Evidence first
v4 · 18 Aug – 17 Sep · Evidence first · see v6, the current engine →
With no model in the loop, v4 spent a month testing whether any of the strategies were real. Almost none were, and most of the month went into showing that carefully enough to believe it.
Figure 01 · v4 · 18 August – 17 September 2026
No model, stricter evidence
The agent layer is gone; grids, folds and frozen monitors do the judging.
- Data
- Orders
- Next-open fill
- Verdict
Grids instead of a model
The walk-forward went from six folds to ten, because a single window could flip a verdict. The sweep farm tried 51 candidates without a single model call. Live results could not help either: no book had reached a t-statistic of 2, so none was tuned on them.
The harder finding was the ceiling. Of 30 swept candidates, 26 were indistinguishable from their benchmark and none beat it, and the test only had even odds of spotting an edge near seven points a year. Smaller edges were invisible to it. Because the history holds only today's survivors, more compute would not have fixed that; it is a limit of the data.
The audit of 2 September
A full read of the simulator found seven bugs. The worst was that no dividend had ever been credited, which a backfill fixed with 26 credits worth $230.42. Monthly books would have skipped a month that ended on a weekend, and there were $0 fills, a look-ahead rollback and risk gates that trusted a typed price. The split checker had also restated three real crashes, BH, ORCL and NEM, as splits, and a re-audit later reverted 18 more. The engine had 228 tests after that audit and over 2,200 two weeks later.
Monitors that cannot promote
Fill model v4 added named execution profiles, so any result can be rerun at double cost or under a participation limit. Doubling costs at $39,000 barely moved SPY but took 1.85 points a year off the stock basket. By September three forward monitors were running, each frozen before its first signal: sector momentum against SPY, cross-sectional 12-1 momentum against equal weight over 48 paired months, and the Monday experiment, which needs 40 Mondays and ends in May 2027. None of them can change a book; they can only report whether it should continue.
By mid-September the engine was honest and very large. My own review on 18 September gave it 6 out of 10, with research output and scope control both at 3. I decided the next version would start with limits on the engine itself before any agent came back.
- dates
- 18 August – 17 September 2026
- model calls in the loop
- none
- price sources
- Yahoo only · Nasdaq API as verifier, not failover
- books
- 25 → 18 after the AI layer, 21 active by the end
- walk-forward
- 10 folds, up from 6
- sweep farm
- 51 candidates, no model calls
- forward monitors
- sector momentum · XS 12-1 momentum · E1 Mondays
- fill model
- v4 · profiles baseline, 2× cost, participation stress
- tests
- 228 on 2 September → 2,261 by mid-September