The Process + Capstone

Phase 9 · Week 74 · ~20 hours

Where it sits. A genuinely good manager, one running at an information ratio of 0.50, needs 15.4 years of track record before the arithmetic can call the skill real at conventional confidence, and 31.4 years before the test can reliably find it at all. Most careers are shorter than that. So the useful question is not how you prove you are good. It is how you get good at something whose scoreboard will not report back in time, and the answer is that you build a machine and then grade the machine.

The final module of Phase 9 and its capstone is the point where the entire program stops being a set of skills you have and becomes one machine you run. Everything upstream was a component. Phases 1–2 taught you to read and trust the numbers; Phase 3 to value; Phases 4–5 to judge the business and its sector; Phase 6 to hold a philosophy, reason under uncertainty, and build a checklist, a journal, and a temperament; Phase 7 to read the weather; Phase 8 compressed all of it into a repeatable teardown and deep dive; and the three modules just before this one gave you the last components: M9.01 (how much to bet), M9.02 (how the bets fit together into a portfolio), M9.03 (when to sell and how to learn from mistakes). What none of those did on its own is assemble them into a single, closed, repeatable loop: sourcing → analysis → checklist → sizing → portfolio fit → decision-and-journal → monitoring → sell → review-and-calibration → back to sourcing. That is the wheel you will turn, quarter after quarter, for the next forty years. Its spine is the decision journal; its engine of improvement is calibration; its non-negotiable constraint is survival. Its output, the M10 milestone, is two things you will carry for life: your complete written investment process document (the operating manual you run) and one full simulator season of at least ten journaled, calibrated decisions, scored on process, not P&L. A talented analyst becomes a repeatable one here, because in investing the person who runs a sound process for decades beats the person who has a brilliant year.

Suggested hour budget. Core content (the loop; the journal spine; quarterly letters; calibration; skill-vs-luck; survival; drawdowns; the process document): 7h · worked examples (a full journal entry, a calibration score, the process scorecard, a review letter): 3h · practice set (calibration scoring, process scoring, survival, drawdown protocol): 4h · the capstone (write the process document + run the simulator season): 4h · flashcards + mastery check: 2h.

App link. This module is where competency C9 ("investor's temperament and process") is certified, and it wires directly into the machinery you have been feeding since week 3: the portfolio simulator (a full accelerated season through a seeded market cycle, every trade requiring a journaled thesis and pre-committed kill criteria), the decision journal (the same entity, now the expanded investment-decision schema of §4.2), the calibration dashboard (predicted-vs-realized, the reliability curve, the Brier score; it has been quietly collecting your journaled confidences since M6.05, and you will learn to read it here), and the process scorecard (§4.5: the app scores each resolved decision on process quality independent of its P&L, which is what the leaderboards rank). The gate is the M10 milestone: the written process document plus a simulator season run on-thesis. The app grades the season on process consistency and calibration, never on the money, exactly as working agreement #8 demands.

Learning objectives

By the end you can:

1. Assemble the full investment loop: name its nine stages in order, state which module built each and what each hands to the next, and explain why running them as a closed, feeding loop (rather than as isolated skills) is what makes a process compound over decades.
2. Operate the decision journal as the spine of the process: write the expanded investment-decision entry (thesis, disconfirming evidence, base rate, numeric confidence, kill criteria, sizing and portfolio-fit rationale, monitoring metrics, review date) before the outcome, never edit it after, and explain precisely why pre-commitment defeats hindsight bias.
3. Write a quarterly review letter: produce the format (positions, decisions, mistakes, calibration, what changed my mind), explain every holding and every buy/sell as if to a demanding investor, and articulate why the act of writing forces a clarity that thinking alone does not (the Buffett/Marks practice).
4. Compute and read your calibration: bucket your probabilistic predictions, plot the reliability curve, diagnose whether and where you are over- or under-confident (by confidence level and by domain), compute a Brier score, and convert the diagnosis into specific corrections (haircut confidence, widen ranges, downsize, avoid a domain).
5. Separate skill from luck with a process scorecard: score a decision on its process on a fixed rubric independent of its outcome, place any resolved decision in the decision-quality × outcome-quality grid, explain "resulting" and the paradox of skill, and run two scoreboards (process and P&L) knowing which one to optimize.
6. State and enforce survival-first discipline: explain why avoiding ruin is the precondition for compounding (ergodicity), enumerate the forms of forced liquidation (margin, ALM mismatch, redemptions, derivatives, illiquidity-plus-a-cash-need), and specify the leverage and position-limit rules that guarantee no market move can make you sell at a time not of your choosing.
7. Run a drawdown behaviorally, by protocol: walk the exact steps your process prescribes when a position or the whole portfolio is deep in a drawdown, so that a calm, pre-committed past self makes the hold/add/sell decision rather than a panicked present one.
8. Produce your complete written investment process document: the operating manual covering philosophy, circle, sourcing criteria, analysis workflow, checklist, sizing and portfolio rules, sell rules, review cadence, and anti-goals. Then run one full simulator season scored on process, completing the M10 milestone.

Prerequisites & connections

Builds on. This module is an assembly, so it depends on nearly everything, but four modules are load-bearing. M9.01 (risk and position sizing): fractional Kelly, conviction-versus-downside, hidden correlation, and the ruin constraint are the sizing and survival stages of the loop, and they become standing rules here. M9.02 (portfolio construction): hurdles and opportunity cost, cash as a position, the concentration policy, and correlation are the portfolio-fit stage. M9.03 (selling and mistakes): the five legitimate sell reasons, the error of selling compounders early (Freeman-Shor), and your mistake taxonomy are the sell and part of the review stages. M8.01–M8.05: the timed teardown and the 2–3 day deep dive are the analysis stage; the one-page thesis you write there is the raw material of the journal entry here. Beneath those: M6.05 installed the checklist, the decision journal, and calibration/Brier as concepts and produced your Checklist v1, and those components now get wired into a running loop while the calibration data M6.05 began collecting finally gets graded; M6.03 supplies the decision science the whole loop enforces (process over outcome, expected value, base rates, Bayesian updating, ergodicity/ruin); M0.04 installed the original eight-field journal template, which §4.2 expands into the full investment-decision schema; and M3.05 (the reverse DCF, "what is priced in?") is the valuation test every journal entry must clear.

Feeds forward. Phase 10 is nothing but this loop, run for real and unaided: M10.01 and M10.02 are two full independent deep dives on never-studied companies (US, then India) that execute this exact process end to end and produce institutional-grade theses; M10.04 sets up the lifelong regimen (a public track record, a paper portfolio, the staying-current stack), which is this loop turning forever, and runs the final competency audit against Section 1 of the master map, where C9 is certified largely here. Your written process document becomes the standing operating manual for all of Phase 10 and the decades after it; your calibration record and process scorecard become the objective, luck-proof evidence that your judgment is actually improving, which over short horizons is the only kind there is.

The one-sentence version of this module. Take every skill the program built, wire it into a single closed loop whose spine is a decision journal written before outcomes and never edited after, whose engine of improvement is calibration (are your 70%s true 70% of the time?), whose scoreboard is process quality rather than P&L (a good decision can lose, a bad one can win), and whose non-negotiable floor is survival (no leverage or structure that can force you to sell at the bottom) — then write that loop down as an operating manual and run it, quarter after quarter, because the compounding of a sound process is the only edge no one can take from you.

Core teaching content

4.1 From components to a machine: the loop

You have, at this point, a workshop full of excellent tools. You can spread a filing by hand, run a quality-of-earnings screen, build a DCF and reverse it, grade a moat four ways, run a sector playbook, read the macro weather, size a bet with fractional Kelly, and construct a portfolio. A workshop full of tools is not a craftsman. The craftsman is the person who has a repeatable sequence: who knows what to do first, what feeds what, where to pause, and how each finished piece makes the next one better. That sequence is the investment process.

Here is the whole loop. Read it as a cycle, not a list: the last stage feeds the first, and every turn of the wheel leaves the wheel itself a little better.

Three things about that cycle carry the weight.

First, it is a loop, and the loop is where the compounding lives. A beginner runs stages 1 through 8 and stops: buys, holds, hopes, and moves on. The professional treats stage 9 as the most important stage, because stage 9 is what makes the next turn of the loop better than the last: every mistake becomes a permanent checklist line (stage 3 improves), every resolved prediction tightens your calibration (stage 6's confidence numbers get more honest), every deep dive widens your circle (stage 1's funnel gets richer and more variant), and every quarterly letter forces the clarity that surfaces the next idea. Run this loop for ten years and you are not ten years older doing the same thing. You are running a materially better process than you started with, built from your own logged experience. That is the flywheel, and it is why an analyst's real edge is cumulative and nearly impossible to copy: it is compounded from a private record only you possess.

Second, the stages are coupled: each is a gate the idea must pass, not an independent activity. The output of analysis (a thesis and a range) is the input to the checklist; the checklist's vetted thesis is the input to sizing; sizing's downside floor is an input to portfolio fit; and so on. This coupling is why a workshop of tools is not a process: the tools have to be chained in the right order with the right hand-offs, or the idea leaks value between stages. The commonest failure is not a weak tool, because your tools are strong by now. It is a broken hand-off: brilliant analysis that never gets checklisted, or a well-checklisted thesis that gets oversized because the sizing stage was skipped, or a perfectly sized position that quietly correlates with everything else you own because portfolio-fit was an afterthought. The loop is only as strong as its weakest coupling.

Third, the loop has a spine, an engine, and a floor, and those three are what remains to build. The spine is the decision journal (stage 6): it runs through every stage, because the journal entry is where the analysis, the checklist flags, the sizing rationale, the portfolio-fit call, and the kill criteria all get written down in one place, before the outcome, as the permanent record the whole process is built to produce and consult. The engine is calibration and review (stage 9): the mechanism that turns experience into improvement. The floor is survival: the constraint that sits under every stage and vetoes anything that could force you out of the game, however attractive its expected value, because a loop only compounds for a player who is still playing. Spine (§4.2–4.3), engine (§4.4–4.5), floor (§4.6–4.7), and then the operating manual that codifies the whole thing (§4.8).

4.2 The decision journal: the spine of the process

If the loop is the skeleton, the decision journal is its spine: the single structure that runs through every stage and holds the whole process upright. You met it twice already: M0.04 installed the original eight-field template as part of your knowledge system, and M6.05 argued why it works (it defeats hindsight bias) and had you file your first entries. The third step expands that template into the full investment-decision schema you will use for real capital and real positions, and makes the case for pre-commitment airtight. Everything downstream (calibration, the process scorecard, the quarterly letter) is only as good as the honesty of what you wrote here, before you knew.

Why pre-commitment defeats hindsight bias: the mechanism, stated exactly. Human memory is not a recording that gets played back. It is a reconstruction that gets rebuilt each time you recall it, and silently edited by everything that has happened since. After a stock triples, the reconstruction inserts a confidence you did not have ("I always knew it was special"); after it halves, it inserts a doubt you did not have ("I had a bad feeling"). This is hindsight bias, and its damage is specific and fatal. It destroys the one input that learning from a noisy, delayed-feedback game absolutely requires: an accurate record of what you actually believed, and how sure you were, at the moment you decided. Without that record you cannot tell a good decision that got unlucky from a bad decision that got lucky, so you cannot learn which of your processes to keep and which to kill, and you will "learn" the wrong lessons with total, unearned confidence.

There is exactly one defense, and it is mechanical rather than psychological: write the reasoning down before the outcome exists, and never edit it after. You cannot out-remember hindsight bias, because memory is compromised at the source. You can only route around it, by creating a fossil, a dated and unalterable record of your prior state of mind that your future, outcome-poisoned memory cannot touch.

The journal is that fossil. Its value is precisely that it cannot flatter you: it will, over the years, embarrass you into calibration, which is the only way calibration ever happens.

The expanded template: the investment-decision entry. M0.04's eight fields were sized for a study prediction ("I predict HUL's volume growth was 2–6%"). A real capital-allocation decision needs more, because it must also record how much you committed, why that size, how it fits the portfolio, and exactly what you will watch, so that the monitoring and sell stages have something objective to consult. This is the schema the app's journal uses and the one you will run for life. Copy it into 90-templates/investment-decision.md:

DECISION JOURNAL ENTRY — <TICKER / Company>
Date: <YYYY-MM-DD>   |   Decision: <BUY / ADD / TRIM / SELL / PASS / WATCH>
Analyst state: <one honest line — calm? excited? time-pressured? in a drawdown?>

1. THESIS (three sentences max — if it needs more, you don't yet understand it).
   What it does & why it's a GOOD business (ROIC vs WACC + the moat), and
   WHY IT IS MISPRICED (the variant view — what does the market have wrong?).

2. THE 2–3 THINGS THAT MUST GO RIGHT. The load-bearing assumptions the
   thesis rests on. If these hold, you win; if they break, you don't.

3. DISCONFIRMING EVIDENCE / THE BEAR CASE — in my own hand, stronger than a
   bear could put it. The strongest current facts against me, stated fairly.

4. BASE RATE (the outside view). Of companies like this — same growth claim,
   same leverage, same turnaround/roll-up/moat profile — what fraction
   actually delivered? (Mauboussin Base Rate Book / your own screens.)

5. WHAT THE PRICE IMPLIES (reverse DCF, M3.05). The growth/margin today's
   price is assuming, and why my view is better-founded than that.

6. CONFIDENCE: <NN>%  — that the thesis proves right by the review date
   (defined below). A number, not "fairly confident." This is the
   calibration record; be honest, not brave.

7. WHAT WOULD CHANGE MY MIND — the kill criteria. Specific, observable,
   pre-committed: "if <metric> falls below <X>" / "if <disclosure> appears."
   Decided NOW, while calm and unpositioned, for a future panicked self to obey.

8. MONITORING METRICS & CADENCE. The 3–5 numbers that actually test the
   thesis (the KPIs from the sector playbook), and how often I'll check them.

9. SIZING RATIONALE (M9.01). Target size = <NN>% of portfolio. Why:
   conviction vs. downside; the realistic downside floor (and why it's REAL);
   fractional-Kelly / limit that produced this number.

10. PORTFOLIO FIT (M9.02). What it correlates with in the book; which
    marginal idea/cash it displaces; why it clears the hurdle.

11. PRE-MORTEM. "It's 3 years later, this was a disaster — what happened?"
    The one or two most likely failure paths, in specifics.

Review date: <YYYY-MM-DD>   |   (Reviews are APPENDED below, separately dated.
                                 The original above is NEVER edited.)

Field by field, the logic: 1–2 are your compressed answer to the five questions and the load-bearing assumptions: the thesis and its joints. 3–5 are the pre-commitment against your own confirmation bias: the bear case in your own hand (if you cannot state it better than a bear, you do not understand the situation), the base rate that prices your specific story against the historical batting average of stories like it, and the reverse-DCF check that turns "it's a great company" into "here is what the price already assumes and why I disagree." 6 is the single number that makes calibration possible. Repeated across dozens of entries, it is the raw data of §4.4. 7–8 are the instruments of the monitoring and sell stages: kill criteria written by a calm past self for a panicked future self to consult (the antidote to sunk-cost and loss-aversion), and the specific metrics that will test the thesis, so "is it still valid?" is an objective question, not a mood. 9–10 connect the journal to the sizing and portfolio stages so the record is complete. 11 is the pre-mortem (M6.03), which converts each failure path into something to watch or size for.

Two rules, both from the working agreements, both absolute. Write before you know. An entry composed after the fact is hindsight cosplay, worth nothing, and worse than nothing because it feels like learning. Never edit after the outcome. You may append a later, separately dated review (that is where you record what happened and grade yourself), but you never touch the original. The original's entire value is that it is a fossil; a fossil you can re-carve is a forgery.

The journal as a Bayesian instrument. The entry is not written once and filed. Between the decision and the review date, evidence arrives (an earnings print, a regulatory event, a competitor's move) and each piece should move your confidence, in an appended, dated update, in the direction and to the degree the evidence is diagnostic (M6.03's Bayes). "Strong opinions, weakly held": conviction enough to size and act, humility enough to revise when the likelihoods say to. A thesis that never moves between decision and review is not steadfast. It is unmonitored. A well-kept journal breathes.

Worked Example 1 fills every field on a real Indian company at genuine scale. Study its texture before you write your own: the specificity of the kill criteria, the honesty of the confidence number, the way the bear case is genuinely argued rather than strawmanned.

4.3 Quarterly review letters: writing to think

The journal is written one decision at a time, at the moment of decision. The quarterly review letter is written all at once, at a fixed cadence, about the whole portfolio and all the quarter's decisions together. It does something the per-decision journal cannot: it forces you to explain every position you own and every action you took as a coherent, defensible whole, to an audience that will not accept hand-waving. That audience is, at minimum, your future self; ideally it is a real or imagined demanding investor, someone who has given you their capital and their trust and is entitled to know, in plain language, what you did with it and why.

This is not a program invention. It is the deliberate, decades-long practice of the best capital allocators alive, and they are unanimous about why they do it. Warren Buffett has written a long, candid annual letter to Berkshire shareholders every year since the 1970s, explaining the businesses, owning the mistakes by name ("I was wrong," about Dexter Shoes, about airlines, about the price paid for Precision Castparts), and teaching as he goes. He has said he writes them imagining he is reporting to his sisters, intelligent non-specialists who have their entire net worth with him and have been away for a year. Howard Marks has written his Oaktree memos since 1990, and his account of why is the sharpest statement of the principle in finance: he writes not primarily to inform readers but because the act of writing forces him to develop and test his own thinking. A half-formed view survives in the head but collapses the moment you try to defend it on the page.

The discipline is the point. As the writer Flannery O'Connor put it and every honest analyst rediscovers: I write because I don't know what I think until I read what I say. Writing is not the transcription of finished thought; writing is how thought gets finished. A thesis you have only thought is a thesis you have not yet stress-tested. The page is where the vague parts, the assumptions you were skating over, and the position you cannot actually justify all become visible, because the sentence will not complete unless the reasoning does.

The format. A quarterly letter has a stable structure, which is part of its power: the same skeleton each quarter makes the changes legible and makes it impossible to quietly skip the uncomfortable parts:

1. The scoreboard, honestly framed. Portfolio result for the quarter and cumulatively, immediately reframed away from short-run P&L as a skill signal (§4.5): one quarter's return is almost all noise, and you say so explicitly, every time, so you never train yourself to read it as a verdict.
2. Position by position: the standing thesis. For every holding: what it is, why you own it, whether the thesis is intact / strengthening / weakening, and how the monitoring metrics moved this quarter versus what the thesis needed. This is the section that forces the work: you cannot write "thesis intact" for a position whose numbers are quietly deteriorating without either lying (which the record will later expose) or noticing (which is the point).
3. Decisions taken: every buy, sell, add, trim, and pass. Each with its reasoning, linked to the journal entry, and each judged against the reasoning you wrote at the time, not the reasoning hindsight would prefer.
4. Mistakes and what I got wrong. The section that separates a review letter from a marketing document. Named, specific, honest: the thesis that broke, the kill criterion you were slow to honor, the position you oversized, the idea you passed on that you shouldn't have. Buffett's letters are studied partly because of this section; a letter without it is propaganda.
5. Calibration and process. This quarter's resolved predictions versus your stated confidences (§4.4), your running Brier score, and any process-scorecard patterns (§4.5). Together they are the objective, luck-proof read on whether your judgment is improving.
6. What changed my mind. Where evidence updated a thesis, where you revised a view, what you learned. Intellectual honesty made visible.

Why the cadence is quarterly, and why it is written even when no one reads it. Quarterly is frequent enough to catch a deteriorating thesis before it becomes a disaster and to keep the calibration record fresh, but infrequent enough that you are reviewing businesses over a meaningful interval rather than reacting to price noise. It enforces the long-horizon, business-owner framing the whole program is built on. And you write it even if the only reader is you in five years, because the value is in the writing, not the reading: the clarity, the forced honesty, and the record are produced by the act of composition regardless of audience. Worked Example 4 is a model letter, abbreviated, but showing the format and, above all, the texture of the mistakes section.

4.4 Calibration: the engine of process improvement

You cannot, over any short horizon, tell whether your investment judgment is good by looking at your returns. Returns are dominated by luck over months and years (§4.5 proves this), so a great process can trail and a terrible one can lead for a long time. That is the deepest problem in the business, and what follows is the idea that makes "grade the process" operational rather than pious. You need a scoreboard for judgment that is immune to luck. Calibration is that scoreboard. It does not ask "did you make money?" It asks a question luck cannot answer for you: "Of all the times you said you were 70% sure, did the thing happen about 70% of the time?" That question has a right answer, it is checkable from your journal, and it measures the one thing you actually control, the honesty and accuracy of your probability estimates, with no reference to any single outcome's luck.

Predicted probability versus realized frequency: the whole concept. A forecaster is calibrated when their stated confidences match reality in aggregate: across everything they called 70%, about 70% come true; across everything they called 90%, about 90% come true. Note what this is not: it is not being right a lot. A weather forecaster who says "30% chance of rain" and is perfectly calibrated is supposed to be wrong (it rains) 30% of the time on those days. Being wrong 30% of the time is exactly what a good 30% forecast looks like. Calibration is not accuracy; it is the truthfulness of your uncertainty. And it is trainable. Tetlock's Superforecasting is the empirical proof that ordinary people become dramatically better forecasters precisely by keeping score this way and adjusting, which is why the whole program routes your confidences through a journal and a dashboard: not to shame you, but because keeping score is the training.

The reliability curve. Plot your stated confidence on the horizontal axis and your actual hit rate on the vertical axis, one point per confidence bucket. Perfect calibration is the 45° diagonal (stated = actual). The shape of your curve is a diagnosis:

• Points below the diagonal → overconfidence. Your "90%" comes true only 65% of the time; you claim more certainty than you have. This is the single most common and most expensive analyst pathology, and it is usually worst at the high-confidence end, exactly where you size largest, so it does maximum damage.
• Points above the diagonal → underconfidence. Your "60%" comes true 80% of the time; you are timid, and you are leaving conviction (and size, and return) on the table. Rarer, but real, and often found in careful, anxious analysts.
• The tell is usually in the tails. Most people are roughly calibrated in the middle (a "60%" really is about 60/40) and badly overconfident at the top: the "sure things" that come true two times in three. Your calibration curve makes this visible as a fact, not a suspicion.

Reading it by domain, not just by level. The single most useful refinement, and one most people never do: bucket your calibration by domain, not only by confidence level. You may be beautifully calibrated on banks (a sector you have studied to the bone) and wildly overconfident on technology (your circle's edge, §M6.03). You may be calibrated on "is this a good business?" and hopeless on "will this catalyst happen on schedule?" The aggregate curve hides this; the domain-split curve reveals it. It tells you, with data, exactly where your circle of competence really ends (which is a more honest map than the one you drew from introspection in M6.03) and which kind of judgment to stop making or to make smaller. Overconfidence is rarely uniform; it is concentrated, and calibration-by-domain is how you find the concentration.

The Brier score: one number to track. When you want a single scalar for the whole record (to watch it fall over years, or to compare yourself this year to last), use the Brier score, the mean squared error of your probabilistic forecasts. For each resolved prediction take your probability p (a decimal) and the outcome o (1 if it happened, 0 if not); the Brier score is the average of (p − o)²:

Brier = (1/N) × Σ (pᵢ − oᵢ)²

It runs from 0 (perfect) to 1 (confidently, uniformly wrong). The benchmark you must beat to claim any skill is 0.25, the score you get by mindlessly saying "50%" to everything, and beating it is harder than it sounds, because the squaring means a confident wrong call (0.90 confidence on something that does not happen contributes (0.90 − 0)² = 0.81) is punished savagely, while an honest 50/50 contributes only (0.50 − 0)² = 0.25 whichever way it resolves. This is the mathematics of humility: one overconfident blowup can sink an otherwise good record, which is the quantitative case for never claiming 90% unless you have earned it. (A subtlety worth knowing: the Brier score decomposes into two parts. The calibration component measures how well your stated probabilities match outcomes. The resolution or discrimination component measures how well you separate things that happen from things that don't by giving them different probabilities. A forecaster who says "50%" to everything is perfectly calibrated in a trivial sense but has zero resolution and no skill. You want both: honest probabilities and the courage to differentiate. Track the Brier; understand it is rewarding calibration and discrimination together.)

A worked calibration score in miniature (the full ten-prediction exercise is Worked Example 2 and Practice Problem 4; do both by hand). Three resolved theses: p = 0.90 (happened, o=1), 0.70 (did not, o=0), 0.60 (happened, o=1). Squared errors: (0.90−1)² = 0.01; (0.70−0)² = 0.49; (0.60−1)² = 0.16. Brier = (0.01 + 0.49 + 0.16)/3 = 0.66/3 = 0.22, just better than coin-flipping, and dragged there almost entirely by the confident 0.70 miss. The lesson is visceral: your confident misses are what wreck your score, so the highest-value correction is almost always at the top of your confidence scale.

From diagnosis to correction: closing the loop. Calibration is the engine of improvement only if you act on the reading. The corrections are mechanical and immediate:

• Overconfident at the top (the usual finding): haircut your high-confidence calls, training yourself to say 70% where you feel 90%, and, because overconfidence does its damage through oversizing, feed the haircut straight into smaller position sizes (M9.01). This single adjustment, repeated, is one of the highest-return habits in the program.
• Overconfident in a specific domain: shrink size there, widen your ranges there, or, if the miscalibration is severe and persistent, move that domain from the "edge" to the "outside" of your written circle. The data has told you where you cannot see; believe it.
• Underconfident: the rarer, gentler problem. Permit yourself larger sizes on your genuine high-conviction ideas, because you have earned the confidence and are taxing yourself for false modesty.
• Whatever the reading, keep scoring. Calibration improves only through the feedback loop; a falling Brier over years is the objective, luck-proof evidence that your judgment is genuinely getting better, and over short horizons it is the only kind that exists. It is the closest thing an investor has to a training-load chart.

Common mistakes & how experts think differently

Owning the components but never wiring the loop. The most common failure at this stage is subtle: the learner has every skill (can value, can checklist, can size) but runs them as disconnected activities rather than a coupled loop, so nothing compounds. The tell is a hand-off that never happens: analysis that never gets checklisted, a checklist that never feeds sizing, a decision that never gets journaled, a quarter that never gets reviewed. Experts obsess over the seams between stages, because a process is only as strong as its weakest hand-off, and because stage 9 (review) feeding stage 1 (sourcing) is the entire source of long-run improvement.

Treating the journal as a diary instead of an instrument. Beginners write journal entries that are narratives ("bought X, feeling good about it") rather than instruments: falsifiable, numeric, with pre-committed kill criteria. A diary records feelings; an instrument produces calibration data and governs future drawdown decisions. The difference is whether the entry contains a number (confidence) and a trigger (what would change my mind), written before the outcome. Without those two things the entry cannot be scored and cannot defend you later; it is a mood with a date on it.

Writing the quarterly letter as marketing. The letter's whole value is the mistakes section and the honest calibration read; a letter that only lists winners and reframes losers as "temporary" is propaganda, and worse, it trains you to lie to yourself, the one reader you can never afford to fool. Experts (Buffett, Marks) are studied precisely because they name their errors; the willingness to write "I was wrong about X, here is why" is the discipline, not a blemish on it. Write the letter you would be uncomfortable showing a rival, then show it to your journal.

Confusing calibration with accuracy, and grading it on outcomes. Two errors here. First, thinking calibration means "being right a lot": a perfectly calibrated 30% forecast is supposed to be wrong 70% of the time; calibration is the truthfulness of your uncertainty, not your hit rate. Second, judging a single prediction's calibration by its outcome: one resolved bet tells you almost nothing (a 70% call that fails is fully consistent with excellent calibration); only the aggregate is meaningful, meaning whether your 70%s came true ~70% of the time across dozens. Experts think in the aggregate and by domain; amateurs "result" on each one.

Chasing the P&L scoreboard instead of the process scoreboard. This is resulting at the portfolio level, and it is the master error the whole module exists to fix. Over short horizons the P&L is dominated by luck (the paradox of skill), so optimizing it directly means optimizing for luck. You will abandon a sound process after an unlucky year and cling to a reckless one after a lucky one, which is exactly backwards. Experts run both scoreboards, know the process score is the one they control and the one that causes long-run returns, and have the temperament to keep running a good process through a bad patch and to kill a bad process that is currently winning.

Leverage that "can't" force liquidation — until it does. The rationalization is always the same: "my leverage is modest / my lender is reasonable / this asset can't fall that far." Every forced-liquidation blow-up in history was run by someone who believed a version of that sentence. Experts treat any callable leverage, any short-funded long position, any redeemable outside capital as a live path to forced selling at the bottom, and they close those paths structurally rather than trusting that the bad scenario won't arrive. Survival is a first-order constraint, and you only need to be forced out once.

White-knuckling drawdowns instead of running the protocol. The beginner tries to feel calm in a crash (willpower, which loses) or, worse, tries to think clearly for the first time while down 40% (impossible; that is when reasoning is most compromised). Experts do neither: they run a pre-written protocol (cooling-off rule, open the original entry, check kill criteria, branch, log) so a calm past self makes the decision. Temperament is not summoned in the moment; it is installed in advance as procedure.

A process document that is aspirational poetry. "I invest in wonderful businesses at fair prices with a margin of safety and a long horizon" is a vibe, not a process. It contains no operable rule, vetoes nothing, and bends to whatever you feel like doing. A real process document is specific and veto-shaped: exact screens, exact size limits, exact sell reasons, an explicit anti-goals list. The test of every line is: could someone else execute this, and does it ever say no? If a line cannot be executed by a stranger or never forbids anything, it is decoration.

Optimizing the loop for the one great call. Beginners fantasize about the ten-bagger that makes their career. The process you are building is optimized for the opposite: a high batting average of sound, survivable decisions over decades, most of them unremarkable. Experts know that the compounding comes from not blowing up and from many good-process decisions, not from one heroic bet, and they would rather have a boring, repeatable 12% than a thrilling, unrepeatable 40%-then-ruin. The loop is a machine for consistency, and consistency, compounded, is what wins.

Worked examples

These four carry the module's core instruments end to end. WE1 is Indian (a full journal entry); WE4 is US (a review letter); WE2 and WE3 are jurisdiction-neutral computations. Work each with a pen before reading on.

WE1: A complete decision-journal entry (Indian company)

Setup (figures illustrative, modeled on a large Indian paints/adhesives compounder at ~FY25 scale; pull the actual filings before using any number). After a full deep dive, you are deciding whether to open a position in "an Indian decorative-paints leader". Treat it as modeled on the scale and economics of an Asian Paints. It trades at roughly 55× trailing earnings; ROCE has sat around 25–30% for a decade; a new entrant (a large industrials conglomerate) has just entered the category with deep pockets. Here is the entry, written before you place any order and before you read a single broker target.

DECISION JOURNAL ENTRY — Indian Decorative Paints Leader (illustrative)
Date: 2026-07-15   |   Decision: WATCH (not yet BUY — see confidence & price)
Analyst state: Calm; genuinely admire the business; wary that admiration ≠ edge.

1. THESIS. A dominant decorative-paints franchise with a distribution moat
   (100k+ dealers, tinting machines placed in-store = switching costs +
   efficient scale in a fragmented-demand category) earning ~27% ROCE, far
   above its ~12% cost of capital, and reinvesting in reach. It is a GOOD
   business. Is it MISPRICED? At ~55x, only if durability is under-appreciated
   AND the new entrant fails to dent returns — which is NOT clearly my variant
   view. Leaning that the price already embeds the quality.

2. THE 2–3 THINGS THAT MUST GO RIGHT. (a) The distribution/tinting moat holds
   the new entrant to low-single-digit share over 5 yrs; (b) volume growth
   stays mid-to-high single digit as housing/repainting cycles run; (c) ROCE
   stays > 20% (no margin war).

3. DISCONFIRMING EVIDENCE (my own hand, steelmanned). A cash-rich conglomerate
   with its own retail/distribution reach is the most credible entrant the
   category has ever seen; paints is a repeat-purchase, brand+distribution
   game they understand. If they buy share with price, ROCE and the multiple
   both compress — a 55x stock re-rating to 35x on a growth/return scare is a
   ~35% loss BEFORE any earnings miss. History says incumbents usually hold,
   but "usually" priced at 55x is a thin margin of safety.

4. BASE RATE. Of dominant consumer franchises at ~25%+ ROCE and this scale,
   what fraction still earn > 20% ROCE a decade later? Reasonably high for
   true distribution moats (fade is slow) — but of those trading at ~55x,
   the fraction that BEAT the market from that multiple over 5 yrs is far
   lower (Mauboussin: high starting multiples are a headwind regardless of
   quality). Two different base rates; the second is the binding one.

5. WHAT THE PRICE IMPLIES (reverse DCF). At ~55x with a ~12% cost of equity,
   the price implies roughly low-to-mid-teens FCF growth sustained ~15 yrs
   with no ROCE erosion. Plausible for this business — but it is the BULL
   path priced as the base case, leaving no margin for the entrant scenario.

6. CONFIDENCE: 55% that, bought at today's price, this beats my ~12% hurdle
   over 5 yrs. (High confidence it's a great business; LOW confidence it's a
   great purchase HERE. The distinction is the whole point.)

7. WHAT WOULD CHANGE MY MIND (kill criteria). BUY trigger: price falls to
   ~40x (a ~25% de-rating) OR the entrant is 3 yrs in with < 3% share and
   incumbent ROCE still > 20% (durability confirmed, uncertainty resolved).
   AVOID/EXIT trigger if owned: incumbent gross margin falls > 300bps in a
   year on competitive pricing, OR dealer count / tinting-machine base stops
   growing, OR the entrant passes ~8% share.

8. MONITORING METRICS. Quarterly: volume growth (want mid-single+), gross &
   EBITDA margin (watch for a pricing war), ROCE (annual), dealer count &
   tinting-machine placements, and the entrant's disclosed paints revenue.

9. SIZING RATIONALE. On a BUY (only if a trigger hits): starter 3%, because
   conviction in the business is high but the downside (multiple compression)
   is real and only partly capped; a 55x entry has no valuation floor, so
   size stays small until price gives margin of safety. Full 5–6% only near
   40x.

10. PORTFOLIO FIT. Correlates with my other Indian consumer-quality names
    (HUL-like, Titan-like) — this is a factor cluster; adding it concentrates
    "Indian premium-consumer-quality," so it must clear not just the hurdle
    but the correlation limit. Likely displaces the weakest consumer name on
    the bench, not cash.

11. PRE-MORTEM. It's 2029, this lost 40%. What happened? The conglomerate ran
    a multi-year price/promotion war; volumes held but margins and ROCE fell;
    the market decided the moat was narrower than believed and re-rated 55x ->
    32x; my "great business" was true and my "great purchase at 55x" was false.

Review date: 2026-10-15 (and on any kill-criterion trigger).

What to notice. The decision is WATCH, not BUY. The honest confidence number (55%) and the missing margin of safety produced a pass at this price, which is a complete and disciplined output of the loop. The bear case is genuinely argued, not strawmanned. The kill criteria are specific and observable (numbers and thresholds, not feelings), so a future you can check them mechanically. And the entry cleanly separates "great business" (high confidence) from "great purchase here" (low confidence), which is the exact discipline that stops admiration from masquerading as edge. This entry, filed and unedited, is now a fossil: in October, and in 2029, you can grade your past self honestly.

WE2: Scoring your calibration (the ten-prediction exercise)

Setup. A learner's journal has ten resolved predictions, each with a confidence stated in advance and a known outcome (o = 1 if it happened). Score the calibration.

Step 1: build the calibration (bucket) table. Group by stated confidence; actual hit rate = came-true ÷ count.

Step 2: read the reliability curve. Plot stated (x) vs actual (y). The 50% and 60% points sit on or just above the diagonal, which is fine. The 80% point sits below the diagonal (stated 80, actual 60), the classic overconfidence at the top: the "near-sure" calls come true only three times in five. The middle and bottom are honest; the top is not.

Step 3: compute the Brier score. Brier = (1/N) Σ (p − o)².

• 80% bucket: three hits contribute (0.8−1)² = 0.04 each → 0.12; two misses contribute (0.8−0)² = 0.64 each → 1.28. Subtotal 1.40.
• 60% bucket: two hits (0.6−1)² = 0.16 → 0.32; one miss (0.6−0)² = 0.36. Subtotal 0.68.
• 50% bucket: one hit (0.5−1)² = 0.25; one miss (0.5−0)² = 0.25. Subtotal 0.50.
• Total = 1.40 + 0.68 + 0.50 = 2.58. Brier = 2.58 / 10 = 0.258.

Step 4: interpret. A Brier of 0.258 is slightly worse than the 0.25 "always say 50%" benchmark. That is a humbling result, and the cause is entirely the two confident 80% misses (0.64 each = 1.28 of the 2.58 total). The squaring means confident wrong calls dominate the score; a couple of them wipe out the credit earned everywhere else.

Step 5: the corrections. (a) Haircut the top: treat your instinctive "80%" as ~65% until the record says otherwise. (b) Downsize high-confidence bets: since overconfidence damages through oversizing, the 80%-bucket ideas should carry smaller positions than their confidence suggested, not larger (M9.01). (c) Keep scoring: this is ten data points; calibration is trainable, and the same table in two years, with the 80% point pulled up to the diagonal, is objective proof the training worked. The exercise converted a vague worry ("am I overconfident?") into a located, quantified, correctable fact, which is exactly what stage 9 of the loop is for.

WE3: Skill versus luck: the process scorecard on two decisions

Setup. Two decisions have resolved. Decision P (a mid-cap industrial): full deep dive, checklist run, bear case written, sized at a disciplined 4% with a verified downside floor, kill criteria set. Then it lost 30% when a cyclical downturn (a risk the entry explicitly named and sized for) arrived. Decision Q (a small-cap tip from a friend): no real analysis, no checklist, no journal, oversized at 12% on excitement. It tripled. Your friend says Q was obviously the smart move. Score both on process, blind to outcome.

Step 1: score the process (0/1/2 per dimension, out of 16), from the journal/record alone.

Step 2: place each in the decision × outcome grid. Decision P: good decision, bad outcome (top-right), a well-reasoned, well-sized bet where a correctly-identified 30%-ish risk simply happened; variance, not error. Decision Q: bad decision, good outcome (bottom-left), an unanalyzed, oversized gamble that got lucky.

Step 3: the verdict, and why it is the opposite of the friend's. The friend is resulting (M6.03): judging the decisions by their P&L, which in a luck-heavy domain is nearly information-free over one outcome each. The honest scoreboard is the process score, and it says: keep P's process and repeat it (a 15/16 process that lost is supposed to lose sometimes, being a well-sized bet on a real, identified risk, and the loss was survivable exactly because it was sized right), and kill Q's process (a 1/16 process that won is a warning, not a validation: it is a habit that will, run repeatedly, eventually deliver the oversized loss that ends you, and the win just delayed the lesson). If you learn the friend's lesson, you will "learn" to stop doing deep work and start taking oversized tips, training yourself, via one lucky outcome, into eventual ruin. You reinforce the process that deserves it and kill the one that doesn't, regardless of which made money this time. That is what it means to grade the process.

WE4: A quarterly review letter (US company, abbreviated model)

Setup. An abbreviated model letter. It shows the format and, above all, the texture of the mistakes section, which is the load-bearing part. Figures illustrative.

Q2 2026 Review — to myself, and to anyone whose capital I steward 1. Scoreboard, honestly framed. The portfolio returned +4% this quarter, +11% trailing twelve months. I am recording this only to note that one quarter's return is almost pure noise and tells us nothing about whether the process is sound. The real scoreboard is sections 4 and 5 below. If I ever start leading this letter with the return as if it were a verdict, that itself is a process failure. 2. Positions (standing theses). Freight railroad (US, ~6% position): own it for the efficient-scale + duopoly cost moat and pricing power; thesis intact: operating ratio improved ~120bps, volumes soft but pricing held, exactly as the thesis needs. Payments network (US, ~7%): two-sided network moat; thesis intact and strengthening: cross-border volume reaccelerated. Regional lender (US, ~3%): thesis weakening: deposit costs rose faster than the entry assumed; monitoring metric (net interest margin) is at the low end of my tolerance band, one more bad quarter trips a kill criterion. 3. Decisions this quarter. Bought the payments network (journal 2026-04-18): reverse-DCF implied ~9% growth, my base-rate-anchored estimate ~12%, a genuine variant view; sized 5% starter. Trimmed the railroad from 8% to 6%: not because the thesis broke but because it had grown past my single-name comfort into a position that was making the portfolio's result about one stock. Passed on a hot AI-adjacent name: outside my written circle; no edge; a clean pass. 4. Mistakes and what I got wrong. The lender. I under-weighted deposit-beta sensitivity at entry. My bear case (own hand) mentioned it but my confidence number (70%) did not reflect how binding it was; a more honest entry would have been 55% and a 2% starter, not 3%. That is a calibration miss on a sector where my domain data (§below) already flags me as overconfident, and I should have applied the haircut and didn't. Logged; the checklist gains a line: "For any spread lender, stress deposit beta explicitly and let it set the confidence number." The one I let get away: I passed on a quality compounder eighteen months ago because it "looked expensive" on trailing P/E without running the reverse DCF; it has compounded ~40% since. The error was using a lazy multiple instead of my own process, a process failure that happened to cost only opportunity, this time. 5. Calibration & process. Nine predictions resolved this quarter; my 80% bucket came true 3-of-5 again (still overconfident at the top, so the correction isn't sticking yet and I am hard-coding the haircut into the journal template). Running Brier 0.19, down from 0.23 a year ago. Domain split: well-calibrated on the industrials/payments names, overconfident on lenders, which is now a written note in my circle-of-competence section, moving "regional banks" from inside toward edge. Average process score on this quarter's decisions: 14/16 (the lender entry, at 11/16, dragged it down, and tellingly it is also the weakest position; process and outcome are agreeing here, which is reassuring but not something I'll count on over one quarter). 6. What changed my mind. The payments cross-border reacceleration raised my confidence on that name from 65% to 78% (appended to the journal, dated). Nothing else moved enough to act on.

What to notice. The mistakes section is specific and self-implicating: a named calibration miss, a named opportunity lost, each converted into a rule or a circle-of-competence revision. The scoreboard is deliberately de-emphasized. And the letter closes the loop: section 4's mistake becomes a new checklist line and a circle revision, which will change stage 1 and stage 3 next quarter. That is a review letter doing its job: not reporting the past, but improving the process.

WE5: Performance attribution: was the edge allocation or selection? (Indian portfolio)

Setup (synthetic). Your quarterly review (§4.3) reports a good number. Good how, and from what? You ran a ₹50,00,000 India equity book for the period against a Nifty-like benchmark. To learn where the return actually came from, and therefore what your edge is, decompose the excess return into an allocation effect (did over/under-weighting whole sectors help?) and a selection effect (did your stocks beat the benchmark within each sector?), using the classic Brinson-Hood-Beebower identity, computed per sector and summed:

• Allocation = (portfolio weight − benchmark weight) × benchmark sector return
• Selection = benchmark weight × (portfolio sector return − benchmark sector return)
• Interaction = (weight difference) × (return difference)

The three effects sum exactly to the total excess return, and the reconciliation is your arithmetic check.

Step 1: the two headline returns. Benchmark = 0.40×10 + 0.30×5 + 0.30×8 = 7.9%. Portfolio = 0.50×12 + 0.20×6 + 0.30×9 = 9.9%. Excess = +2.0%, which, on ₹50,00,000, is ₹1,00,000 of outperformance the review must now explain.

Step 2: allocation effect (the sector bets). Per sector, (wₚ − w_b) × R_b: Financials (0.10)(10) = +1.0; IT (−0.10)(5) = −0.5; Consumer (0)(8) = 0. Allocation total = +0.5%. Over-weighting Financials (a strong 10% benchmark sector) helped; under-weighting IT (a weak 5% sector) helped a little more; net +0.5%.

Step 3: selection effect (the stock-picking). Per sector, w_b × (Rₚ − R_b): Financials 0.40×(12−10) = +0.8; IT 0.30×(6−5) = +0.3; Consumer 0.30×(9−8) = +0.3. Selection total = +1.4%. Your holdings beat the benchmark inside every sector.

Step 4: interaction, and the reconciliation. (wₚ − w_b) × (Rₚ − R_b): Financials 0.10×2 = +0.2; IT −0.10×1 = −0.1; Consumer 0 = 0 → +0.1%. Sum of effects = 0.5 + 1.4 + 0.1 = +2.0% = the excess exactly. In rupees: selection ₹70,000, allocation ₹25,000, interaction ₹5,000.

What to notice. The story is selection, not allocation: ₹70,000 of the ₹1,00,000 came from picking better stocks within sectors, only ₹25,000 from sector tilts. That is a fact about your edge, and it feeds the loop directly: it is the quantitative cousin of calibration-by-domain (§4.4). If, quarter after quarter, your value-add is selection while your allocation effect is noise, your written circle (§4.8) should say "edge = bottom-up selection; keep sector weights near the benchmark," and your hours should go to stock analysis, not sector timing. Attribution converts "I did well" into "I did well at this specific thing", which is the only version stage 9 can actually learn from.

WE6: Turnover and tax drag: why selling early is expensive (US)

Setup (synthetic). Two investors hold the same asset earning the same 8% pre-tax return for 20 years. Patient buys and holds, selling only at the very end and paying long-term capital-gains tax of 20% once, on the whole gain. Restless turns the position over every year, realizing the gain annually and paying the short-term/ordinary rate of 35% each time. Same security, same pre-tax return; the only difference is turnover. How far apart do they finish? (This is the arithmetic beneath M9.03's "don't sell compounders early" and stage 8 of the loop.)

Step 1: Patient (buy-and-hold, tax deferred). Pre-tax, $1 grows to 1.08^20 = 4.661×. Gain = 3.661. Tax at the end = 20% × 3.661 = 0.732. After-tax terminal = 4.661 − 0.732 = 3.929×, an after-tax CAGR of 3.929^(1/20) − 1 = 7.08%.

Step 2: Restless (100% turnover, taxed short-term every year). Realizing the gain each year makes the annual compounding factor not 1.08 but 1 + 8%×(1 − 0.35) = 1.052. Over 20 years: 1.052^20 = 2.756×, an after-tax CAGR of 5.20% (by construction).

Step 3: the gap. Same 8% pre-tax asset: Patient ends at 3.929×, Restless at 2.756×, so Patient has ~42.5% more terminal wealth. On a $1,000,000 start, that is $3,928,766 vs $2,756,226, a $1.17m difference, and ~1.88 percentage points more per year, after tax, forever, purely from not selling.

What to notice. Two forces compound together, and the review should name both. First, deferral: an unrealized gain is an interest-free loan from the tax authority, and the tax you have not yet paid stays invested and compounds for you (worth a great deal even at the same rate). Second, rate: churning realizes short-term gains taxed at the high ordinary rate instead of the lower long-term rate. High turnover pays both penalties every year, and because the drag compounds, a "small" annual tax bill becomes a seven-figure career gap. This is why turnover is a first-class number on the quarterly scoreboard. And the disposition-effect habit of cutting winners early (M9.03) is a quantifiable, compounding tax on your returns, not only a behavioral slip.

Practice set

Do each problem fully in writing before reading the solution (working agreement #3). Problems 1–2 are guided; 3–8 independent; 9–12 timed/synthesis. Numeric answers to the stated tolerance.

Problem 1 (guided: the loop). Here are the nine loop stages, shuffled: Portfolio fit · Review & Calibration · Sizing · Sourcing · Decision + Journal · Monitoring · Checklist · Analysis · Sell discipline. (a) Put them in order. (b) For the hand-off from Analysis, from Sizing, and from Review, state exactly what each stage passes to the next. (c) Which single stage, if skipped, stops the whole process from improving over time, and why?

Solution. (a) Sourcing → Analysis → Checklist → Sizing → Portfolio fit → Decision + Journal → Monitoring → Sell discipline → Review & Calibration (→ back to Sourcing). (b) Analysis hands forward a one-page thesis plus a valuation range (the input the checklist vets). Sizing hands forward a target position size with a verified downside floor (the input portfolio-fit tests for correlation and the marginal-idea hurdle). Review hands forward a better checklist (a new line per mistake), tighter/more honest confidence numbers, and a deeper, more variant bench, which is to say it improves stages 1, 3, and 6 for the next turn. (c) Review & Calibration (stage 9). It is the only stage that feeds back: without it the loop still functions but never learns: the checklist never grows, calibration never tightens, mistakes recur. Skipping stage 9 turns a compounding process into a static one; you would be ten years older running the same flawed process, not a better one.

Problem 2 (guided: journal critique). A learner files this entry: "2026-05-10. Buying CementCo. Great management, industry consolidating, stock looks cheap versus history. Going in at a full position. Confident." Against the expanded investment-decision template (§4.2), list every field that is missing or defective, and state why each omission specifically weakens the process.

Solution. Missing or defective, field by field: (1) Thesis: "great management, consolidating industry" is not a thesis; it never says how it makes money, whether ROIC > WACC, or why it is mispriced (the variant view). (2) The 2–3 things that must go right: absent. (3) Disconfirming evidence / bear case in own hand: absent; no defense against confirmation bias. (4) Base rate: absent; "cheap versus its own history" is not an outside view and is a classic value trap ("cheap for a reason"). (5) What the price implies (reverse DCF): absent; "looks cheap versus history" is a lazy relative multiple, not a reverse-DCF check (WE4's exact error). (6) Numeric confidence: "confident" is not a number, so it produces no calibration data. (7) Kill criteria: absent; nothing pre-committed, so a future drawdown will be decided by panic, not by a calm past self. (8) Monitoring metrics: absent; "is the thesis still valid?" will be a mood, not a check. (9) Sizing rationale: "full position" with no conviction-vs-downside logic and no verified floor; likely an oversize. (10) Portfolio fit: absent (correlation with other cyclicals?). (11) Pre-mortem: absent. As written, the entry cannot be calibrated, cannot govern a drawdown, and cannot distinguish a good decision from a lucky one. It is a mood with a date on it.

Problem 3 (repair to template). Rewrite the Problem 2 entry as a proper investment-decision entry, inventing reasonable specifics for a synthetic Indian cement company (label them illustrative). Include at least a real confidence number, two specific kill criteria, and a sizing rationale tied to a downside floor.

Solution (one acceptable version; specifics illustrative).

Date 2026-05-10 · Decision: BUY starter · State: calm. Thesis. Regional cement leader with a cost moat (limestone reserves + captive power + freight-advantaged plant locations) earning ~14% ROCE vs ~11% WACC; consolidating region is lifting utilization and pricing. Mispriced because the market extrapolates a recent weak-demand patch; I think mid-cycle margins are structurally higher post-consolidation. Must go right. (a) Regional utilization rises toward 80%+; (b) pricing discipline holds among the top-3 players; (c) no debt-funded capacity land-grab at the peak. Bear case (own hand). Cement is cyclical and capital-heavy; consolidation can reverse into a share war; a debt-funded expansion at the top would crater ROCE. Cheap-vs-history often means the cycle is rolling over. Base rate. Of consolidating regional cement markets, a minority sustain pricing discipline for 5+ years; the modal outcome is periodic price wars, so I anchor on "cyclical, mean-reverting margins" and demand the specific evidence (utilization, top-3 discipline) before straying bullish. Price implies (reverse DCF). At the current EV/EBITDA, the price implies roughly trough-to-mid-cycle EBITDA/tonne holding, which is undemanding if consolidation sticks; my variant view is that mid-cycle is higher than the market's trough extrapolation. Confidence: 60% that this beats my hurdle over 4 years. Kill criteria. (i) Top-3 regional realization falls >5% QoQ for two quarters (price war); (ii) the company or a major peer announces debt-funded capacity at >X net debt/EBITDA (supply glut at the peak). Either → exit. Monitoring. Quarterly: regional utilization, EBITDA/tonne, net debt/EBITDA, top-3 realization spread. Sizing. Starter 3%. Downside floor ~ −30% (cyclical trough multiple on mid-cycle EBITDA): real but not catastrophic, and not leveraged; full 5% only if utilization confirms. Fractional-Kelly on a 60/─30 profile supports a modest size, capped by the cyclical-sector limit. Portfolio fit. Correlates with my other India cyclicals; must clear the cyclical-cluster limit, displaces the weakest cyclical on the bench. Pre-mortem. 2030, −40%: a price war + a debt-funded expansion at the peak crushed ROCE; I mistook a cyclical uptick for a structural re-rating. Review date 2026-08-10.

The point of the exercise: every field converts a vague conviction into a checkable, falsifiable, survivable commitment.

Problem 4 (the calibration-scoring exercise). A learner's journal has ten resolved predictions with confidences and outcomes (o = 1 if it happened): #1 90% (o=1); #2 90% (o=1); #3 90% (o=0); #4 70% (o=1); #5 70% (o=1); #6 70% (o=0); #7 70% (o=1); #8 60% (o=1); #9 60% (o=0); #10 60% (o=1). (a) Build the calibration table (bucket, count, came-true, hit rate). (b) Diagnose over/under-confidence and where. (c) Compute the Brier score and compare to the 0.25 benchmark. (d) Give the two most important corrections.

Solution. (a) Calibration table:

(b) The 70% and 60% buckets are essentially well calibrated (75% and 67% versus stated 70% and 60%, within noise on tiny samples). The 90% bucket is badly overconfident: things called near-certain came true only two times in three (67% vs 90%, a 23-point gap), concentrated at the top of the scale: the classic and most expensive pattern. (c) Brier: 90% bucket → (0.9−1)²·2 + (0.9−0)²·1 = 0.02 + 0.81 = 0.83; 70% bucket → (0.7−1)²·3 + (0.7−0)²·1 = 0.27 + 0.49 = 0.76; 60% bucket → (0.6−1)²·2 + (0.6−0)²·1 = 0.32 + 0.36 = 0.68. Total = 2.27; Brier = 0.227, just better than the 0.25 know-nothing benchmark, dragged down almost entirely by the single confident 90% miss (0.81). (d) Corrections: (i) haircut high-confidence calls: treat your "90%" as ~70% until the record earns back the confidence; (ii) downsize the sure things, because overconfidence damages through oversizing, high-confidence ideas should not carry your largest positions until the 90% bucket actually hits ~90%. Keep scoring; the fix is verified only when the top of the curve rises to the diagonal.

Problem 5 (calibration vs resolution). Over six binary predictions of which exactly three occur, Forecaster A says "50%" to all six; Forecaster B says 80% to the three that occur and 20% to the three that do not. (a) Compute each Brier score. (b) Both are "calibrated" in some sense. Explain the difference and why B has skill and A does not.

Solution. (a) A: every term is (0.5−o)² = 0.25, so Brier = 0.25. B: three hits at 0.8 → (0.8−1)² = 0.04 each; three misses at 0.2 → (0.2−0)² = 0.04 each; sum = 6 × 0.04 = 0.24; Brier = 0.04. (b) A is trivially calibrated (says 50%, and 50% come true) but has zero resolution/discrimination: A never separates the events that happen from those that don't, so A has no skill; the 0.25 is the know-nothing score. B is calibrated and has resolution, assigning high probabilities to what happens and low to what doesn't, which is exactly what skill looks like, and the Brier (0.04) rewards it. The lesson: calibration alone is not skill; you also need the courage to differentiate (resolution). A forecaster who hedges everything to 50% can never be embarrassed and can never be useful.

Problem 6 (process scorecard). A learner describes a decision: "I found it on a screen (it's in my circle), did a full deep dive and reverse DCF, ran my checklist and resolved the one flag in writing, wrote a bear case and a base rate, set a 65% confidence and two kill criteria, and sized it at 4% with a verified −25% downside floor. Then it fell 20% and I'm annoyed." (a) Score it on the 8-dimension process scorecard. (b) It lost money. Where does it sit in the decision × outcome grid, and what should the learner do?

Solution. (a) Circle 2; full analysis 2; checklist+flag resolved 2; falsifiable thesis + bear + base rate 2; numeric confidence 2; margin-of-safety/reverse-DCF 2; sized per policy + real floor 2; kill criteria + monitoring 2 → 16/16. (b) A 20% fall is within the pre-identified −25% downside floor and there is no indication a kill criterion triggered, so this is a good decision with a (so far) bad, but survivable and anticipated, outcome (top-right of the grid). What to do: nothing driven by the price. Run the drawdown protocol (open the entry, check the kill criteria; if nothing structural broke, a 20% dip is a lower price for the same business), keep the process, and do not let annoyance ("resulting" on an unrealized loss) corrupt a 16/16 decision. If anything, a well-analyzed business 20% cheaper with the thesis intact is a candidate to add, subject to limits.

Problem 7 (skill vs luck 2×2). Classify each and say which process to keep or kill: (a) Full process, sized right, thesis played out, +50%. (b) Full process, sized right, a named-and-sized risk hit, −30%. (c) Tip, oversized, no journal, +120%. (d) Tip, oversized, no journal, −70%.

Solution. (a) Good decision, good outcome (deserved): keep the process; don't get overconfident. (b) Good decision, bad outcome (variance): keep the process; a well-sized bet on an identified risk is supposed to lose sometimes, and it was survivable because it was sized right; log a well-decided miss. (c) Bad decision, good outcome (dumb luck): kill the process; the win is a warning that this habit, repeated, delivers (d) eventually. (d) Bad decision, bad outcome (deserved): the obvious lesson; kill the process. The non-obvious, high-value calls are (b) keep-despite-loss and (c) kill-despite-win; getting those two right, against the pull of "resulting," is the whole skill.

Problem 8 (survival: forced liquidation). For each, state whether it carries a forced-liquidation (ruin) path and name the mechanism; then rank from safest to most dangerous: (a) buy quality equities with cash not needed for 10 years, no leverage; (b) the same, on 2× margin; (c) a fund of concentrated illiquid small-caps funded by monthly-redeemable client capital; (d) an NBFC funding 10-year loans with 90-day commercial paper; (e) a 3-month call option on your best thesis; (f) 55% of an unlevered portfolio in one name.

Solution. (a) No forced path: you choose when to sell; the only risk is being wrong on the merits (survivable, unlevered). (b) Yes, margin call: a drawdown forces selling into the fall to restore collateral. (c) Yes, redemption run: a drawdown triggers client redemptions, forcing sales of illiquid names at the bottom. (d) Yes, ALM run: short funding refuses to roll, forcing asset dumps to meet maturing CP (the IL&FS/DHFL mechanism). (e) Yes, the expiry clock: you can be right and still have the option expire worthless; the clock forces the outcome. (f) Yes, concentration ruin: one blow-up in a 55% position is a permanent, unrecoverable loss even with no leverage. Ranking, safest → most dangerous (roughly): (a) safe; then (f) [ruinous only if that one name blows up]; (e) [bounded loss but a hard clock]; (b) and (c) [drawdown-triggered forced selling]; (d) [the most classic and violent forced-liquidation structure, a run]. The lesson: only (a) is structurally sound; every other structure hands your sell-timing to someone or something else.

Problem 9 (timed, 8 min: drawdown protocol). You own a 5% position, now down 45%. Walk, in order, the exact steps your process prescribes before you may act, for both branches (thesis intact vs kill-criterion triggered). Then compute the gain needed to recover.

Solution. Steps, in order: (1) Do not act on the down-day: enforce the cooling-off rule; no sell unless a kill criterion has unambiguously triggered. (2) Open the original journal entry and read the kill criteria (field 7) and monitoring metrics (field 8) as written by your calm past self. (3) Ask: has anything structural broken, or is this price/sentiment? Check the monitoring metrics against current reality; did a load-bearing assumption fail or a kill criterion trigger? (4) Branch. Thesis intact (sentiment only): the drawdown is a lower price for the same business, so hold, or add if sizing/conviction warrant and no limit is breached; log the decision as a dated journal update. Kill criterion triggered (thesis broken): sell, because the reason you owned it is gone, and the recovery math is now irrelevant (holding to "get back to even" is the sunk-cost trap). Recovery math: to recover −45% you need L/(1−L) = 0.45/0.55 = 81.8%. The point: the price move never decides; the pre-written thesis test decides.

Problem 10 (timed, 10 min: write the mistakes section). A learner had, this quarter: a lender position where they under-weighted deposit-beta risk at entry and are now down; and a quality compounder they passed on 18 months ago (on trailing P/E, without a reverse DCF) that has since risen 40%. Write the "Mistakes and what I got wrong" section of their quarterly letter (§4.3 item 4), converting each into a rule or a circle revision.

Solution (one acceptable version). *"Two errors this quarter. (1) The lender: at entry I under-weighted deposit-beta sensitivity — my own-hand bear case named it, but my 70% confidence number did not reflect how binding it was; an honest entry would have been ~55% and a half-size starter. This is a calibration miss in a domain (spread lenders) where my own data already flags overconfidence — I should have applied the haircut and didn't. Corrective actions: (a) new checklist line — 'For any spread lender, stress deposit beta explicitly and let it set the confidence number'; (b) move 'lenders' from inside to edge in my written circle until the calibration data improves. (2) The one that got away: I passed on [Compounder] 18 months ago because it 'looked expensive' on trailing P/E — a lazy multiple instead of my own reverse-DCF process; it has since compounded ~40%. The error was a process failure (skipping my own valuation step), which happened to cost only opportunity this time. Corrective action: no pass on valuation grounds without a completed reverse DCF."* The grade is on honesty, specificity, and conversion to a rule — not on the P&L.

Problem 11 (timed synthesis, 10 min: reading two scoreboards). After one year a learner has: portfolio +18% (great); average process score 9/16 (mediocre); calibration shows 90%-bucket hits at 55%; two of the year's gains came from oversized tip-style bets that worked. Your friend says "18% — clearly the process is working, don't change a thing." Rebut, and state what you would actually change.

Solution. The friend is resulting at the portfolio level: reading a one-year P&L (dominated by luck, the paradox of skill) as a verdict on skill. The process scoreboard, which is what you control and what causes long-run returns, is flashing red: a 9/16 average process score means most decisions were sloppy; a 90%-bucket hitting 55% is severe overconfidence; and two big gains came from bad-process bets (oversized tips) that worked this time: dumb luck that, repeated, delivers the eventual oversized loss that ends the record. The +18% is not evidence the process works; it is largely evidence the year was lucky, and clinging to it would be learning exactly the wrong lesson. Changes: (i) enforce the full loop on every decision (the 9/16 says stages are being skipped, likely checklist and journaling); (ii) hard-cap position sizes and ban tip-driven entries (kill the bad-process habit while it is winning, which is the hardest and most important discipline); (iii) hard-code the confidence haircut and downsize high-confidence bets (fix the 90%→55% calibration). Optimize the process score; let the P&L be the lagging, noisy confirmation.

Problem 12 (synthesis: diagnose the broken loop). An investor does superb deep dives, keeps a journal, sizes carefully, and reviews quarterly, yet keeps getting surprised by theses that quietly broke months earlier, and finds their portfolio is really one big bet on "India consumer premiumization" despite owning fifteen names. Identify the two broken hand-offs and prescribe the fix.

Solution. Two seams are broken. (1) Monitoring (stage 7) is disconnected from the journal. The deep dives and journal entries are excellent, but nobody is checking the monitoring metrics and kill criteria (fields 7–8) against reality between decisions, so broken theses are discovered late. Fix: schedule the monitoring cadence written in each entry, and treat a tripped kill criterion as an event that forces a review, not something noticed by accident. (2) Portfolio fit (stage 5) was skipped or run superficially. Fifteen names that are all one factor ("India consumer premiumization") is the hidden-correlation trap (M9.02): thirty names, one bet. Fix: run the correlation/factor check at stage 5 for every addition, cap per-factor exposure, and recognize that diversification is by independent driver, not by ticker count. The tools are all strong; the hand-offs between analysis→monitoring and sizing→portfolio-fit are the failures, which is the central lesson here: a process is only as strong as its weakest coupling.

Problem 13 (guided: performance attribution). You ran an India equity book for the year and edged your benchmark. Decompose the excess return into allocation, selection, and interaction (allocation = (wₚ − w_b) × R_b; selection = w_b × (Rₚ − R_b); interaction = (wₚ − w_b) × (Rₚ − R_b), summed across sectors), then say in one line where your value-add, and your value-destruction, came from.

Solution. Benchmark = 0.20×15 + 0.50×5 + 0.30×10 = 8.5%; portfolio = 0.10×15 + 0.55×7 + 0.35×11 = 9.2%; excess = +0.7%. Allocation: Energy (−0.10)(15) = −1.5; Financials (0.05)(5) = +0.25; Tech (0.05)(10) = +0.5 → −0.75%. Selection: Energy 0.20(15−15) = 0; Financials 0.50(7−5) = +1.0; Tech 0.30(11−10) = +0.3 → +1.3%. Interaction: Energy 0; Financials 0.05×2 = +0.1; Tech 0.05×1 = +0.05 → +0.15%. Sum = −0.75 + 1.3 + 0.15 = +0.7% ✓ (reconciles to the excess). The one-line read: your stock selection added +1.3%, but your sector bets cost you −0.75%, chiefly by under-weighting Energy, which returned a strong 15% in the benchmark. Your edge is bottom-up selection; your top-down sector calls are a net drag, so the process should push you toward benchmark-like sector weights and let selection do the work.

Problem 14 (guided: IPS constraint check). Your written process document / IPS (§4.8) sets these hard limits: max single position 10%, max sector exposure 30%, minimum cash 5%, top-5 positions ≤ 50%, and no callable leverage (gross long ≤ 100%). A ₹1,00,00,000 (₹1 crore) portfolio holds: Alpha (Consumer) 12%, Beta (Consumer) 11%, Gamma (Consumer) 10%, Delta (Financials) 10%, Epsilon (Financials) 8%, Zeta (IT) 9%, Eta (IT) 7%, Theta (Pharma) 8%, Iota (Pharma) 5%, Kappa (Auto) 6%, Lambda (Auto) 4%, and 10% cash. Check every constraint, name each breach in rupees where relevant, and state the single trim that fixes the most breaches at once.

Solution. Positions sum to 90% + 10% cash = 100% ✓. Checking each limit: (1) Single position (10%): BREACHED by Alpha (12% = ₹12,00,000 vs the ₹10,00,000 cap → ₹2,00,000 over) and Beta (11% → ₹1,00,000 over); Gamma and Delta sit exactly at 10% (compliant). (2) Sector (30%): BREACHED by Consumer at 12 + 11 + 10 = 33% (₹33,00,000 vs ₹30,00,000 cap → ₹3,00,000 over); Financials 18%, IT 16%, Pharma 13%, Auto 10% all clear. (3) Top-5 (≤ 50%): BREACHED: the five largest are 12 + 11 + 10 + 10 + 9 = 52%. (4) Cash (≥ 5%): OK at 10%. (5) Leverage: OK, gross long is 90%, nothing borrowed. So three of five constraint categories are breached. The highest-leverage fix: trim the two Consumer names Alpha and Beta, because that one action simultaneously (a) clears both single-position breaches, (b) pulls Consumer back under the 30% sector cap, and (c) drops the top-5 under 50%. One disciplined trim, three constraints restored, which is exactly why the review runs the constraint check mechanically, as pass/fail, not by feel. (Note: 10% cash is not a violation even though it is on the high side, since the cash policy permits it as a residual of opportunity, §4.8 item 7.)

Problem 15 (independent: turnover and tax drag). A US taxable portfolio averages $1,000,000 over the year. During the year you buy $350,000 and sell $300,000 of securities; of the sales, $80,000 was realized gain, all short-term, taxed at 37%. (a) Compute the turnover ratio (standard fund definition: min(purchases, sales) ÷ average value). (b) Compute this year's tax and express it as a tax drag in percent of the portfolio. (c) What would this year's tax have been had you simply held?

Solution. (a) Turnover = min(350,000, 300,000) ÷ 1,000,000 = 300,000 ÷ 1,000,000 = 30%. (b) Tax = 37% × 80,000 = $29,600, which is 29,600 ÷ 1,000,000 = 2.96% of the portfolio, subtracted from this year's compounding base. (c) Had you held, the gain is unrealized, so this year's tax is $0; the $29,600 stays invested and compounds (deferral). The lesson: 30% turnover cost ~3.0% of the portfolio in tax this year alone, and because that drag recurs and compounds (WE6), turnover is not a free activity. It is a recurring tax on returns, which is why the review tracks it and why "don't sell a compounder without a reason" (stage 8) is arithmetic, not sentiment.

Problem 16 (independent: a full portfolio review computation). Your book began the year with these weights and produced these one-year returns: A 25% (+30%), B 20% (+10%), C 20% (−5%), D 15% (+15%), E 10% (−20%), Cash 10% (+6%). The benchmark returned 8.0%. (a) Compute each position's contribution to return (weight × return) and the total portfolio return. (b) State the excess versus the benchmark. (c) Identify the largest contributor and detractor, and compute the largest contributor's share of the total return. (d) What one review question does part (c) force?

Solution. (a) Contributions (percentage points): A 0.25×30 = +7.50; B 0.20×10 = +2.00; C 0.20×(−5) = −1.00; D 0.15×15 = +2.25; E 0.10×(−20) = −2.00; Cash 0.10×6 = +0.60. Total = +9.35%. (b) Excess = 9.35 − 8.00 = +1.35%. (c) Largest contributor A (+7.50 pp); largest detractor E (−2.00 pp). A alone produced 7.50 ÷ 9.35 = 80.2% of the entire year's return. (d) It forces the review's central honesty question: was that concentration of return skill or luck? One position drove four-fifths of the result, so A's journal entry and process score carry almost all the year's evidence (was A a high-process, well-sized decision, or an oversized bet that happened to work, per WE3's grid?), and the calibration on A's thesis is where the learning lives. Meanwhile E, the biggest drag, is the first name to run the drawdown protocol on (has a kill criterion tripped, or is the thesis intact?). A review that stops at "+9.35%, beat the benchmark" learns nothing; the contribution breakdown is where the learning starts.

Problem 17 (independent: the IPS required-return feasibility check). A retired US investor's IPS lists investable assets of $2,000,000, a first-year spending need of $80,000, expected inflation of 3%, a tax rate on returns of 15%, and a long horizon with capital to be preserved in real terms. (a) Compute the real spending rate. (b) Build the required nominal, pre-tax return (combine spending and inflation multiplicatively, then gross up for tax). (c) If the IPS's expected portfolio return is 8.0%, is the plan feasible, and if not, name three levers to fix it.

Solution. (a) Real spending rate = 80,000 ÷ 2,000,000 = 4.0%, the real return needed just to fund spending without touching real principal. (b) Nominal after-tax required = (1 + 0.04)(1 + 0.03) − 1 = 1.0712 − 1 = 7.12%; grossing up for the 15% tax on returns, pre-tax required = 7.12% ÷ (1 − 0.15) = 8.38%. (Additive shorthand: 4% + 3% ≈ 7% after-tax ≈ 8.24% pre-tax, close, and it makes the same point.) (c) The expected 8.0% is below the required 8.38%, so the plan is (marginally) infeasible: run as stated, it slowly erodes real purchasing power (a ~0.4-point annual shortfall). Three honest levers: cut the spending rate (e.g., to $75,000, a 3.75% real draw), lengthen the horizon or relax "preserve real capital" (permit a gentle principal drawdown), or raise the expected return by taking more risk (only if the IPS's risk section actually permits it). What you must not do is quietly assume a higher return to make the arithmetic close. That is the return objective bending to the wish, the exact self-deception a written IPS exists to prevent.

Problem 18 (timed, 12 min: the full quarterly review, synthesized). You are writing your Q-review (§4.3) for a $2,500,000 US book. The data: portfolio +6.0%, benchmark +4.0%; attribution allocation +0.4%, selection +1.4%, interaction +0.2%; turnover 15% with $50,000 of realized short-term gains taxed at 37%; and of the quarter's resolved predictions, 8 of 12 made at "80% confidence" came true. Compute (a) the excess return and confirm the attribution reconciles, naming the dominant effect; (b) the tax drag in percent; (c) the 80%-bucket hit rate and the calibration verdict; then (d) write the three-sentence conclusion that closes the loop.

Solution. (a) Excess = 6.0 − 4.0 = +2.0%; attribution sums to 0.4 + 1.4 + 0.2 = +2.0% ✓, and the dominant effect is selection (+1.4%): the quarter's edge was stock-picking, not sector tilts. (b) Tax = 37% × 50,000 = $18,500 → drag = 18,500 ÷ 2,500,000 = 0.74% in a single quarter (a meaningful bite if it recurs). (c) Hit rate = 8 ÷ 12 = 66.7% versus a stated 80%, which is overconfident at the top by ~13 points, the classic and most expensive pattern. (d) A model conclusion: "The +2% was almost entirely selection, so my edge stays bottom-up — I will keep sector exposure near benchmark and spend my effort on names. Turnover cost 0.74% in tax this quarter; I will let winners run unless a kill criterion trips, because that drag compounds. And my 80% calls came true only two-thirds of the time, so I am hard-coding the confidence haircut and downsizing high-conviction positions until the calibration curve earns the confidence back." The grade is on whether the numbers are right and whether each becomes a rule: attribution → where to spend effort, turnover → let winners run, calibration → haircut and downsize. That conversion is stage 9 feeding stage 1.

Applied mini-project: Phase 9 Capstone: your written investment process + a simulator season (the M10 milestone)

This is the capstone of Phase 9 and the M10 milestone (week 68). It has two parts and produces two durable artifacts you will run for the rest of your investing life: your complete written investment process document and one full simulator season of journaled, calibrated decisions scored on process, not P&L. Together they certify competency C9.

Part A: Write your complete investment process document (the operating manual). Author the ten-section document specified in §4.8, filled with your rules: specific, operational, and veto-shaped, not aspirational. Required sections: (1) philosophy & edge (three sentences); (2) circle of competence (inside/edge/outside, informed by your calibration-by-domain data where you have it); (3) sourcing criteria (universe + the exact first-cut screen); (4) analysis workflow (your M8 procedure); (5) the checklist (your Checklist v1, adopted from M6.05); (6) sizing rules (the conviction→size function + hard limits); (7) portfolio construction rules (concentration policy, the marginal-idea hurdle, correlation limits, cash policy); (8) sell rules (the five reasons + the compounder rule + kill-criteria-govern-drawdowns); (9) review cadence (quarterly letter, calibration + process scoring, checklist-growth rule, annual re-underwrite); (10) anti-goals (what you will not do). The test of every line: could a stranger execute it, and does it ever say no?

Part B: Run one full simulator season (≥10 journaled, calibrated decisions). In the app's portfolio simulator (or a paper portfolio if offline), run one accelerated season through a seeded market cycle (e.g., a dot-com, 2008, 2020, or high-inflation regime; pick one that will test your temperament, not a calm bull). The protocol:

1. Every decision (buy/add/trim/sell/pass) gets a full journal entry in the §4.2 expanded template, written before the outcome, including a numeric confidence and pre-committed kill criteria. Minimum ten journaled decisions across the season.
2. Run the whole loop on each: source → analyze → checklist → size (within your Part-A limits) → portfolio-fit → decide+journal → monitor → sell by your rules.
3. Live through the seeded drawdown by protocol (§4.7): when the crash hits, do not improvise. Run the drawdown protocol, consult your kill criteria, and log every decision. Surviving a bear market on-thesis is the point of the exercise.
4. At season end, close the loop: (a) resolve your predictions and score your calibration (bucket table + reliability read + Brier); (b) score each decision on the process scorecard (8 dimensions, 0/1/2), blind to P&L; (c) write a season-end review letter (§4.3 format), with an honest mistakes section; (d) harvest at least one new checklist line from your logged mistakes.

The process-scoring rubric for the season (this is how the season is graded: on process, never on the money it made or lost):

Overall capstone scoring (0 = absent/wrong · 1 = present but shallow · 2 = rigorous and specific).

Pass = ≥13/16 (≥80%), with the journal-discipline and survived-on-thesis criteria both at 2. File the process document and the season's journal, calibration record, and review letter in your knowledge system (M0.04). This is the artifact you carry into Phase 10, where you run this exact process, unaided, on two real companies. Certifies C9.

Reading & resources

Core (do these).

• Howard Marks, The Most Important Thing (2011), and the Oaktree memos (oaktreecapital.com/insights, free). The memos are the living example of the quarterly-letter discipline and Marks's own statement of why he writes (writing forces and tests thinking); read "You Bet!" (probabilistic thinking, decision quality vs luck), "Risk," and "The Illusion of Knowledge." The spine of §4.3 and §4.5. [Book Paid / Memos Free] [Beginner→Intermediate]
• Warren Buffett: the Berkshire Hathaway shareholder letters (berkshirehathaway.com, free). Decades of the quarterly/annual review letter done at the highest level: businesses explained plainly, mistakes owned by name, process made visible. Read three or four consecutive years to feel the format and the honesty of the mistakes sections. [Free] [Beginner]
• Annie Duke, Thinking in Bets (2018) and How to Decide (2020). The clearest treatment of process-over-outcome ("resulting"), decision journals, and thinking in probabilities; How to Decide is a practical workbook for exactly the journaling and calibration habits here. [Paid] [Beginner]
• Philip Tetlock & Dan Gardner, Superforecasting (2015). The empirical science of calibration and the Brier score, and the proof that forecasting is a trainable process of keeping score and adjusting, the backbone of §4.4. [Paid] [Intermediate]

Deepening.

• **Michael Mauboussin, *The Success Equation: Untangling Skill and Luck (2012)**: the "can you lose on purpose?" test and the paradox of skill (§4.5); pair with his free Counterpoint Global Base Rate Book for the outside-view inputs the journal's field 4 needs. [Book Paid / Base Rate Book Free] [Intermediate]
• Lee Freeman-Shor, The Art of Execution (2015): the evidence on selling (running winners, cutting broken theses) behind the loop's stage 8; a fast, data-driven read (also in M9.03). [Paid] [Beginner]
• Edward Thorp, A Man for All Markets (2017): Kelly sizing, edge, and a lifetime of survival-first discipline from someone who never blew up; the human face of §4.6 (also in M9.01). [Paid] [Beginner→Intermediate]
• Charles Ellis, Winning the Loser's Game (8th ed.): investing as a game won by avoiding unforced errors, not by heroics; the philosophical case for optimizing a boring, repeatable process over the one great call. [Paid] [Beginner]
• Peter Bernstein, Against the Gods: The Remarkable Story of Risk (1996): the history of how humanity learned to measure risk and probability; the intellectual backdrop to calibration and expected value. [Paid] [Intermediate]
• Nassim Taleb, Fooled by Randomness and Skin in the Game: luck vs skill in markets, and the ergodicity/ruin argument under §4.6; read for the temperament to distrust short-run track records. [Paid] [Intermediate→Advanced]
• Atul Gawande, The Checklist Manifesto (2009): revisited from M6.05 as the pre-commitment logic behind the written process document, not just the checklist. [Paid] [Beginner]

Free / durable.

• CFA Level III: the portfolio-management and Investment Policy Statement (IPS) readings. The institutional version of your written process document; the IPS structure (objectives, constraints, governance) maps closely onto §4.8 and is worth reading for the discipline of codifying a mandate in writing. [Free curriculum outlines / Paid full] [Intermediate]
• Farnam Street (fs.blog), free. The widely used decision-journal template and essays on calibration, process, and skill vs luck; a good ongoing companion to the loop. [Free] [Beginner]
• Sanjay Bakshi: Behavioural Finance & Business Valuation (BFBV) notes and the Fundoo Professor blog, free. The India-specific treatment of process, temperament, journaling, and base rates applied to Indian companies. [Free] [Intermediate]
• Gautam Baid, The Joys of Compounding (2020): the Indian synthesis of philosophy, temperament, the personal knowledge/decision-journal system, and running a process for the long term; ideal for this phase. [Paid] [Beginner]
• Going deeper: Carl R. Bacon, Practical Portfolio Performance Measurement and Attribution (2nd ed.). The practitioner's standard behind WE5's attribution arithmetic: allocation vs selection vs interaction, the arithmetic-versus-geometric debate, and how single-period effects are linked across a multi-period review; the rigorous version of "was my edge the sector calls or the stock-picking?" [Paid] [Advanced]
• Going deeper: Bogleheads wiki, "Tax-efficient fund placement," "Turnover," and "Tax-adjusted asset allocation" (bogleheads.org, free). A clear, free treatment of WE6's turnover/tax-drag arithmetic: why deferral behaves like an interest-free loan and why low turnover compounds into materially higher after-tax wealth; the quantitative backing for stage 8's "don't sell compounders early." [Free] [Beginner→Intermediate]

Program companions. M9.01 (sizing, Kelly, ruin: the sizing/survival stages), M9.02 (portfolio construction: the portfolio-fit stage), M9.03 (selling, mistake taxonomy: the sell/review stages); M6.05 (the checklist, journal, and calibration concepts this module wires into a loop); M6.03 (the decision science the loop enforces); M0.04 (the original journal template §4.2 expands); Phase 8 (the analysis stage); and Phase 10 (where this process is run, unaided, on real companies).

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 nine-stage loop, the decision journal as a pre-outcome fossil, quarterly review letters, calibration and the Brier score, the process scorecard and the decision x outcome grid, survival discipline, the drawdown protocol, and the written process document",
  "use": "The whole module is built on a record that must not be contaminated, so the rule for AI here is strict and worth stating before the uses: it may read the journal, it may never write in it, and it may never see the outcome while grading the process. Within that: (1) OUTCOME-BLIND PROCESS SCORING. Hand it a journal entry with the ticker, the dates and every price redacted, and ask it to score the eight rubric rows of Section 4.5 with the sentence it scored each row on. Redaction is the whole method - the scorer that knows the stock tripled cannot score the process, and neither can you. (2) EXTRACTING THE RESOLVED PREDICTIONS. Calibration dies from the friction of finding, in three years of prose, every sentence that contained a probability and a resolvable claim; that extraction is mechanical and a model does it well. (3) QUARTERLY LETTER FIRST DRAFT from the position data and the journal, so the writing time goes into the thinking rather than the assembly. (4) PATTERN SEARCH ACROSS ENTRIES: 'which pillar type has broken most often in my closed positions?' is a question about your own record that you cannot answer by memory, because memory is the thing this module says is compromised.",
  "tools": ["Chat assistants — Claude, ChatGPT for redacted process scoring, prediction extraction and the letter draft", "Coding assistants — for the calibration curve, the Brier decomposition, the Brinson identity and the regression attribution below", "Document/PDF AI — for pulling the resolved facts out of filings that settle a journal prediction, so resolution does not depend on your recollection of what happened"],
  "prompt": "Below is a decision-journal entry with the company name, the dates and all prices removed. Score it 0/1/2 on each of these eight process dimensions <PASTE RUBRIC>, and for every score quote the exact sentence from the entry you scored it on. If a dimension is absent from the entry, score 0 and say 'absent' - do not infer it from the quality of the rest. Do not guess what the company is. Do not comment on whether the thesis was right. Separately, list every sentence in the entry that contains a probability and a claim that could later resolve true or false, with the probability and the claim stated as a pair.",
  "verify": "The journal is a fossil and its value is that it cannot be edited, so the single rule that outranks every use above: the model's output goes in a separate file, never into the entry. An AI-improved journal entry is not a record of what you thought, it is a record of what you and a model thought later, and it destroys the only artefact in this program that hindsight bias cannot reach. Redaction has to be verified, not assumed - a model that infers the company from the sector and the year is scoring with the outcome in hand, so spot-check by asking it to name the company and confirming it cannot. Score a sample of entries yourself and compare; a scorer that agrees with you on the entries you scored generously is flattering you. And never accept a bad-luck verdict or a probability from it: Section 4.5's off-diagonal cells are where all the learning is, and they are exactly the cells a fluent, agreeable reader will smooth over.",
  "diy": "The gate is unaided: you write a full journal entry with a falsifiable thesis, a numeric confidence and pre-committed kill criteria; you compute a Brier score and read a reliability curve; you score a decision on the sixteen-point rubric and place it in the decision x outcome grid; and you run the Brinson identity by hand and reconcile the three effects to the excess return. No tool is in that room."
}

Modern Data Analysis

By hand first. You computed a Brier score, read a reliability curve for the overconfidence signature at the top of the scale, scored two decisions on the sixteen-point rubric and placed them in the grid, decomposed an excess return into allocation, selection and interaction that reconciled exactly, and worked the turnover tax drag to a 1.88-point annual gap. Keep all of it. The Brinson reconciliation in particular is a check you should be able to run in your head, because an attribution that does not sum to the excess return is wrong and code will not tell you so.

Today's workflow. Three things change when the record lives in a data structure rather than in a notebook.

• Calibration becomes a curve you watch fall, split by domain. §4.4's most useful refinement, bucketing by domain rather than just by confidence, is arithmetically trivial and organisationally impossible on paper, because it requires every prediction to carry a tag and a resolution date. A journal with structured fields gives it to you for free.
• Attribution gets a second, risk-adjusted layer. WE5's Brinson identity answers "allocation or selection?" It does not answer the question §4.5 actually cares about, which is "was any of it skill?" The lens below shows that the selection effect can be entirely manufactured by beta.
• And the sample-size question becomes answerable. §4.5 says the P&L is dominated by luck over short horizons and that "long enough" is many years. How many? That is a power calculation, and it has an exact answer that ought to change how you weight your own evidence.

Tools & sources (IN + US). pandas for the journal as a dataframe with one row per prediction and one per decision, numpy and statsmodels for the attribution regression and its standard errors, matplotlib for the reliability curve and the rolling Brier. India: NSE/BSE bhavcopy and the NSE sectoral indices (Nifty Bank, Nifty IT, Nifty FMCG, Nifty Auto, Nifty Metal) as the benchmark sector returns the Brinson identity needs, the Nifty 50 / Nifty 500 TRI as the market series for the regression (use total-return, not price, or your alpha is a dividend yield), NSDL/CDSL statements and broker contract notes for the actual weights and dates, RBI's 91-day T-bill rate for the risk-free leg, and quarterly results plus concall transcripts to resolve journal predictions against filed fact rather than recollection. US: the Ken French Data Library for the market, size, value, profitability, investment and momentum series that turn a one-factor alpha into a multi-factor one; FRED (DGS3MO) for the risk-free leg; EDGAR 8-K and 10-Q for resolution evidence; the GIPS standards (CFA Institute, free) for how a track record must be presented if it is ever shown to anyone else; Mauboussin's Success Equation and his Counterpoint Global notes on luck-versus-skill for the framework §4.5 is built on.

# Two questions the module poses and cannot settle on paper. R3, recomputed in-session.
import numpy as np
rng = np.random.default_rng(904)
S, Q, BETA, rf = 3, 40, 1.25, 0.015                 # 3 sectors, 40 quarters, quarterly rf
bw = np.array([.40, .30, .30])                       # NO sector bets: allocation is 0 by design
sel_m, alph, tstat = [], [], []
for _ in range(4000):
    Rb = rng.normal(0.025, 0.09, (Q, S))             # benchmark sector returns
    Rp = rf + BETA*(Rb - rf) + rng.normal(0, 0.03, (Q, S))   # ZERO stock-specific alpha
    sel_m.append((bw*(Rp - Rb)).sum(1).mean())       # Brinson selection effect
    x = (bw*Rb).sum(1) - rf; y = (bw*Rp).sum(1) - rf
    X = np.c_[np.ones(Q), x]; b, *_ = np.linalg.lstsq(X, y, rcond=None)
    r = y - X@b; se = np.sqrt((r@r)/(Q-2) * np.linalg.inv(X.T@X)[0, 0])
    alph.append(b[0]); tstat.append(b[0]/se)
print(np.mean(sel_m), np.mean(alph), np.median(tstat))
# +0.00250/qtr (+1.00%/yr) "selection"   vs   alpha -0.00003/qtr, median t = 0.000

# how long before a REAL edge clears t = 1.96?
for IR in (0.2, 0.3, 0.5, 0.75, 1.0):
    print(IR, round((1.959964/IR)**2, 1), round(((1.959964+0.841621)/IR)**2, 1))   # 95% sig; 80% power
# 0.20 96.0 196.2 | 0.30 42.7 87.2 | 0.50 15.4 31.4 | 0.75 6.8 14.0 | 1.00 3.8 7.8  (years)

Verify. Every worked figure reproduces. §4.4's miniature Brier is 0.22 from squared errors of 0.01, 0.49 and 0.16, with the confident-wrong call contributing 0.81 against an honest 50/50's 0.25. WE5: benchmark 7.9%, portfolio 9.9%, excess +2.0%; allocation +1.0 / −0.5 / 0.0 = +0.5%, selection +0.8 / +0.3 / +0.3 = +1.4%, interaction +0.2 / −0.1 / 0.0 = +0.1%, summing to exactly +2.0%, and on ₹50,00,000 that is ₹70,000 / ₹25,000 / ₹5,000. WE6: 1.08²⁰ = 4.6610, after-tax 3.9288× for a 7.081% CAGR, against Restless's 1.052²⁰ = 2.7562× at 5.20%, a 42.54% terminal-wealth gap, $3,928,766 versus $2,756,226 on a $1m start, a $1,172,540 difference and 1.881 percentage points a year. Practice 18: excess +2.0% with the three effects summing to +2.0%, tax drag 0.74%, hit rate 66.7%. Practice 13's contribution set totals +9.35% for an excess of +1.35%, with position A supplying 80.2% of the year's return.

Quantitative lens

The selection effect can be entirely manufactured by beta, and the Brinson identity cannot see it. This is the sharpest thing in the addendum, because WE5's conclusion ("the story is selection, not allocation; your edge is bottom-up stock-picking") is exactly the conclusion a book with no stock-picking edge at all will also produce. Simulate a manager who takes no sector bets whatsoever (portfolio weights identical to benchmark weights, so the allocation effect is zero by construction) and holds, within every sector, a portfolio with beta 1.25 and zero stock-specific alpha. Over forty quarters the Brinson selection effect averages +0.250% per quarter, or +1.00% a year, and it will be written up in the quarterly letter as stock-picking. Split it by market direction and the mechanism is naked: +1.137%/quarter when the market rose, −0.957%/quarter when it fell. It is not selection. It is leverage on the sector return, and the Brinson identity classifies it as selection because the identity's within-sector term makes no risk adjustment at all. Run the regression instead, portfolio excess return on benchmark excess return, and the same records return a mean alpha of −0.003% per quarter with a median t-statistic of 0.000, and only 2.6% of the ten-year records clear t > 1.96, which is the false-positive rate you should expect and not a whisper more. The practical instruction is not to abandon Brinson, which answers the question it was built for. It is that §4.5's question (was it skill?) needs a risk adjustment that WE5 does not contain, and that a selection effect which is large in up quarters and negative in down quarters has already told you what it is before you run any regression at all.

How long is "long enough": the number §4.5 declines to give. The claim above is that skill surfaces in results only over "many years, far longer than the interval at which you need feedback." The exact statement is that the t-statistic on a track record grows as IR × √T, so the years needed to clear t = 1.96 are (1.96/IR)². At an information ratio of 1.00, exceptional and sustained by almost nobody, that is 3.8 years for bare significance and 7.8 years for 80% power. At 0.75, 6.8 and 14.0 years. At 0.50, a genuinely good manager, 15.4 and 31.4 years. At 0.30, 42.7 and 87.2 years. At 0.20, 96 and 196 years. Read the third row again: a manager with real, sustained skill at an IR of 0.5 needs a career to prove it at conventional confidence, and needs two careers for the test to be reliably able to find it. The same arithmetic on a simulated ten-year quarterly record: a true alpha of 3%/yr clears t > 1.96 in 76% of ten-year records, 2%/yr in 41%, and 1%/yr in 16%, which is to say a genuine one-point edge is, over a decade, statistically invisible five times out of six. And on individual decisions the picture is the same: distinguishing a true 55% hit rate from a coin flip needs 384 resolved decisions at 95% confidence and 782 for 80% power; a 60% hit rate needs 96 and 194. At the pace of a concentrated fundamental investor, that is decades of decisions. This is the quantitative content of the paradox of skill, and it is the reason §4.5's process scorecard is not a consolation prize.

Which is why calibration is the efficient scoreboard, and now it has a number. A perfectly calibrated forecaster whose confidences spread across 50–95% scores an expected Brier of 0.1825, well below the 0.25 coin-flip benchmark §4.4 names. How many resolved predictions before that gap is statistically visible? 30 predictions gets you there 50% of the time, 50 gets 67%, and 100 gets 90%. A hundred resolved predictions is two or three years of a journal that logs three or four resolvable claims a month, against the ten to thirty years the return series needs for a comparable read on a comparable edge. The advantage is real and it is roughly a factor of three to five, not a factor of a hundred, but it points the same way §4.4 does, and for the same reason: a prediction resolves against a fact, while a return resolves against a market. Calibration is the faster scoreboard because it asks a question with less noise in it, not because it asks an easier one.

Honest limits, and they matter here more than anywhere in Phase 9 because the numbers above are about how much you are allowed to believe. Every power calculation assumes the returns are i.i.d. and roughly normal, and investment returns are neither. They are autocorrelated, fat-tailed and regime-dependent, all of which lengthen the required horizon, so the years quoted are floors rather than estimates. The regression attribution corrects for one factor; a book that is long quality and short value has a "skill" measurement that is really a factor tilt until you regress on those factors too (M9.02 §4.6), and each added factor spends degrees of freedom you do not have. The information ratios above are inputs, not measurements: nobody knows their own IR, and the circularity is the point: you would need the track record to estimate the IR you would need to evaluate the track record. The beta-contamination result uses a single fixed beta of 1.25 and a single benchmark; a book whose beta drifts with the cycle produces a messier and generally more flattering picture. And the Brier calculation assumes predictions resolve independently, which theses in a concentrated book emphatically do not, since five predictions about the same credit cycle are close to one prediction, so the effective sample size behind a calibration curve is smaller than its row count, exactly as M9.01 §4.10's N_eff says of a portfolio. Nothing above rescues a short record. What it establishes is narrower and more useful: the process scoreboard is not a substitute for evidence you could have had faster. It is the only evidence available on the timescale at which learning happens.

Do it in code: structure the journal so every prediction carries a probability, a domain tag and a resolution date; plot the reliability curve by domain, not just by level; run the Brinson identity and the regression, and treat any gap between them as beta until proven otherwise; and before you conclude anything from a track record, compute the t-statistic it would need and the years it would take. See G1/AA1.03, G1/QM1.02 and G2/QD2.03.

Where this goes next: galaxy cross-links

• G1 · AA1.03 Performance Attribution & GIPS: the institutional standard behind WE5: multi-period linking, arithmetic versus geometric attribution, and the presentation rules that exist because a track record is the easiest thing in finance to present misleadingly.
• G2 · QD2.03 Performance Attribution in Code: the same decomposition as a reproducible pipeline, including the risk-adjusted layer this lens shows the Brinson identity cannot supply.
• G1 · QM1.02 Hypothesis Testing & Regression: the t-statistic on alpha, what its standard error depends on, and why a multi-factor regression on a short record spends degrees of freedom faster than it buys insight.
• G1 · QM1.01 Probability, Distributions & Estimation: power and sample size, which is the entire content of "how long before skill is distinguishable from noise," plus what a Brier score is as an estimator.
• G2 · QD2.04 Reproducible Research & Deployment: the journal, the calibration curve and the quarterly letter as artefacts that regenerate from named inputs, which is what stops a decision record from quietly becoming a story about the past.

Teach it back & journal

Feynman prompt. Write a one-page explainer for a smart 15-year-old titled "How a serious investor runs the same loop for forty years — and why they grade themselves on the plan, not the score." You must convey, without jargon: that good investing is not a series of clever guesses but one repeatable routine (find, study, check, size, decide, write it down, watch, sell, review) run over and over, where the review is what makes the next lap better than the last; that you must write down what you believe and how sure you are before you find out what happens, because otherwise your memory quietly rewrites the past to make you feel smart, and then you can never learn; that the only fair way to judge a decision in a game full of luck is by whether the thinking and the bet size were sensible, not by whether that one time it made money (a good plan can lose, a bad plan can win); and that the one unbreakable rule is never let anything force you to sell at the bottom, no borrowed money and no promises you can't keep, because you can only keep winning if you stay in the game. If your page needs the words "calibration," "Brier," or "ergodicity," you don't own it yet.

Journal prompt. Three entries, and keep them: these are the seeds of the process you will run for decades. (1) Your philosophy and edge, in three sentences (the opening of your process document): what kind of investor you are, and honestly where your edge comes from: behavioral, analytical, or (rarely and probably not) informational. Just as important, say where you have no edge and will not pretend to. (2) Your anti-goals. Write the five things you will absolutely not do, decided now while you are calm: the leverage you won't take, the tips you won't act on, the sizes you won't exceed, the panic-sell you won't make without a triggered kill criterion. In a future drawdown you will consult this list instead of your fear. (3) Your calibration prediction. Before you run your simulator season, predict, with a number, whether your resulting calibration curve will show you overconfident, underconfident, or well calibrated, and in which domain you expect to be worst. Then run the season and score it. Grading your own past self on that prediction, honestly, is the whole discipline compressed into a single act.

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