How It Works

A daily-updated brain for the entire family balance sheet — the public-market accounts, the boat dealership, and the real estate — in one place. It runs on four inputs (each detailed below and listed in full under Data Sources): a curated roster of macro and sector experts whose new interviews and two paid letters are read every morning; live market and economic data (prices and fundamentals from Financial Modeling Prep, macro series from the Federal Reserve's FRED, Bitcoin on-chain feeds), refreshed daily and live on the page; a daily event layer (news on every holding, the US economic calendar, marine-industry headlines); and the family's own books, each syncing on its own clock. The experts and data shape a single Living Thesis with explicit probabilities; the numbers do the ranking and the math; the family makes the decisions. Every figure is checked against the live number, and the system builds the case rather than making the call.

We track 23 YouTube channels, 15 named guest voices, two paid newsletters, and individual research voices on X. Channels split into tier 1 (core macro and sector voices, fully transcribed every run) and tier 2 (high-volume channels like Bloomberg Technology and a16z, keyword-filtered to only the clips that touch our themes). Guest voices (Lyn Alden, Gromen, Bianco, Visser, Gundlach, Druckenmiller, Lacy Hunt, Raoul Pal and others) are caught by name whenever they appear on any tracked show. The two paid newsletters (Lyn Alden premium and Northstar & Badcharts) are read in full through an authenticated browser, never scraped. Individual X voices, like Peter Mantas for biotech, are pulled on demand. The roster is curated, not comprehensive: quality of thinking over volume.

Each morning the system reads every feed, transcribes new videos, summarizes them, and grounds the numbers against live data before anything reaches the thesis. Transcripts are fetched through a residential proxy (YouTube rate-limits the unattended run's shared server IP; the proxy gives each request a fresh address); if a must-have episode still resists, a fallback reads it from the rendered page in a logged-in browser. Every run reports which transcripts came through and which were blocked, so a missed source is never invisible.

The roster supplies opinions; a separate layer supplies facts. Each morning the system pulls the last day's news for every holding, a general feed filtered against the standing theses and the dealership, the day's top market-wide stories, the week's economic calendar, and upcoming earnings for names we own. It also sweeps SEC EDGAR for every document a held company filed — annual and quarterly reports, 8-K current reports including earnings releases, share offerings, big-stake notices — because what a company files is ground truth in a way a headline never is: each new material document is read the same morning, its takeaways land on the Portfolio page with a link to the official filing, and any thesis whose underlying assumption the document touches is checked and, if the assumption moved, rewritten the same run. The discipline is strict: every item must map to a thesis, a trigger level, a holding, or the dealership, or it is dropped, and a headline alone never moves a probability — news raises questions; the data and the tracked analysts answer them. The exception is hard fact (a central-bank decision, a signed deal, an earnings report), which updates the record immediately. Looking forward, a rolling three-week catalyst calendar lists the scheduled, price-moving events known in advance — Fed decisions, the major inflation and jobs reports, earnings of names we own or watch, options-expiry and index-rebalance days, Bitcoin and biotech milestones — each dated and tagged with the positions it tests, so no scheduled event arrives as a surprise.

We verify rather than trust. Prices, fundamentals and analyst estimates come from Financial Modeling Prep; macro series like CPI and fed funds from FRED; Bitcoin on-chain metrics from dedicated feeds; the authoritative MicroStrategy and STRC figures from the company's own dashboard at strategy.com. When an expert cites a level or multiple, we check it against the live number first — and the macro series are themselves the basis for the consumer and housing theses, where the data leads and no expert is needed.

The thesis is the brain of the system. Each day's synthesis updates a small set of core theses, each carrying an explicit conviction percentage, the experts who support it, the dissent against it, and the specific signal that would change our mind. Disagreement is preserved rather than averaged away, because that is usually where the information is. A revision log records what changed each day and which sources drove it, newest first. Two standing rules: for the MicroStrategy and STRC leverage read we lean hardest on the quantitative risk experts at True North (Jeff Walton), while the broader Bitcoin cycle view draws on the full roster; and the thesis is written to be right, never to be shareable — the marketing function reads it but can never edit it.

Levels are never news; changes are. Borrowing from the alpha-capture programs the big quantitative funds run on human experts (the design owes a debt to a former Two Sigma builder of exactly such a system, interviewed on Odds On Open in July 2026), the system now keeps a longitudinal record of every tracked voice's stated view on every asset — not just what they think, but when it changed. Each morning's read of the roster appends observations to a ledger: who, which asset, how bullish or bearish on a five-point scale, and the verbatim quote that backs it (no quote, no entry). A small engine then computes the deltas, and only the deltas print. The patterns it watches, each grounded in how expert opinion actually behaves: a voice with no prior view taking a side for the first time (historically the strongest follow-them signal — someone who finally got off the fence did the work); a voice already at maximum conviction getting even louder while the rest of the roster leans the same way (historically a fade — when everyone is comfortable, the move is mostly priced); and the first credible defector breaking from a crowded consensus (the strongest contrarian information there is). An asset-level gauge also flags when the whole roster's average lean gets extreme. The briefing shows only state-changes, and a weekly scorecard tracks how each signal type has actually performed, with the honest caveat that our panel is a few dozen voices, not the thousands a fund would normalize across — so these are attention flags that direct deeper work, never automatic trades.

One scheduled job runs at 6am: it ingests the roster, syncs the live portfolio, pulls the event layer, updates the Living Thesis and every page's view block, writes the day's briefing, then fetches fresh market data, rebuilds all eleven pages, and publishes the site. A Sunday review goes deeper: it re-scores every conviction from scratch, consolidates the week's prose, audits the charts, and publishes a state-of-the-thesis note. Scheduled jobs run while the app is open; if it is closed, they run on the next launch.

One writing standard across every page: the reader is a smart business executive, not a market professional. People get a role on first mention, every term of art is explained inline the first time it appears, and every data point carries its implication in the same sentence — the numbers are never dumbed down, only the jargon. In the Briefing, any mention of a standing thesis is a click-to-open popup showing that thesis's full current text, parsed live so it can never go stale.

The anchor is the power-law trend: over its full history Bitcoin's price tracks a straight line against time on a log-log scale, which gives a long-run fair value, and today's price against that line is the first read. We then blend a composite z-score across the Mayer multiple, MVRV, the Puell multiple, the fear-and-greed index, and a net-liquidity measure — negative means cheap, positive expensive — with the buy trigger around minus 1.3 standard deviations and the trim zone around plus 2. The MicroStrategy and STRC leverage complex is tracked separately, because the fragility there is the leverage and the dollar-reserve runway, not spot Bitcoin.

The AI, Oil & Gas, and Biotech screens share one engine. Biotech keeps its own page; the AI and Oil & Gas screens run HEADLESS since July 2026 — their pages were retired, their scores feed the Portfolio page's Model column, and their standing views render as the Sector Views section of the Investment Thesis. Each name is scored within its own layer of true peers, apples to apples, on five axes: Value, Quality, Trend and entry, Analysts, and Roster conviction. Scores are robust z-scores (built from the median and median absolute deviation, winsorized) so one outlier can't distort a small layer. Within a layer, a score at or above plus 0.5 sigma screens BUY and at or below minus 0.5 sigma screens TRIM — a relative ranking and long-horizon entry tool for one-to-five-year holds, not a trading signal. The axes: Value is cheapness on the metrics that fit the sector; Quality is durable profitability and a sound balance sheet; Trend and entry rewards long-term relative strength but penalizes being stretched above the 200-day average, so it favors pullback entries; Analysts blends price-target upside, the net buy-versus-sell rating, and estimate revisions; Roster is a deliberately sparse conviction tilt from the tracked experts. A market-temperature gauge atop the Biotech page shows how extended that universe is, as context.

The valuation column leads with the best sector-appropriate multiple rather than forcing one template on every name: AI and tech use forward P/E and PEG (PEG on a multi-year growth rate); energy uses EV/EBITDA and free-cash-flow yield; commercial biotech uses P/E where it exists, else EV/EBITDA. Forward P/E is deliberately not the headline everywhere — it needs analyst coverage and positive earnings, and a P/E on a tiny number is distorted — so we lead with whatever is available, keep forward P/E as a secondary line, and drop it from the scoring when missing so a blank never biases the result.

Most of biotech is pre-revenue, so P/E and margins are meaningless there, and we split each name by stage. Commercial, profitable names keep full fundamentals. Clinical, pre-revenue names are scored on balance-sheet survivability through two repurposed columns: Pipeline (enterprise value against cash on hand — below net cash means the market assigns the pipeline negative value, either deep value or distress) and Survival (cash runway, the months left at the current burn rate, plus share dilution). Cash runway is the single most important number for a clinical biotech, because it decides who reaches their next catalyst without a dilutive raise. This keeps the read honest: a name a roster voice loves can still rank Screen− if it is the most expensive and most dilutive of its peers. A planned upgrade adds a catalyst axis from clinicaltrials.gov (trial phase and expected readout date); true risk-adjusted NPV needs paid pipeline data, so for now we approximate with these objective proxies and say so.

Every page opens with a view block, rewritten each morning from the updated thesis so data and narrative always agree. It is structured, not prose: a one-line stance, the conviction percentage, a short read, the live trigger levels with what a break of each would mean, the named voices and dissent, what would change the view, and when it last changed — every level checked against live data. The discipline: a view you cannot reduce to a stance, a number, and a falsifiable trigger is not yet a view. (The Macro page is the exception — its block is a fresh analysis of the day's data.) The thesis text itself states the current view only; history and conviction paths live in the Revision Log and the Daily Briefings.

Which prose is refreshed daily, and which is deliberately not (2026-07-31). The macro dashboard carries three layers of writing above the same tiles, and they follow opposite rules on purpose. The tiles themselves and the "What the board is saying" narrative beneath them (`knowledge/macro_summary.md`) are rewritten EVERY run from that morning's `macro_data.json`, so both carry live prints. The per-group card footers and the per-tile click-through notes live in the page builder and are EVERGREEN: they explain what an indicator means, how to read it, and where its decision gate sits (WTI $89, mortgage 5.5%, savings 4%, payrolls 75k/month, high-yield 4%), and they must never quote a current print. The reason is a failure we shipped twice: hardcoded prints in builder prose silently contradict the live tiles they sit under within days, and the reader has no way to know which one is current. So numbers live where they refresh, interpretation lives where it does not, and named thresholds — which are decisions, not data — sit with the interpretation.

The site opens on the day's Briefing. Beyond the Investment Thesis and the two model pages (Bitcoin, Biotech — the AI and Oil & Gas screens run headless, surfacing through Portfolio and the Thesis's Sector Views), the tabs are:

• Briefing — the morning capital-allocation command center: a plain-English "This Morning" line, a "Capital Command" gauge of the three engines (business, public markets, real estate) against their targets with a where-the-next-dollar-goes read, then the day's market moves, the top stories moving markets, what to watch, the lesson, and sources. History lives in the Notion briefings database.
• Macro — the conditions dashboard: every indicator must feed a thesis, mark a trigger, or be one of the roster's named tells (policy and liquidity, rates, inflation, labor, credit, dollar and hard assets, energy), each with 2/5/10/20-year charts and a ten-year range gauge. Its centerpiece is the trigger board: every falsifiable thesis level as a live ARMED/TRIGGERED row, so the page answers "which conviction is the market voting on right now." A throttled source carries its last-good value rather than failing the page.
• Family — every family asset rolled up against a target allocation, with a net worth chart by asset class, a liquidity ladder, owner and manager breakdowns, and a standing action queue. The math runs on the investable base; personal residences are shown but excluded (a home is consumption), and real estate splits into dealership, investment, and personal so each is judged on its own job.
• ACM — the boat dealership as a capital-allocation problem: nine years of statements in four pictures (revenue, profit, inventory versus floor-plan debt, return on equity), the current balance sheet, an Industry watch of demand-driver news, and a standing view on where the next dollar earns most. A weekly dealer-system export adds inventory and unit-sales trends — the family's own ground-level read on big-ticket demand, feeding thesis 8 often months ahead of national marine data.
• Portfolio — the live book: NAV and P&L on live prices, a performance scorecard (return versus the S&P, Nasdaq, world, Bitcoin, gold, bonds and a 60/40 mix, the alpha over each, and the risk taken: Sharpe, beta, volatility, drawdown), the annual track record, factor-cluster exposure, allocation versus target with the dollars to close each gap, every holding joined to its model signal, a Filings & earnings section (every SEC document a held company filed in the last 45 days, with the morning run's takeaways and links to the official filings), and a flags panel. The sheet stays the source of truth for targets.

Family and ACM carry the same redactions as everywhere: category and entity labels, never account numbers or property addresses. One chart rule across the site — macro composites come from Caliban (3Fourteen Research), the Bitcoin visuals are generated natively from our model, and every chart must support the specific claim beside it or it comes out.

Risk is measured in factor clusters, not account sleeves, because positions that crash together count as one exposure: MicroStrategy, the spot Bitcoin ETF and the STRC preferred are a single Bitcoin-complex position, not three diversified holdings. The daily run recomputes cluster exposures and a small set of explicit stress scenarios (betas in an editable config) against the live book, and every briefing carries them, so concentration is always visible next to the day's views. Target weights in the portfolio sheet remain the allocation authority; a fuller framework of binding limits exists in draft and is parked by choice.

A June 2026 recalibration made the division of labor explicit. The sector models are SCREENS — relative rankings labeled SCREEN+ and SCREEN−, never buy/sell verdicts — and their job is triage: deciding which of ~110 names earns deeper work, gating clinical biotech on survivability math, and flagging divergences between the model and the tracked experts (the highest-information rows on the site). The Bitcoin model keeps verdict language because it is a single-asset valuation framework with a structural anchor and explicit thresholds, not a cross-sectional ranking. When a name graduates from the screen, it goes through the name-level diligence checklist: a Third Point-style gate of 14 universal points plus a sector module (about 20 to 26 total) — the differentiated claim with a number, business quality, what the price implies, dated catalysts, regime and cluster fit, the steelmanned bear case, management's say-do record, and a pre-committed invalidation — output as a one-page memo that doubles as the decision-journal entry. Decisions themselves are made at the sleeve level, where the thesis, the cluster lens, and the sheet's targets live.

This is not investment advice. Conviction is a probability-weighted synthesis of other people's views, not an independent forecast unless noted, and disagreement is preserved rather than resolved. The sector models are relative-ranking screens and long-horizon entry tools, not market timing. Risk-adjusted NPV for biotech and a full catalyst calendar require data we do not yet pay for, and the pages flag where we approximate. The scheduled jobs only run when the app is open. The system is built to be transparent about all of this on purpose: a screen you can audit is worth more than a black box you have to trust.

Everything the system reads, grouped by the role it plays. The rule: a source earns its place by feeding a thesis, a model, a trigger, or the family balance sheet — and this registry is updated in the same change whenever a source is added or removed.

Opinion layer — what smart people think

Curated experts whose reasoning feeds the Living Thesis. Disagreement is preserved, never averaged away.

Data layer — what the numbers say

Hard market and economic data; used to verify every expert-cited figure and drive the models.

Event layer — what just happened

Facts to test against the theses, never a second opinion stream. Map-or-drop: every item must tie to a thesis, trigger, holding, or the dealership. Headlines alone never move a probability.

Family layer — our own numbers

Private data; the only sources the outside world cannot give us. Drives the Family, ACM, and Portfolio pages.