Publish the trail, not just the result.

Abstract · condensed

The first paper moves Do The Right Math from a decision philosophy into a live U.S. equity experiment. It frames disciplined decision-making through explicit players, strategies and payoff ranges, then applies that structure to an event-driven portfolio process in which AI augments search and interpretation while final portfolio actions remain human-governed. The study follows the DTRM Fund from June 2025 to March 2026 and compares it with the S&P 500, a recurring LLM-only portfolio and the AI-managed AIEQ ETF. Across five completed cohorts, DTRM reported a 32.99% compounded return versus 9.17% for the ChatGPT Fund, 8.75% for the S&P 500 and 5.98% for AIEQ. The paper's central claim is architectural rather than algorithmic: outcomes depend not only on prediction quality, but on how information is translated into accountable action under uncertainty.

Abstract · condensed

Paper II narrows the question from portfolio architecture to the value of a frozen probabilistic downside guardrail around an existing deterministic ranking. Three quantile models estimate lower-tail and central outcomes, while the human-defined rule decides how that uncertainty constrains selection. Direct probabilistic ranking did not improve the baseline, and an absolute downside veto failed out of sample. A preregistered relative P10 veto then produced positive incremental Top-K target value across three temporal tests, with positive news-event-cluster bootstrap intervals in each cohort, although ticker-cluster sensitivity was wider and not uniformly conclusive. The result is intentionally narrower than a claim that humans always beat AI: calibrated downside information can improve a fixed decision process, but its effect is regime-dependent. The broader research thesis is that increasingly similar machine systems may share correlated failure modes, while human value may persist through different objectives, constraints and models of error.

Abstract · condensed

Paper III separates prediction from decision even more explicitly. The predictive models remain frozen; only the rule that converts their probabilistic state into action changes. MM1 is a preregistered MaxMin selection layer that receives the existing point estimate, calibrated lower-tail estimate, baseline rank and probabilistic guardrail. It builds an asset-specific downside width and selects the exact-K portfolio that maximizes worst-case mean value under a budgeted adversarial uncertainty set. On the preregistered V5 temporal holdout, MM1 improved realized Top-K target-model value by 0.05725 and hit rate by 5.49 percentage points relative to the promoted Phase-2 rule. The identical frozen policy then improved value by 0.11065 and hit rate by 15.86 points on the non-overlapping prospective V6 holdout. The paper does not claim absolute profitability or universal robustness; it shows that decision quality can improve without retraining the predictor when uncertainty is acted on differently.