This is the investment framework I use to decide what to own. I have written the rules down in enough detail for them to be wrong, and I have applied them to the 200 largest companies in the S&P 500. I publish the hard cases and the boundary calls with the verdicts.
The decision hierarchy
Every verdict comes from the same ordered sequence. A name that fails an early stage cannot be rescued by a later one.
Stage 0: the conduct veto
One thing outranks price and quality. A company that obstructs a safety or investigative process in documented and active ways is withholding the evidence a regulator or court needs to adjudicate harm. I don't invest in these companies when I am aware of such instances. Tesla's record meets that test, so I rate it AVOID on conduct alone, independent of valuation. Tesla filed delayed and incomplete crash reports to NHTSA under the Standing General Order, and those filings are now the subject of a federal audit query. A U.S. Senate letter to the regulator documented a practice, unique among reporting automakers, of redacting crash details as confidential business information. Trial reporting in a fatal-crash lawsuit that Tesla lost in 2025 documented the withholding of Autopilot crash data.
The veto covers active obstruction. I give a prominent flag in the rationale to misconduct that has been surfaced, litigated, penalized, and remediated. I flag Wells Fargo's fake accounts, J&J's talc record, Exxon's climate-science concealment, GM's concealed ignition-switch defect, and 3M's decades of internal PFAS knowledge under this rule. I can price a harm after it has surfaced. I cannot price a harm while a company hides its evidence. Cummins is the hardest test of this boundary. Cummins engineered emissions defeat devices to deceive a regulator's test, and I do not apply the veto because the conduct was caught, penalized at near-record scale at $1.675B, and ended.
The veto covers obstruction. Litigation alone leaves a company outside it. Nearly every mega-cap carries a litigation record, and a veto that fired on settlements would exclude most of the companies in the index.
The veto covers evidence. Reputational distaste and polarizing founders fall outside it.
Stage 1: the harm-exclusion test
A business is excluded on harm grounds only when it meets all three prongs. (a) Certain, prevalent harm from intended use: the product seriously harms a large fraction of ordinary users used exactly as intended, at closed-science certainty. (b) Involuntary or socialized harm: a material share falls on non-consenting bystanders or the public. (c) Concealment or obstruction conduct: the industry's posture toward the harm is concealment. The prongs are conjunctive.
Tobacco meets all three. Tobacco causes roughly 480,000 US deaths a year from intended use and more than 41,000 non-smoker deaths from secondhand exposure, and it carries roughly $600B in annual socialized costs. The industry concealed that record for decades. I exclude Altria and Philip Morris at any price, and I include their smoke-free transition in that exclusion. I grant no exemptions for execution or returns.
Robinhood is the case where I decline to exclude under this test. The allegations against Robinhood follow a familiar pattern. Critics accuse its engagement-engineered product, gamified trading, of harming its own users. No prong binds in full. Prong (a) fails because the behavioral-finance evidence on gamified trading is contested and concentrated in susceptible subgroups, and most users trade without serious injury. Prong (b) fails because trading losses fall on the trader who chose to trade, and no bystander bears an equivalent of secondhand smoke. Prong (c) does not bind because Robinhood's conduct record, a then-record FINRA penalty over supervision failures and misleading communications plus state actions over gamification, was caught and penalized under regulatory supervision. Tobacco's record was decades of coordinated concealment. I therefore pass Robinhood to Stage 2 like everything else, and I rate it HOLD there on valuation. I treat the flags in its rationale as sell triggers if the record worsens.
Sugar fails prong (a). Harm from sugar is dose-dependent. A serving of Coca-Cola can sit inside a health-neutral dose, and a cigarette has no health-neutral dose.
I flag fossil fuels and do not exclude them. Prong (b) binds, and prong (c) binds historically. Fossil-fuel harm is aggregate and planetary, and it lacks the high-certainty individual kill rate that tobacco inflicts on its users. That boundary is a judgment about values. The exclusion also binds at the level of primary economics: it removes manufacturers, and it leaves general retailers who incidentally stock cigarettes in place.
Stage 1: the defense threshold
I exclude weapons and defense companies where defense is their primary business: RTX, Lockheed Martin, General Dynamics, Northrop Grumman, and Palantir. Palantir sells software, and its revenue is majority government work including targeting and battlefield AI, so its primary economics are defense and I exclude it as well. I flag minority exposure at Boeing, GE Aerospace, Honeywell, Howmet, and TransDigm. At mega-cap scale, trace defense exposure is nearly universal, so I record it with a flag without disqualifying the company.
Stage 2: valuation, the primary screen
(1) I read multiples against margin structure and durability; raw ratios in isolation mislead. NVIDIA at roughly 25x sales on extraordinary margins is a coherent and demanding price. Pre-veto Tesla at roughly 16x sales on automaker margins rested on narrative. (2) I never buy cyclicals at cycle-high earnings. The current test cases are the AI-adjacent names trading far above their own histories: Ciena at roughly 14.5x sales against a 2 to 3x historical norm, Lumentum at roughly 25x, and Comfort Systems at roughly 6.4x sales for a mechanical contractor. Demand for all three is real. The entries are indefensible at these prices, and all three are AVOIDs. (3) A low multiple can hide unresolved existential risk. Intel carries an unproven foundry bet, and Pfizer carries a patent cliff. I read their low multiples as traps until those risks resolve. (4) A verdict allocates capital. Where one thesis has several expressions, the BUY concentrates in the premier expression. I give a second name its own BUY only when it independently clears every stage, as Ross does beside TJX and Progressive does beside Chubb. (5) Close multiples can carry opposite verdicts. Monolithic Power at roughly 24.7x holds while Lumentum at roughly 25x is avoided. The deciding variable is the durability and quality of the earnings behind the multiple. (6) Every BUY carries a price ceiling above which it decays to HOLD. Every HOLD carries a conversion level. For event-driven names, I state that no price alone converts them, because a cheaper price on an unproven turnaround increases the loss when the turnaround fails. (7) For compounders, the sell trigger is thesis-break. For cyclicals, price is the sell trigger. For flagged names, the conduct conditions double as sell triggers.
Stages 3 and 4: resonance, founders, stewards
Among fairly priced businesses, I prefer category-defining products built with trust and craft that I use and evangelize without irony. The founder factor splits in two, and neither score credits bluster. Conviction authenticity asks whether I believe the person. Organizational capability asks whether the machine ships regardless. By scoring the two separately, I arrive at my verdict on Tesla: high capability, zero conviction credit, conduct veto. I read stewardship provenance against lifecycle. Insider-stewards help compounding companies and harm declining ones. Outsider change-agents run the reverse pattern, and Niccol at Starbucks and Ortberg at Boeing are the correct provenance for a broken franchise.
I buy trusted, crafted, cash-generating businesses at prices that leave a margin of safety. I prefer products I would evangelize. I state my exclusions as tests, I name every known flag, and I write down the conditions for changing my mind in advance.
The verdicts
The list covers 200 listings and 199 distinct businesses, counting both Alphabet share classes. The first hundred listings contain 27 BUY, 61 HOLD, and 12 AVOID. The second hundred contain 10 BUY, 85 HOLD, and 5 AVOID.
The 37 BUYs
| Ticker | Company | Thesis in one line |
|---|---|---|
| GOOGL / GOOG | Alphabet | The definitional resonant default at a discount to mega-cap peers. Antitrust ruling named and monitored, with no veto. Verdict re-underwritten July 2026 for the AI-capex transition. |
| MSFT | Microsoft | Three interlocking enterprise franchises at elite margins; the AI hedge and the AI bet at once. |
| AMZN | Amazon | A logistics-and-compute utility priced on a retailer's income statement. The accounting shadow is the mispricing. |
| META | Meta | Consumer-internet scale at the group's most modest multiple of cash generation. Disclosure: my employer, listed under the same rules as every other name, without further commentary. |
| BRK.B | Berkshire Hathaway | The equity that gets stronger in a crash. The culture is the asset. |
| JPM | JPMorgan Chase | Scale leadership in everything it does. Fortress capital practiced as strategy. |
| V | Visa | Basis points on global consumption at ~80% margins; the toll-bridge archetype. |
| MA | Mastercard | The duopoly's second seat, independently clearing on identical economics. |
| JNJ | Johnson & Johnson | AAA-grade diversified healthcare at a modest multiple; talc record named in full. |
| ABBV | AbbVie | Replaced a Humira-sized cliff in real time. Demonstrated execution at an unheroic price. |
| MRK | Merck | Paid to underwrite a known cliff: the Keytruda fear is already in the price. |
| LIN | Linde | The most reliable compounding machine in industrials; take-or-pay oligopoly economics. |
| WFC | Wells Fargo | Post-turnaround catalyst (asset cap lifted); fake-accounts record named in full. |
| AXP | American Express | A payments company that is also an aspirational brand; closed-loop economics. |
| BLK | BlackRock | Toll collector on indexing itself; flows compound faster than fees compress. |
| TJX | TJX Companies | The buying organization is the moat; counter-cyclical treasure-hunt economics. |
| TMO | Thermo Fisher | Picks-and-shovels of medicine at a cycle-trough-flavored price. |
| ISRG | Intuitive Surgical | The switching cost is surgeons' hands; the one name where quality earns a premium entry. |
| GLW | Corning | Materials-science craft with an AI optical inflection the market long ignored. |
| ABT | Abbott | Diversified medtech tortoise with a genuine consumer product (Libre) as the engine. |
| UBER | Uber | The network-effects survivor of mobility, finally cash-generative. The most on-thesis name in public markets. |
| CB | Chubb | Underwriting craft institutionalized; the premier P&C expression. |
| SPGI | S&P Global | Four toll bridges in one company. |
| BKNG | Booking Holdings | The demand aggregator of global travel; the bear case has lost for fifteen years. |
| SYK | Stryker | Medtech's most consistent compounder; Mako builds an installed-base annuity. |
| PGR | Progressive | A data company selling insurance; earns the second P&C BUY on its own edge. |
| DHR | Danaher | The operating-system acquirer of life sciences at a post-hangover fair price. |
| INTU | Intuit | Category-defining defaults priced for AI disruption; FTC flag named. |
| CME | CME Group | A monopoly on derivatives liquidity that converts volatility into revenue. |
| ADBE | Adobe | The clearest quality-at-a-discount case in software; the disruption thesis is the risk being paid for. |
| ICE | Intercontinental Exchange | Exchange, data, and mortgage rails. Half the revenue recurs regardless of volumes. |
| MRSH | Marsh & McLennan | The toll-bridge principle applied to risk itself. |
| ABNB | Airbnb | Founder-led, category-defining, and financially mature. Clears every stage at once. |
| MCO | Moody's | The ratings duopoly's second seat with a growing analytics annuity. |
| CRH | CRH | What cities are made of, at a reasonable price. The built-environment core holding. |
| ROST | Ross Stores | The off-price thesis, independently cleared: same moat, distinct customer. |
The 17 AVOIDs, by ground
| Ticker | Company | Ground |
|---|---|---|
| TSLA | Tesla | Conduct veto: delayed and redacted crash reporting to NHTSA (under federal audit query) and court-documented data withholding in a fatal-crash suit. No price converts this. |
| MO / PM | Altria · Philip Morris | Harm-exclusion test, all three prongs. |
| RTX / LMT / GD / NOC | RTX · Lockheed · Gen. Dynamics · Northrop | Defense primes: categorical exclusion. |
| PLTR | Palantir | Defense-primary economics, and the most extreme multiple in the two hundred (~62x sales). Two independent grounds. |
| INTC | Intel | Value trap: cheapness reflects an unproven foundry bet. Re-entry on customers committing. |
| BA | Boeing | Cash-burning turnaround already priced for recovery; re-entry on sustained free cash flow. |
| SNDK / STX / WDC | Sandisk · Seagate · Western Digital | Storage cyclicals capitalized at cycle-peak earnings. |
| APP | AppLovin | Extreme multiple atop named earnings-quality questions. |
| CIEN / LITE / FIX | Ciena · Lumentum · Comfort Systems | AI-adjacent cyclical euphoria: multiples several times historical norms. |
The 146 HOLDs include beloved franchises like Apple, Costco, Netflix, Nike, and Coca-Cola. They are wonderful or fair businesses whose prices offer no margin of safety today. Each carries a written conversion level or event trigger.
What would change these verdicts
Every flagged name's conduct conditions double as its sell triggers. I will lift the Tesla veto only on documented restoration of regulator data access, and I will still apply the valuation stage after any lift. I will revisit Boeing when it shows sustained free cash flow. I will change my verdict on J&J on any evidence of current concealment. I will convert my verdicts on Intel, Pfizer, and the managed-care complex on events. Every premium-priced HOLD becomes a BUY at a price that restores a margin of safety. The watchlist is the HOLD list.
Concentration — the list is a position too
The five stages judge one company at a time. A portfolio is a list, and a list has properties no single verdict can see. The clearest is shared exposure: several BUYs, each earned on its own merits, riding the same economic driver.
The framework already handles admission. A second expression of a thesis I already hold gets its own BUY only by clearing every stage independently, never by resemblance to the first. What the list holds after admission is a separate fact. It gets the same treatment as every other flag in this framework: named, in the open, never silently weighted.
The July 2026 buy list, read as a list, holds three concentrations worth naming.
Alphabet, Microsoft, Amazon, and Meta are four businesses. At the snapshot they were 62 percent of the combined market value of the 36 distinct businesses on the buy list. Their theses share one chain: heavy capital spending on AI infrastructure, converting into cloud and advertising revenue. The Alphabet note below works that chain through for one company. The same failure, spending that does not convert, breaks all four at once, and their price ceilings will be crossed in a cluster rather than one at a time.
Fourteen of the 37 buy listings are financial companies. Visa and Mastercard collect the same toll on the same traffic. CME, ICE, S&P Global, and Moody's each take a cut of market activity itself. Each of the fourteen cleared the stages on its own record. Together they mean more than a third of the list by count moves with credit conditions, payments volume, and markets, whatever the individual verdicts say.
The buy list also counts its largest business twice. GOOGL and GOOG are two listings of one company, so the list's combined value carries Alphabet's $4.8 trillion twice. The 200 listings in this exercise hold 199 distinct businesses, and the buy tally of 37 holds 36.
The practice that follows is modest. Name the driver each position rides. Count clusters by names and by value. When a verdict inside a large cluster changes, treat the cluster as a standing flag on that decision. No cluster gets a cap, and cluster membership neither rescues nor condemns a company. The stages decide companies. The list is where I check what the decisions added up to.
Live thesis note: Alphabet and the capex question (July 2026)
Alphabet is the largest change since these verdicts were written. In Q2 2026 Alphabet reported its first negative free-cash-flow quarter since listing in 2004, roughly negative $5.9B. Capital expenditure of $44.9B in the quarter produced the shortfall, and full-year guidance rose to $195 to $205B. Revenue grew 24% in the same quarter. The company behind the original BUY converted cash at an extraordinary rate as a software business. At the margin, that company is becoming an infrastructure owner: data centers, TPUs, power. A different kind of business needs its own underwriting.
The verdict holds on new terms. The underwriting question is whether roughly $200B a year of AI capital earns returns above its cost. My working split assigns roughly 60% to the buildout earning attractive returns, roughly 25% to a moderate overbuild, and roughly 15% to meaningful destruction of value. These are judgments. The demand evidence currently favors the first branch. Cloud revenue grew 82% on a $514B backlog, and management states that demand still outpaces capacity. Backlog records contracted future revenue. Backlog alone proves nothing about future returns on capital.
The auction decides the outcome, and model quality takes second place to it. Alphabet built its moat on the efficiency of matching commercial intent to advertisers, and Search is the surface where that matching happens. AI compounds today's capex on top of unmatched distribution if AI preserves or strengthens that matching. World-class models save nothing if AI answers replace monetizable queries and monetize worse. Secondary conditions are each necessary. Utilization must stay high, because idle accelerators earn nothing. AI pricing must hold against many-way competition. Hardware must earn faster than it depreciates. Cloud must mature into a true second profit engine, and the Q2 margin expansion is evidence on that point. Much of the announced spend remains reversible. Future-year plans are largely deferrable, purchase commitments are partially deferrable, and poured concrete is fixed. Management holds real option value to throttle the spend if returns disappoint, and markets sometimes price that flexibility at zero.
If the spend turns defensive, on evidence such as stalled backlogs, fading pricing power, or capacity exceeding demand, I decay the BUY to HOLD. If search monetization deteriorates without offsetting AI revenue, I decay the BUY to HOLD. If utilization and pricing disclosures imply returns below the cost of capital, I decay the BUY to HOLD. I watch free cash flow and its trajectory against guidance, purchase commitments and construction in progress in the filings, Cloud growth and operating margin, search monetization per query, and any disclosed utilization or AI pricing data. If those turn, I decay the BUY on the terms I stated.
Sources and records
The conduct claims above rest on public records. The Tesla items are sourced first because they carry the veto. The remainder are settled enforcement matters of public record, cited with enough detail to locate each record. Allegations still in litigation are identified as allegations.
Tesla / NHTSA
- NHTSA Office of Defects Investigation — Audit Query into Tesla's delayed and incomplete ADAS crash reporting under Standing General Order 2021-01 (opened August 2025; NHTSA documents note "numerous incident reports" filed months late).
- The Wall Street Journal investigation into Tesla Autopilot crashes, built from police records and heavily redacted federal filings (August 2024).
- U.S. Senate letter to NHTSA (June 2026), noting Tesla is the only reporting company that regularly redacts crash details as confidential business information.
- Benavides v. Tesla (S.D. Fla.): 2025 jury verdict against Tesla in a fatal Autopilot crash, with trial reporting documenting that collision data was withheld from police and plaintiffs until independently recovered.
Alphabet capex note
- Alphabet Q2 2026 results (July 22, 2026): capital expenditure of $44.9B, free cash flow of approximately −$5.9B, full-year capex guidance raised to $195–205B, Cloud revenue +82% with backlog above $514B — company earnings release and call; Reuters coverage.
Settled enforcement and litigation records (public)
- CDC, Smoking & Tobacco Use fast facts: ~480,000 US deaths/year; ~41,000 from secondhand exposure; ~$600B annual costs (cdc.gov/tobacco).
- DOJ/EPA Clean Air Act settlement with Cummins: $1.675B civil penalty over emissions defeat devices (announced December 2023).
- U.S. v. General Motors: $900M deferred prosecution agreement over the concealed ignition-switch defect (S.D.N.Y., September 2015).
- 3M: public water systems PFAS settlement (announced 2023, court-approved 2024) and Combat Arms earplug settlement (~$6B, 2023); documented internal PFAS research reported extensively in litigation and press.
- Reuters investigation into internal J&J talc records (December 2018); ongoing talc litigation and settlements.
- U.S. v. Google: district-court monopolization ruling (D.D.C., August 2024) and subsequent remedies order.
- Wells Fargo: CFPB/OCC consent orders on unauthorized accounts (September 2016); Federal Reserve growth restriction (2018) and its termination (2025).
- FINRA settlement with Robinhood: ~$70M, then the largest in FINRA history (June 2021); Massachusetts Securities Division action over gamification (settled).
- Intuit/TurboTax: $141M multistate settlement (2022) and FTC final order on "free" advertising (2024).
- DoD Office of Inspector General reports on TransDigm parts pricing, with negotiated refunds (2019, 2021).
- National opioid distributor settlement: McKesson's share ~$7.4B over 18 years (finalized 2022).
- Norfolk Southern / East Palestine: $600M class-action settlement (approved 2024) and consent decree with the United States (2024).