Playbook心得

2026 ran two sequential, opposite routs — software crashed in March on AI-displacement fear while AI-capex names ran up, then AI-capex crashed in July — so "tech is down" is never a usable statement without naming which act and which layer.

Two companies can execute the identical asset-light strategy and get opposite results, because "asset-light" only says the asset left the balance sheet — it says nothing about where the profit on that asset went.

Measure valuation against a name's own historical multiple band, never against its 52-week high — a stock can be 40% off its high and still trade above every multiple it has ever sustained.

The watchlist's "Trim NNx fwd" convention computes its dollar level as NN x (price / forward P/E), so it scales with whatever EPS the market currently forecasts. On a compounder that is the point — the trim rises with earnings.

Two standard WACC inputs break on REITs in the same direction — the after-tax cost of debt (a REIT pays no entity tax, so the interest deduction is worth nothing) and CAPM's cost of equity (net-lease betas rank inversely to…

[portfolio-specific detail removed from the public build]

A biotech in its first years of commercialisation books upfront and milestone payments from ex-US licensing deals as revenue. Those payments are lumpy, near-100% gross margin, and can exceed early product sales.

When a company publishes a DCF-derived asset metric (Net Earning Assets, NAV, embedded value, portfolio fair value) it also chooses the discount rate.

When a company reports a self-defined revenue segment — "Strategic Revenue", "Core", "Platform", "Adjusted" — the segment boundary is a management choice, not an accounting standard, and the segment's growth rate can be inflated…

The day-one move on a corporate cyberattack disclosure carries almost no information — across ten incidents it clustered between -3.6% and +1.0% regardless of eventual cost.

When a tax-law change forces a company to write a valuation allowance against its deferred tax assets and it later releases that allowance, GAAP net income takes a huge charge in one quarter and an equally huge credit in another.

When a company that has been accumulating losses reaches sustained profitability, it releases the valuation allowance held against its deferred tax assets, booking a large one-time non-cash tax BENEFIT that pushes that quarter's…

When a levered company gives a leverage target and a date, do the arithmetic: (net debt today − target multiple × forecast EBITDA) ÷ (free cash flow − dividends).

/analyze-from-before builds its delta ledger from the baseline report, then researches the window since the baseline date.

Some component suppliers disclose their weighted-average selling price change and their unit volume change every year in MD&A.

When a company divests or discontinues a segment, vendor feeds restate the recent years to continuing operations but frequently leave the oldest years at the old total-company basis — so any CAGR computed across the boundary…

[portfolio-specific detail removed from the public build]

When a stock-funded acquisition uses a fixed exchange ratio with no collar, target holders carry the acquirer's price risk from announcement to close. If they wanted cash and were prorated into stock, they sell it immediately.

When a company raises its revenue guide and its non-GAAP operating-income guide in the same release but leaves the free-cash-flow guide unchanged, the unmoved line is the informative one.

When management reports a segment's growth as "N times faster than overall growth" rather than as its own percentage, the multiple rises automatically as the denominator falls.

When management's EPS algorithm is built from revenue growth plus a guided annual margin expansion plus buybacks, the margin term has a hard ceiling that the guidance never states and no ratio surfaces.

When a company acquires, pays for, or partners its way into a rival's audience, the third-party panel metric it reports (Comscore uniques, app installs, traffic share) may silently start including the rival's properties from the deal date.

Remaining performance obligation is quoted as though it were contracted profit, but three disclosures routinely gut it and all three sit in different parts of the filing.

Form 4 code F means shares withheld BY THE ISSUER for tax.

The watchlist writes trims as "Trim NNx fwd" so the dollar level tracks earnings between analyses. The site computes that dollar as NN x (price / vendor P/E).

When a company widens a non-GAAP metric it must restate the prior-period comparative on the new basis.

[portfolio-specific detail removed from the public build]

A redaction gate built from patterns scoped to specific nouns will pass any disclosure phrased with a different noun, and will report success while doing it.

When a regulatory investigation is known only from press reporting and the company has never disclosed it in a 10-K or 10-Q, there is no company-sourced status at all, so the staleness clock runs from the last INDEPENDENT report…

To test whether a marked equity stake will be written up or down next quarter, compare the investee's CURRENT SHARE PRICE against the PER-SHARE PRICE used at the last mark date, and multiply by the shares held.

A single quarter's free cash flow, taken from a quarterly statement column and described as "TTM," will look like a decline against the prior full year whenever the business is seasonal — because one quarter of a four-quarter…

Comparing a company's next-quarter margin GUIDE against its most recent sequential PEAK manufactures a "permanent step-down" out of ordinary quarter-to-quarter noise; the test that settles it is the year-over-year margin and the…

When a subscription or engagement business loses users, it loses them in order of engagement — the casual, marginal, least-monetizable leave first.

For a business with a hard seasonal pattern, the H2/H1 (or Q4/Q3) revenue ratio is remarkably stable across years and across macro conditions, because the seasonality is driven by customer budget calendars rather than by demand levels.

An entry zone is a derivative of a fair value, so it goes stale exactly when the fair value does — and a zone left above the current price makes every scan report the name as "still below its add zone," which reads as not yet…

A fair value that is not refreshed does not fail symmetrically. On a falling name it gets re-tested quickly because the loss demands attention.

When a manufacturer cannot ship, the variable that decides whether customers come back is not the severity of the cause, the attack type, the ransom, or the headline.

A tangible equity unit (TEU) splits into a prepaid stock purchase contract booked in shareholders' equity and an amortizing note booked as debt.

Third-party measurement panels (Tinuiti, Sensor Tower, Similarweb, Superads and peers) sample a narrow slice — one platform surface, one region, one client mix — and are routinely quoted as if they described the consolidated company.

A multi-tranche entry ladder collapsed into a single entry: LOW-HIGH frontmatter field becomes a band as wide as the whole ladder, which reports the name as in-zone at nearly any price.

The standard net-lease thesis is that the tenant bears operating risk while the landlord collects contractual rent. That holds only while the tenant's rent coverage is comfortable.

When a prior analysis gates an upgrade on a falsifiable numeric test, the next earnings print arrives wrapped in management language that sounds like the test — "growth accelerated again", "momentum returned", "record backlog".

On ADRs, vendor feeds frequently apply the ADR ratio to the price field but not to the book-value-per-share field, and separately fail to convert the local-currency book value to USD.

For at least one reporting period after a spin-off, the company reports pro-forma results excluding the spun entity while vendor consensus still reflects the pre-spin company.

A newly spun-off company's vendor balance sheet is the pre-separation carve-out for weeks or months after the distribution.

In an Up-C (public C-corp over an LLC with exchangeable member units), Yahoo/fin.py break in six places simultaneously: impliedSharesOutstanding double-counts the exchange by adding the non-economic voting stub on top of a base…

A change in the fair value of contingent consideration (an acquisition earnout) is recorded in OPERATING expenses under ASC 805, not below the operating line.

The ADR data traps are properties of the depositary structure, not of foreign domicile.

Vendor feeds have been caught misreporting the exact figures that decide verdicts — FCF, margins, non-operating income — so any number the conclusion rests on gets pulled from the 10-K/10-Q or press release, not from a data feed.

[portfolio-specific detail removed from the public build]

Hotel, restaurant and other asset-light franchisors report a large cost-reimbursement and marketing-fund line inside IFRS/GAAP revenue.

A company with flat operating income and flat free cash flow can post double-digit EPS growth indefinitely by retiring shares, and no screener distinguishes that EPS from operating EPS.

Cabot Corp's CMP slurry business left in 2000 and now sits inside Entegris — CBT today is ~63% carbon black for tires, so including it in a semiconductor-materials basket flatters the group's numbers with an unrelated business.

A company can add a billion dollars of productive assets in a year, report a falling capex line, and publish a positive free-cash-flow margin throughout — by acquiring the equipment under finance leases and equipment-financing…

[portfolio-specific detail removed from the public build]

In razor-and-blade capital-equipment markets (EP catheters, surgical robots, sequencers, analysers, lithography), the installed console is usually read as a one-way moat. It is not.

Measured across 141 watchlist entries, the 0-10 conviction scale operates in practice as a 3.5-8.5 scale with 62% of names inside a 1.5-point band.

A relative-value comparison is only valid if every name in it is on the same basis.

In a GAAP loss year all dilutive securities are anti-dilutive and excluded, so diluted shares equal basic. In the next profit year options and RSUs enter the denominator all at once.

DYT compares current yield to a historical yield band, but the yield can only rise two ways — the dividend went up or the price went down — and DYT cannot tell them apart, so a collapsing price mechanically produces DYT's…

Yahoo's default close price is dividend-adjusted, so building a historical P/E, P/AFFO or P/B band from it understates every past multiple by the cumulative dividend yield — on a high-yield name this manufactures a false "it has…

Entegris FY2025 margins are corrupted in Yahoo and roic.ai simultaneously — the widely quoted ~41% EBITDA margin is fiction, the true figure is 13–17% — proving that cross-validating two vendor feeds is not an independent check.

Every per-share figure in a vendor feed — EPS, diluted-share CAGR, P/E — is built on the WEIGHTED-AVERAGE share count, which by construction lags a discrete issuance. Market cap is built on the BALANCE-SHEET count, which does not.

A state-sponsored liability backstop (a wildfire fund, an insurance pool, a bailout framework) protects the enterprise, not the equity.

When a vendor issues warrants to its own customers as a commercial inducement, GAAP capitalises a "customer warrant asset" and amortises it as a REDUCTION OF REVENUE over the life of the related contracts.

As of July 2026, every AI-adjacent field swept traded above its own historical multiple band despite large drawdowns, because the drawdown is smaller than the preceding run-up — so drawdown-driven screens in AI-adjacent sectors…

The standard moat test reads margin compression as moat erosion, so stable margins read as an intact moat. But an incumbent under attack chooses between ceding volume and cutting price, and only the second shows up in margin.

ADR ticker lookups return clean, plausible data for an entirely different company — SCRYY resolves to SCOR SE not Screen Holdings, CPCAY to Cathay Pacific not Capcom — so always confirm the returned company name before using any foreign-listing data.

Forward P/E should sit below trailing P/E whenever earnings are expected to grow.

When a company issues warrants to a government or strategic investor and books them as a derivative liability rather than as equity, the mark-to-market runs through income — so the GAAP loss GROWS as the stock rises.

Labour-based services firms (IT services, BPO, consulting) are all being marked down on the same thesis — that generative AI deflates the seat-based model faster than it creates new demand — and management commentary cannot…

During a supply shortage a company can post its fastest revenue growth in years while losing unit share in every segment, because price and mix carry the whole increase.

Across the March 2026 screen series the same ranking held at every pass — application software offered the deepest value, then cybersecurity and infrastructure software, with hardware consistently weakest on risk/reward — so…

Compute incremental gross margin as the change in gross profit divided by the change in revenue, and compare it to the average gross margin.

"Customer retention >95%" counts accounts, not dollars.

A company can move supplier or data costs from contra-revenue (netted against the top line) into operating expense.

Screening sources keep surfacing companies that have been acquired, taken private, delisted or wound down — check this list before researching any unfamiliar name, because the research is wasted and the "cheap valuation" is often a stale artifact.

Net margin should always sit below operating margin — interest and tax subtract. When a screen shows net margin above operating margin, the excess did not come from the business.

The net-margin-below-operating-margin tripwire only detects a non-operating gain large enough to invert the inequality.

Logging a sale as "fewer shares, same average cost" is wrong whenever the broker sells specific lots rather than an average, which is the normal case.

A foreign company's reported ROE is denominated in its functional currency, not in USD — discounting a local-nominal return stream at a USD cost of equity credits the full nominal local return in the numerator while charging only…

On a US-listed ADR, Yahoo's epsCurrentYear/priceEpsCurrentYear can be quoted in the company's LOCAL reporting currency (e.g. RMB, BRL) while forwardEps/forwardPE is quoted in USD against the ADR price.

For an ADR that reports in a currency other than USD, Yahoo/fin.py computes enterprise value by subtracting the LOCAL-CURRENCY cash and investments from a USD market cap. The result is wrong and often negative.

For certain filers, vendor normalization maps segment-level SG&A into "Cost of Revenue" and leaves only corporate G&A in operating expense, so the income-statement gross margin lands 6-8 points below the reported figure.

Extrapolating a company's next earnings date from last year's date is unreliable — the date moves by days or weeks, and syndicated financial sites publish the extrapolation as fact.

Payroll processors report client payroll funds they hold in trust as current liabilities, which crushes the quick ratio to a fraction of 1 regardless of actual liquidity — PCTY reads 0.094 and ADP reads 0.155 for the same structural reason.

"Regulated industry" is not a moat descriptor and carries no directional information on its own.

A valuation supported by addressable-market size rather than realised revenue is a story, not a thesis — require the revenue line to exist and grow before any multiple is applied.

IQVIA-style prescription data measures US drug sales at list price while the company reports net of rebates, so the two figures for the same product in the same year can differ by the entire gross-to-net spread — up to ~70% of…

Short interest is conventionally quoted as a percent of FLOAT, but float excludes insider-held shares. On a founder-controlled company the float can be half the shares outstanding, so the same short position reads roughly double.

Shortlists published into a fear-priced tape have run 8-27% within two weeks on all three occasions recorded (2026-07-14, 07-29, 08-04), so the analysis layer is reliable.

For companies that co-fund assets with JV or SCIP-style partners, vendor "free cash flow" (operating cash flow minus capex) and the company's own "adjusted free cash flow" can differ by more than the entire capex line, because…

When a company reduces stock-based compensation by shifting the compensation mix toward cash, SBC falls and free cash flow falls by the same amount, so SBC-adjusted FCF is mathematically unchanged — yet the headline "SBC as a…

Syndicated market-research summaries reverse the direction of strategic minority investments, reporting the issuer as the acquirer.

In 2026 the AI capital cycle shows up in financial statements as two simultaneous and opposite distortions: non-operating marks on private AI stakes INFLATE reported earnings, while capex or AI-delivery COGS DESTROY the free cash…

Yahoo and .mcp/fin.py populate the country field from the issuer's principal executive office address, not from its jurisdiction of incorporation.

When successive re-underwritings of the same name each cut fair value but each still land ABOVE the prevailing price, and the price goes through every one, the pattern is diagnostic of the method rather than of the market.

The cash a company spends withholding shares to cover employees' RSU tax bills is a financing outflow that looks exactly like a buyback, and vendors disagree about whether to put it in the "share repurchase" line.

Under ASU 2016-01 a company must mark its equity holdings to fair value through net income, so a large private stake that IPOs or revalues can inject tens or hundreds of billions of non-operating, non-cash income into GAAP…

Yahoo and roic.ai report bank revenue as a net figure (roughly net interest income plus net fee income), which can understate the issuer's own reported total revenue by a third — making every vendor-derived P/S, revenue CAGR and…

Yahoo's enterpriseValue is not recomputed on the same clock as marketCap.

Yahoo's enterpriseValue is market cap + debt − cash. It adds no non-controlling interest.

For REITs that book acquisitions as sales-type or financing-receivable lease investments rather than as business combinations, Yahoo and roic.ai report an acquisition line of exactly zero while billions are being deployed, and…

Consensus forward EPS in a vendor snapshot is a rolling average of analyst estimates and does not update same-day after a guidance cut — but price does, instantly.

Yahoo's forwardPE/forwardEps does not state which fiscal year it estimates, and it is frequently neither the current year nor a consistent one across tickers — sometimes a year already three-quarters reported, sometimes a year out.

A company that converts owned assets into optioned assets pays a continuing carry fee to the option counterparty. Under ASC 970 that fee may be capitalized into the asset basis rather than expensed as incurred.

If a platform's discovery algorithm can be tuned toward extraction, then its reported DAUs, hours and bookings measure the setting of that dial as much as they measure user demand — so a growth vintage produced by an…

A de-rating company can generate a full set of value signals - a low trailing P/E, a high dividend yield, a large drawdown from the high - where every one is mechanically produced by the de-rating itself rather than by the price falling below value.

Yahoo Finance misclassifies mark-to-market gains on strategic equity stakes as "Interest Income", inflating apparent non-operating yield to impossible levels — sanity-check interest income against the cash balance.

For platform businesses that pay out to creators, Yahoo's grossMargins field nets content/creator payouts and infrastructure out of gross profit and labels the result a gross margin.

For an issuer that reports in one currency but trades in another (CAD-reporting, USD-listed), Yahoo converts the PER-SHARE fields to the quote currency but leaves the AGGREGATE income-statement fields in the reporting currency…

Yahoo's getstockinfo returns populated grossMargins and operatingMargins for banks, lenders and card networks, which have no cost-of-goods-sold line and therefore no gross margin.

Yahoo's enterpriseToEbitda is not always computed from EBITDA. On issuers carrying heavy acquired-intangible amortisation it tracks the EV/EBIT series instead, overstating the multiple by the entire D&A load — 2x or worse.

Yahoo's getstockinfo snapshot bookValue (per-share) field can stay pinned to the pre-buyback value for at least a quarter after a large — especially debt-funded — share repurchase shrinks shareholders' equity, so it overstates…

When a company spins off a subsidiary, Yahoo records the value adjustment as a fractional "Stock Split" row in getstockactions and retroactively divides every earlier dividend by that factor.

getstockactions can return a dividend row that is exactly 2x the amount the issuer declared, with no split, no special dividend and no currency event to explain it.

Yahoo Finance computes dividendYield and trailingAnnualDividendYield off previousClose, but computes trailingPE off currentPrice, inside the same payload.

Yahoo's epsCurrentYear / priceEpsCurrentYear fields can freeze at a stale vintage and never update, while the live consensus in the same vendor's epstrend table moves materially.

Yahoo's heldPercentInsiders counts common stock only. A founder or sponsor whose entire stake is convertible preferred plus warrants shows as 0.00% insider ownership - the exact inverse of the truth.

Yahoo Finance's insiderpurchases summary reports RSU vesting at $0.00 and option conversions at legacy strikes as open-market "purchases" — on RBLX it showed "+3.5M net shares purchased, 12.3% buy" during a period with literally…

For dual-class companies — foreign private issuers were where this was found, but US domestic filers fail identically — Yahoo's sharesOutstanding field reports only the exchange-listed class, omitting unlisted super-voting…

Yahoo's getstockinfo snapshot returns a trailing "freeCashflow" field that does not always equal the sum of the four most recent quarters in its own getfinancialstatement (quarterlycashflow) response — even when that sum is…

.mcp/fin.py --screen defaulted to 25 rows with no way to override it, so every screen returned exactly 25 regardless of how many names passed the filter, under a "— 25 hits" header that reads like a match count.

.mcp/fin.py prints its balance-sheet block from the latest ANNUAL statement while the snapshot block's price-derived ratios (BVPS, P/B, EV, D/E) come from Yahoo's most-recent-quarter fields.

[portfolio-specific detail removed from the public build]

Whenever roic.ai's cfcapexpenditures is null, it publishes cffreecashflow equal to cfcashfromoper — FCF with no capex subtracted — and silently propagates the error into freecashflowpersh and prtofreecashflow.

CLAUDE.md's Tool Hierarchy lists roic.ai as the escalation path for "20-40yr statement history," but on the currently-authorized account it silently returns a single fiscal year and refuses any request reaching further back.

yfinance's fastinfo.previousclose does not reliably return the prior regular-session close.

Swept 2026-06-12 over two rounds and 25 names as an INDI alternatives search — nothing cleared, and the pass also surfaced unresolved market-cap discrepancies between Yahoo and web sources on SITM, VICR and POWI.

Brazil's central bank runs a consistent, repeated policy of building free public infrastructure that competes away the profit pool of each banking layer in turn — payments, then data, then interest rates, then installment credit…

Swept twice (Apr and Jul 2026) against a deteriorating private-credit backdrop — BDC secondary sales fell 40% amid lengthening default shadows — so the field was screened on dividend coverage and NAV discipline rather than on yield.

Broad "find value anywhere" screens have consistently underperformed field-specific ones in this repo — they surface names already covered elsewhere and, in the global case, produced an FX artifact mistaken for a discount.

Swept Jun and Jul 2026 — the useful distinction is revenue model, not sector: IT (Gartner) sells subscriptions and BAH sells government work, both uncorrelated to the project-based tech-spending cycle that drives ACN and CTSH.

Swept 2026-07-29 — since the user already holds BTC directly, coin-beta names are leverage on an existing position rather than diversification, so the only interesting question is which names have a real business independent of the coin price.

Swept 2026-03-21 for Visa-like asset-light toll-booth businesses — the field's quality is uniformly high so the binding constraint is price, not business selection.

Swept 2026-07-31 via NVO — the field is the largest therapeutic category in pharma history and is simultaneously the clearest case of a huge field with a losing incumbent: obesity volume compounds for decades while value per…

Swept 2026-07-29 — NTES (conviction 7.0) survived as a non-AI-levered survivor, but it carries China structural risk, which is a different risk class from the valuation risk the sweep was screening for.

Swept across five March 2026 screens — consistently the weakest risk/reward layer in the tech selloff (narrower moats, softer FCF) with NTAP/CDW/ZBRA the only names of interest; never refreshed since, so the entire field is stale by four months.

Swept 2026-07-29 and produced the best result of the entire July exercise — DOCS at conviction 7.0 and in its entry zone, because the field is not AI-narrative-levered and so did not participate in the run-up that left every adjacent field expensive.

Swept 2026-07-29 as the "unswept fields" pass and rejected by the user on business-model grounds rather than valuation — the durable finding is that the 2026 defensive rotation bought the label, not the cash flow.

Swept 2026-07-29 and the premise failed — Japan's apparent discount was an FX artifact (EWJ −8.4% ≈ yen −8.41%, while TOPIX set a record close), and Japanese AI names carry zero discount, with Advantest at 60x P/E and 28x P/B.

Swept 2026-06-10 across the full fuel-cycle stack (mining → enrichment → SMR → generation) — the durable finding is that the layers carry completely different risk types, so "nuclear exposure" is meaningless without naming the layer.

Swept 2026-07-29, no entries — 10 of 13 names traded above their own historical bands (LITE at 2.6x its own recent-high P/B) and three of the field's familiar names are no longer independently investable.

Swept 2026-04-28 (IONQ, RGTI, QBTS) — only QBTS cleared the bar for even a speculative position, and only on a 40-45% pullback, because it was the sole name with a demonstrable near-term business rather than a roadmap.

Swept twice (Apr and Jul 2026) on a discount-to-NAV/AFFO-durability basis, with the July run adding US net-lease and industrial names; AFFO coverage rather than headline yield was the operative filter throughout.

Swept 2026-07-29, no entries — the quality names (CGNX, ZBRA and peers) trade at 32-35x earnings despite the selloff, so the field is a watch-for-price list rather than a research problem.

Swept 2026-07-15 to 07-29, no entries — all 11 ATE names traded above their own historical multiple bands despite a −28.6% SOX drawdown, and the WFE names that matter are gated on August prints rather than on price.

Swept 2026-07-29, no gems — materials tracks fab UTILIZATION while equipment tracks fab CAPEX, so "materials is the defensive way to own semis" holds against capex cuts and inverts against utilization declines; in the July 2026…

Swept heavily in June 2026 (50+ names across four screens) and the field's own verdict was that it is "genuinely hard to screen right now" — quality is scarce, the graveyard of dead names is large, and story-over-revenue is the dominant failure mode.

The most heavily swept field in the repo (12+ screens, Mar–Jul 2026) — the March SaaSpocalypse produced genuine below-own-band value in application software, but no screen has ever been scored against outcomes, so the field's real lesson is still unread.

First look at the field, 2026-08-19, via LEN plus a seven-name peer read.