Forward test rules
Locked by the board. A change means a new test with a new version number.
Locked rules (from ops/forward-test-rules.md)
Forward test v1: rules
Status: APPROVED by the board on 2026-10-14 and locked. A change means a new test with a new version number.
Purpose
Build a track record that can't be fitted to the past. It's the evidence that will still count after AI joins, because no model can know these days in advance.
Start
The first trading day after the Phase 1 data fixes (T3 through T8) are merged and verified. Before day one, the CEO records the start date and the code commit here.
Strategies, frozen at the start
| # | Strategy | Source |
|---|---|---|
| 1 | SPY | benchmark |
| 2 | Textbook momentum | genomes/baselines/base-momentum-12-1.yaml |
| 3 | 20% SPY core plus momentum | genomes/baselines/base-core20-momentum.yaml |
| 4 | 60/40 stocks and bonds | genomes/baselines/base-sixty-forty.yaml |
| 5 | The S&P 500 champion from the first evolution run after the fixes | its genome file, copied into forward/ |
Strategies added later (for example the blended team, T19) start their own clock on the day they're added; they're never backfilled.
Each strategy starts with $10,000 of paper money in our ledger, pays the same trading costs as in backtests, decides after each close using only that day's data, and fills at the next open. S&P 500 only for v1; the other arenas join once their member lists are point-in-time.
What "frozen" means
No change to a strategy's settings during the test. If an engine bug turns up that changes results, the fix is applied, the whole forward period is replayed from day one with the fixed code, both versions are kept, and the change is logged in ops/DECISIONS.md.
Records
- The daily ledger and equity of every strategy, committed to git by a workflow.
- Every Monday, the standard forecasts in
ops/FORECASTING.md, committed before their windows start.
Checkpoints and the verdict
At 3 and 6 months we report results only; nothing is decided except the kill switch below.
At 12 months, each strategy except SPY gets one of two verdicts:
- Promising, if all of these hold:
- It beat SPY after costs.
- It beat textbook momentum (not applicable to momentum itself).
- It ranked at or above the 90th percentile of 300 fair random 5-stock portfolios over the same 12 months.
- Its worst drop was no more than 10 points deeper than SPY's.
- No evidence, otherwise.
One year can't prove skill (see "Why a backtest can't prove skill" in the review), so Promising means "keep testing", not "proven".
Separately, a strategy's forecasts count as calibrated if their mean Brier score is below 0.25 (better than a coin flip) over at least 30 resolved questions.
Kill switch
A strategy that falls 35% from its peak stops trading and stays in the record as stopped.
AI
AI can join whenever the board decides (D5). Anything an AI proposes is judged by these same rules, from the day it's proposed, against these frozen strategies.
Start record (CEO, 2026-10-15)
- Start date: 2026-10-16 (first trading day after T3 through T8 were merged and verified; T7 verified by a green real-data sp500 run on 2026-10-15).
- Code commit: a025350fc418c16162732e8410691b016e8b7f76.
- Champion:
forward/champion-sp500-20261014-151728-1.yaml, the top S&P 500 champion of run sp500-20261014-151728, the first evolution run after the fixes. On its unseen holdout (2026-04-16 to 2026-10-14) it returned 0.5% vs SPY 10.3% and textbook momentum 8.6%, percentile 31 against fair random portfolios: no signal. It is frozen anyway, as the rules require.