Can rules-based, AI-managed investing produce risk-adjusted returns above a broad index after trading costs and taxes? The primary metric is Sharpe ratio, net of all costs and after-tax where applicable. The target is above 1.0 — meaningfully better than SPY's long-run Sharpe around 0.5.

Supporting metrics: Sortino (downside volatility), maximum drawdown, Calmar ratio, hit rate, and alpha against a factor model. Benchmarks are chosen per signal — small-cap signals benchmark against Russell 2000, factor tilts against the corresponding factor ETF.

An honest answer requires: at least 40–100 closed trades per signal before drawing conclusions, out-of-sample evidence (not backtest fits), and explicit awareness that with multiple signals running, one is likely to look "significant" by noise alone.

A signal is one complete, self-contained trading strategy — a rulebook that specifies exactly when to buy, when to sell, how much to commit, and the conditions under which it retires itself. No signal trades without a written contract.

Each contract declares: the mechanism hypothesis (why the pattern should work going forward, not just why it worked in the past), entry rules, exit rules, the universe of eligible securities, kill criteria (thresholds that force retirement), benchmark for evaluation, and evaluation cadence. Contracts are version-controlled; any change creates a new version and resets the observation window.

Each signal gets at most two refinements before it must be retired. More than that is curve-fitting to noise.

Every proposed change requires a stated mechanism hypothesis — "it worked in backtest" is not a mechanism. Refinements are validated on data not used to motivate the change (out-of-sample), and parameter stability is checked: if performance collapses when a single parameter moves slightly, the signal was fit to noise.

Refinements happen only at quarterly windows, not weekly. Weekly observation is monitoring; tuning is a formal, scheduled process with a cooling-off period between "I want to change this" and the change taking effect.

The system ingests structured data on pre-committed schedules: SEC filings (10-K/Q/8-K, Form 4, 13F), earnings calendars, macro indicators (VIX, credit spreads, yield curve), congressional disclosures, and end-of-day prices. Each source feeds a specific signal rule that knows what to do with it. Everything else — general news, market commentary, analyst opinions, pundit takes — is deliberately ignored.

This is not laziness. It is the single most important piece of behavioral discipline in the design. "Read the news and decide what's important" is a discretionary layer that re-introduces exactly the biases pre-committed rules exist to defeat: recency, availability, narrative-fitting, confirmation. The documented LLM failure mode in this domain is a model that reads news, reasons itself into a directional conviction, and cannot rationalize exiting it. The structural fix is to remove news from the input set.

Long-term structural shifts — geopolitical risk concentration, AI-driven labor displacement, demographic transitions — are real. The system handles them through structure, not narrative. Geopolitical stress shows up as widening credit spreads and rising volatility, which the regime overlay responds to automatically. Productivity-driven margin expansion shows up as earnings beats, which the earnings-drift signal captures. If a long-term concern can be formulated as a deterministic rule with an entry trigger, an exit, and kill criteria, it can become a signal. Until it can, it is not actionable.

The legitimate path for a new structural thesis is the idea queue: submit it, let it be evaluated for operationalizability, backtested out-of-sample, and considered at the next quarterly refinement window with a cooling-off period before activation. That path takes weeks. The cost is moving slowly. The benefit is producing a track record that is actually evaluable, and not abandoning the system the first time the news cycle feels urgent.

Regardless of any signal's conviction, hard caps apply: no more than 10% of the portfolio in a single ticker, 40% in a single signal, or 60% in a single sector. Leverage is zero by default.

Total exposure scales with the current market-stress regime: full exposure in normal conditions, halved in high-volatility regimes, and a quarter of normal in crisis. The limits are enforced by the execution layer, not by discretion — orders that would violate them are refused before they reach the broker.

The system will not time market crashes correctly. Nobody reliably does. The posture in a crisis is survival, not prediction: smaller positions, more cash, fewer changes, more observation.

In turbulent conditions, the system's trust in its own reasoning goes down, not up. Pre-committed rules are what carry through — the rulebook was written in a calm hour precisely so it can run in an uncertain one. Circuit breakers trip automatically on regime shifts, drawdown thresholds, and severe market moves, pausing new entries until conditions normalize.

The test is not whether the system can predict a crash. The test is whether it can be wrong for eighteen months and continue operating without abandoning its own rules.

Public, with no auth: portfolio value vs. benchmarks, current drawdown, regime state, signal-level health indicators, and the methodology on this page. End-of-day trade activity is published after execution, never before — pre-trade visibility would create exactly the wrong incentives.

Published on a schedule: daily end-of-day activity summary, weekly review every Sunday, monthly performance letter, and quarterly state-of-system. Each published from structured data through templates, not free-form — so the public narrative can't drift from the numbers.

Not published: real-time positions, individual signal rules (mechanism is public; specific thresholds are not), or operator actions. See disclosures for the full non-advice posture.