Methodology
The machine searches. The human defines the game.
Human Clause separates the search signal, the decision rule and the realized outcome. From the 2026-09-28 cohort, the public experiment also compares human selection with a direct robust MM1 portfolio under the same eleven-player geometry.
1. Define
Humans define the objective, horizon, eligible universe, portfolio geometry, evidence boundary and the rules under which the experiment may change.
2. Search
Machine systems process market and news evidence, estimate the ranking state and search the governed decision space at machine scale.
3. Freeze before outcome
A cohort decision is frozen before its effective market session. Later realized returns cannot be used to rewrite that decision or retune the rule that produced it.
Production ranking
A Top 30 whose signal strength depends on the evidence behind it.
The production model evaluates the governed 174-asset universe over a 60-day news window. Recent evidence receives greater weight, with a seven-day recency half-life. The system also measures how concentrated that weighted evidence is rather than treating a large raw article count as automatically independent support.
Those signals produce an evidence confidence. It is not a probability that a stock will rise. It controls how strongly the model is allowed to trust the ticker-specific prediction. Thin or stale evidence shrinks the prediction toward a neutral prior before the public ranking scores are formed.
Trump Score ranks the evidence-adjusted point prediction. Adjusted Scorestarts from that evidence-aware Trump Score and applies the existing volatility/drawdown penalty. The confidence layer is therefore inside the ranking signal, not a decorative field displayed after ranking.
From the 2026-09-28 cohort, Human Clause uses the public Top 30 as the traditional search surface. The final DTRM team remains a human decision: eleven assets in the fixed 1-4-4-2 formation.
Cohort amendment · 2026-09-28
MM1 now optimizes the actual portfolio, not an intermediate Top 25.
Earlier MM1 work used a robust Top-25 set as an exploratory bridge. That exposed a mismatch: robustness was being optimized for 25 names, while the portfolio that actually enters the game contains eleven. A later human reduction could destroy the property MM1 was meant to test. The correction was frozen before the first market session of the new cohort.
DTRM Fund · Human Clause
The machine searches the evidence-aware Top 30. The human selects the final eleven and assigns the fixed football-role weights.
MM1 Fund · Robust Machine
MM1 directly chooses eleven unique assets and their roles under the identical 1-4-4-2 geometry: 2×15%, 4×10%, 4×7% and 1×2%.
ChatGPT Fund
The memory-free advisory baseline independently selects eleven assets and its proportional weights from the governed universe.
What MM1 optimizes
MM1 uses the evidence-adjusted nominal prediction and calibrated P10 downside state for every candidate. For each possible eleven-player assignment it evaluates expected portfolio signal after a frozen budgeted-downside penalty. The uncertainty budget and tie-breaking rule were fixed before the cohort outcome.
Why the comparison is prospective
MM1 receives no invented history. Its public line is born at the 2026-09-25 close at exactly the same normalized value already accumulated by DTRM Fund: 140.310236. The MM1 portfolio becomes effective on 2026-09-28 and diverges only from subsequent market evidence.
This creates a controlled fork: same accumulated starting point, same future market path, same eleven-player DTRM/MM1 geometry, but a different selector. Historical MM1 Top-25 artifacts remain research evidence and are not rewritten.
Risk-adjusted performance
Evaluate the payoff without hiding the path.
The original since-inception risk report remains calculated in the backend from the common normalized series for DTRM Fund, ChatGPT Fund, AIEQ and the S&P 500. MM1 is deliberately excluded from that historical risk report because its legitimate observation window begins only with the prospective 2026-09-25/2026-09-28 fork.
- Maximum drawdown
- The largest observed peak-to-trough decline in the published series. It measures realized capital impairment rather than ordinary day-to-day variation.
- Calmar ratio
- CAGR divided by the absolute maximum drawdown, relating compound growth to the worst realized loss from a prior peak.
- Annualized volatility
- The sample standard deviation of consecutive daily simple returns multiplied by the square root of 252. It measures dispersion, not loss.
- Sharpe ratio · U.S. Treasury 3M
- Daily excess return uses only the latest U.S. 3-month Treasury yield known at the start of each interval, converted to a daily equivalent before annualization.
No look-ahead
Future Treasury values, later market outcomes and later news are never used to price or score an earlier decision. Stale inputs are rejected instead of silently becoming future knowledge.
Ex-ante score, ex-post evidence
Trump Score, Adjusted Score and MM1 construction are ex-ante decision layers. Realized return, drawdown, volatility, Sharpe and Calmar are ex-post observations. The two layers stay separate by design.
Game-theoretic interpretation. The experiment is a repeated decision problem in an adaptive competitive environment: machine search defines a signal, a governed selector chooses an action and the subsequent risk-return path is the observed payoff. This framing does not imply that equity markets are a literal zero-sum game or that observed returns establish causal superiority.