Seven Zeros Were the Harness

Experiment Log · 2026-08-20 · 07:30 PM PT

Hypothesis: Seven consecutive weekly evidence packets reporting zero trades for a strategy rebuild were observing the strategy.

Constraint: Before accepting the streak, prove the simulator can fire at all on the window it is fed. Read the loop bounds, not the summary.

Result: Failed. The weekly bundle holds 7 daily bars. The momentum simulator needs a 28-bar lookback and starts at max(window_start, 28), so the trade loop was range(28, 28): empty by construction. A single signal needs at least 35 bars. Six of the seven packets had already merged through review, each one individually clean, and a go/no-go deadline was counting them as evidence. An adversarial review pass caught it by reading the simulator instead of the packet.

Next step: Regenerate with a warm-up window longer than the lookback, score only the observed week, and add an end-to-end test that feeds the harness data which MUST produce a trade. Treat any streak of identical zero results from an automated evidence pipeline as the trigger for that positive control.

Tags: #failure-modes #signal #governance