How the machine manages the book.

Durable businesses, owned while the wind is at their back.

The Flagship owns 16–26 genuinely durable franchises — strong economics, sound balance sheets, managers who allocate capital well — and owns them while sovereign-wealth flows, hyperscaler capex, defense bills, and enacted policy are at their back. The tailwind is not the thesis; the durable business is. A position holds through drawdowns and weak prints as long as the business and its structural tailwind are intact, and exits when one of them breaks — not because the price moved.

The signal mix is fixed in policy: ~50% business durability (franchise, unit economics, balance sheet, management), ~30% policy & macro tailwind (defense bills, CHIPS/IRA-style programs, the rate backdrop), and ~20% capital-flow confirmation (13F deltas, capex commitments, insider clusters). Technicals and sentiment appear nowhere in that list — they decay too fast and add turnover without alpha.

Positions cluster in four pillars — a diversification frame, not a quota: Compute, AI & Software, Energy & grid, Defense & sovereign capability, Biology & longevity. Builder bias throughout: companies that physically build — fabs, reactors, ships, drug pipelines — are first-class, alongside durable, mission-critical software in the Compute pillar.

The agent decides. Three limits are code.

Since November 2026 the agent decides every trade. Three limits are enforced in code (trades/judge.js) whatever the model writes: no leverage, an 18% ceiling on any one name, and a liquidity floor. The other lines — per-name, theme, sector, AI exposure, the cash band — are shown to the agent on every run; it may go past one only with a stated reason, and the journal records it. Nothing trades on its own: a line crossed wakes the agent, and the agent decides.

Three regimes set the cash band. The agent holds the cash.

Cash is governed by a three-regime composite read weekly from VIX, the yield curve, credit spreads, market breadth, and index extension. The regime call proposes a cash target and a band; the agent sees both on every run and decides the cash it holds.

When deterministic defensive signals fire, the ceiling of the current band extends — Extension (QQQ >+8% above its 200-day MA, or weekly RSI ≥ 70) adds +10pp, Stress (VIX >+25% vs its 200-day, or HY spreads +50bps in 20 days) adds +10pp, and Event risk (a binary macro event within 3 sessions) adds +5pp — hard-capped at 30% of NAV. The floor never moves: the escalator grants permission to defend, never a mandate to sell, and cash above the band with no active signal is rejected in post-validation regardless of the model's reasoning.

Speed comes from detection, not looser rules.

The protocol acts on evidence, not forecasts. A price or volatility shock earns the book permission to defend; a credit or capital-flow break makes de-risking mandatory. Nothing in it predicts a crash: every step waits for a measured signal, and the thresholds, the order in which names are trimmed and the way cash comes back were all written down before the first episode, so nothing is improvised under stress.

The detector is pure data — no model, no cost. It fires when any one of its legs trips: VIX at or above 30, or the one-month VIX above the three-month with VIX at least 20; high-yield credit spreads widening sharply within ten sessions; the S&P 500 at least 7% or the Nasdaq-100 at least 10% below its 20-day high; or the book itself at least 8% below its own 20-day high. The price legs are checked every 30 minutes in market hours, the volatility and credit legs at the close — credit running one session behind the tape, because the public series publishes with a lag. On a fire the regime call re-runs the same session instead of waiting for Monday.

Before going live the protocol was replayed over 2008, March 2020, 2022, August 2024, the DeepSeek shock of January 2025, the April 2025 tariff shock and the summer-2026 chips episode. The replay chose the VIX-below-30 exit over 25, and tightened the volatility leg to require VIX at or above 20 after a run of calm-year false positives. The full replay is in the calibration report.

A score is the output of a written thesis, not the input to a sort.

Every candidate gets a 0–100 conviction score derived from four questions the agent must answer in writing: why this is a durable 20–30 year compounder; what named policy or macro tailwind is at its back; who is confirming it with capital; and what, specifically, would break the thesis. Names below 60 are not investable. Names at 60 and above are sized in proportion to (score − 60), subject to every cap in the ledger above.

Generic theses are refused mechanically at the buy seam: “strong fundamentals,” “compelling AI exposure,” a bare score, or a doc-pointer gets the open dropped and a refusal row appended to the public journal. If the agent cannot articulate the durable business and name its tailwind, the position cannot enter the book.

A 30-day lock protects the book from its own manager.

New positions carry a 30-day floor on trims and closes. Inside that first month, only a true regime change, a real break in the company's thesis, or a policy shock that postdates the entry can shrink the position — not flow softening, not score drift, not a shinier candidate on the watchlist. The lock is mechanical: it sits at every sell-emission seam, and an override requires citing one of the named paths, which is then journaled.

Refusals are public. When the lock (or the named-flow gate above) blocks an action the model wanted, the refusal lands in the journal with the same prominence as a trade — “checked, nothing to enforce” is itself evidence the rules run on every decision.

One deep decision a quarter. Sentinels watching every day.

A 1+ month hold horizon doesn't want a scheduled full model run every week — most weekly reviews re-affirm the same book, which is decision cost without decision value. The scheduled deep run is quarterly; between audits a sentinel layer watches continuously and wakes the model only when something actually changed:

The S&P 500 is not the benchmark.

The book is measured against a custom civilizational-builder basket: 25% per pillar, equal weight within each pillar, rebalanced on the first trading day of each calendar quarter, computed on the same price-return basis as the paper book. The constituent list is versioned and append-only. Beating SPY with a defense-and-uranium book in a tech rally proves little; beating the builders it actually competes with proves the selection.

Don't trust the summary. Check the ledger.

Every regime call, strategist run, scoring delta, refusal, and trade ticket is appended to a public journal whose entries are hash-chained — editing any historical entry breaks its fingerprint and every link after it, and the chain is re-verified on a public endpoint. The reasoning is timestamped before the market opens on it; the costs of producing it are published down to the model call.

One human, in a supervisory seat. The AI makes the calls.

Capital Compass is run by a single independent operator who built the system, pays its API and infrastructure bills, and maintains STRATEGY.md — the investment policy the code enforces. The operator does not pick stocks, size positions, or override individual decisions: every buy, trim, and exit is made by the AI strategist inside the mechanical caps above, and lands in the public journal either way.

The human role is supervision, not management: reviewing the weekly tripwires (§7), authoring the standing research brief the regime call reads, and editing the strategy document — through version control, in public — when the policy itself needs to change. Run failures page the operator; they never silently rewrite the record. No client money is involved, nothing is sold, and there is nothing to subscribe to but the reading.