when MISSES <= k
            if binom.cdf(kk, n, 1 - p0) <= alpha: k = kk
            else: break
        if k >= 0 and binom.cdf(k, n, 1 - p1) >= target: return n
        n += 1
for p0, p1 in [(.50, .80), (.65, .85), (.70, .85), (.75, .85)]:
    n = smallest_n(p0, p1)
    print(f"  show a true clearance rate of {p1:.0%} against {p0:.0%}: n = {n} reps "
          f"(~{n/50:.1f} yr at the module's ~50/yr cadence)")      # 18 / 31 / 49 / 103

# 4. A short perfect streak, read honestly (Clopper-Pearson one-sided 95% floor).
for k, n in [(3, 3), (5, 5), (8, 10), (17, 20)]:
    print(f"  {k}/{n} cleared -> true rate is at least {beta.ppf(.05, k, n-k+1):.1%}")
    #  3/3 -> 36.8% | 5/5 -> 54.9% | 8/10 -> 49.3% | 17/20 -> 65.6%

Verify. Prove any scoring code against this module's own worked figures: WE4's four components must give 89.3% / 82.1% / 87.5% / 91.7% and a four-way AND that clears; WE5 must give bin hit rates of 66.7 / 60.0 / 62.5%, an overall 63.3% against an average stated confidence of 73.0%, an overconfidence gap of 9.7 points, bin contributions of 2.72 / 2.625 / 2.48, a total of 7.825 and a Brier score of 0.261 against the coin-flipper's 0.25; and the fast Gordon inversions must give a 2.0% FCF yield with 10.0% approximate and 9.8% exact implied growth for Aarogya, and 2.0%, 7.0% and 6.9% for Sentinel. All reproduce exactly. Then verify the thing the arithmetic cannot: that the instrument is honest. Every figure in a rep log is self-scored, so it measures your work through a gauge you also control. Re-score two old reps cold against the printed anchors and compare to that day's grade; systematic drift is a finding about the gauge, not about the skill.

Quantitative lens

WE4 says a certification is a chain rather than an average, and WE5 says a win rate is not calibration. Both are right and both stop one step short of the number that makes them actionable. Everything below comes from the block above, on the module's own gate structure and stated inputs; they are properties of that arithmetic, not measurements of anyone's ability.

A four-way AND gate punishes reliability far harder than the average suggests. If you clear each of the four components with probability 85% (a candidate who is genuinely good and occasionally has a bad hour), the probability of certifying on the day is 0.85⁴ = 52.2%. At 80% per component it is 41.0%; at 90%, 65.6%; only at 95% does it reach 81.5%. Invert it and the planning number is stark: to have an 80% chance of certifying you need to be 94.6% reliable on every component, and to have even a coin's chance you need 84.1%. That is the arithmetic behind WE4's warning, and it is sharper than the warning. WE4's three score claims all reproduce: its own scoreboard averages 87.65%; a scoreboard of 23/28, 23/28, 30/32 and 11/12 averages 87.42%, indistinguishable to the eye while every component still clears; and dropping one teardown to 21/28 fails the certification outright at a four-way average of 85.64%, which still looks fine. The probability version says something those three cannot: being marginally above each bar is not a passing position, it is a coin flip, and the only preparation that changes the odds is being comfortably clear on your weakest component. Practising the thing you are already good at does not move p⁴ at all.

And a single certification is a gate, not a measurement. Pass all four components once and the one-sided 95% floor on your per-component reliability is 47.3%. Passing proves you are probably better than a coin at each component and essentially nothing more. That is not a criticism of the certification; a driver's licence has the same property, and §4.8 already calls it one. It is the argument for §4.8's periodic self-recertification, stated as arithmetic: one observation of a four-part event cannot separate a 60%-reliable analyst from a 95%-reliable one, so if you want to know which you are, the only instrument is the ongoing rep record.

Which raises the question the whole module implies and never answers: how many reps before a claim about your skill is distinguishable from luck? The honest answers are humbling at the short end and encouraging at the long one. A short perfect streak proves very little: 3 of 3 cleared bounds your true clearance rate at only 36.8%, 5 of 5 at 54.9%, and 8 of 10 at 49.3%. So the learner who clears three timed teardowns in a row and concludes the skill is theirs has, statistically, learned that they are probably not terrible. Seventeen of twenty gets the floor to 65.6%, which is the first number that would justify any confidence at all. Testing rather than bounding gives the same shape: demonstrating a true clearance rate of 80% against a 50% null takes 18 reps; 85% against 65% takes 31; 85% against 70% takes 49; and 85% against 75% takes 103. Set those against §4.8's own cadence of roughly fifty timed teardowns a year and they become a calendar. Eighteen reps is four months. Thirty-one is seven. Forty-nine is a year. A hundred and three is two years, and that last one is the honest reason §4.8 says the difference between the analyst you are at week 80 and at year four is "roughly 150 more teardowns": fine distinctions in skill need years of reps to become visible, and coarse ones do not.

The practical reading is the same as M0.04's on calibration and it is genuinely liberating rather than discouraging. Your first ten reps can tell you whether you are grossly below the bar, which is the most expensive thing to be wrong about and worth catching immediately. They can tell you nothing finer, so do not try to read a trend into them. And the pattern of your misses, the §4.2 signature, is informative long before the rate is, because a category tally with twelve coverage misses and one calibration miss is a clear instruction from a sample far too small to support any statement about your clearance rate. Read the signature early; read the rate late.

Honest limits, and here they are unusually severe because the model is unusually simple. Reps are not independent draws from a fixed rate, and the violation is the entire point of the module: you improve across the sample, so a single pooled binomial test is the wrong final instrument. The right one segments by period, and segmenting costs the sample size you did not have. Companies differ in difficulty and you choose your own targets, so a clearance rate is a statement about your selection as much as your skill, and §4.2's instruction to practise at the edge of your ability deliberately drives it down. The rubric is self-scored, so the outcome variable is measured with an instrument the subject controls. The 85%-versus-65% framing picks two numbers to make the arithmetic concrete; nothing in this module establishes what a "true clearance rate" of a real analyst is, and I have no panel to establish it from. And the deepest limit is that clearing a rubric bar is not the same as being right about a company: the rubric measures whether you ran the process, which is what §4.2 correctly calls process feedback, and the slow, noisy outcome feedback that decides whether the process was any good takes years and is measured by the calibration machinery in M0.04 and M6.05, not by anything computed here.

Do it in code: keep the rep log as rows from your very first timed rep, tally the miss category before you look at the total, query the signature at twenty reps and the rate at fifty, re-score two old reps cold every quarter to audit your own gauge, and before certification week compute p⁴ at an honest estimate of your weakest component rather than at your average. See G1/QM1.02, G1/QM1.01 and G2/QD2.04.

Where this goes next: galaxy cross-links

• G1 · QM1.02 Hypothesis Testing, Correlation & Regression: the machinery behind the rep-count arithmetic: what a one-sided test is, what power means, why an exact binomial and a normal-approximation planning formula give different answers to "how many," and why a small sample can be honest and uninformative at once.
• G1 · QM1.01 Probability, Distributions, Sampling & Estimation: confidence bounds on a proportion, why 3-of-3 supports almost nothing, and why a compound AND gate collapses so much faster than an average suggests.
• G2 · QD2.04 Reproducible Research, Deployment & the Quant Stack: turning the rep log and the calibration query from a spreadsheet into a versioned pipeline that re-scores itself as the record grows, which is what a multi-decade practice record actually needs.
• G2 · CS2.06 Software Engineering: Testing, Clean Code & AI-Assisted Development: the rubric as an executable check rather than a printed page, and the discipline of a test suite that fails loudly, which is the same discipline as scoring each criterion against its anchor rather than against your impression of the rep.
• G5 · AI0.07 The AI-Augmented Analyst — Applying AI Across Your Work: the general form of this module's hardest line: where a tool belongs in a practice loop whose entire purpose is to make you independent of tools, and why the answer is "on either side of the rep, never inside it."

Teach it back & journal

Feynman prompt. Write a one-page explainer for a smart 15-year-old titled "How to size up any company in an hour." Teach them the five questions in plain words (what does it do and how does it make money; is it a good business; are the people honest and good with money; what could kill it; and is the price already assuming everything goes right), the idea that you spend your hour on the two or three things that decide the answer and let the rest run on autopilot, and one plain-English version of the reverse-DCF idea: "the price is a promise about the future; work out what future the price is promising, then ask if that promise is likely." End with the eight lines you would hand them. Rule: no jargon. If you cannot explain "what the price is implying" without saying "discount rate," you do not yet own it well enough to do it in sixty minutes.

Journal reflection — the certification calibration entry. This is the most important journal entry of the program so far; write it the day you finish the certification and date it. Record: (1) your three certification scores for teardown A (India), teardown B (US), and the deep dive, criterion by criterion, and the single lowest-scoring criterion across all three; (2) your miss-taxonomy signature across this module's reps, meaning which one or two of the six categories you miss repeatedly, and what that reveals about your bias (do you over-trust growth stories? under-scan governance? need 80 minutes for a 60-minute job?); (3) the one failure mode from §4.4 most likely to cost you real money under pressure, and the exact structural fix you commit to running (e.g., "tripwire scan is minute 0–3, every company, no exceptions"); and (4) your lifelong cadence commitment: the weekly teardown, the roughly-monthly deep dive, and the day of the week you protect for them. Phase 9 opens by asking you to size real positions; you will re-read this entry before you do, because the analyst who knows their own error signature sizes very differently from the one who thinks they have none.

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