Recorded engine evidence · 19 October 2026
The jobs succeeded.
The strategy needed work.
A completed backtest is only the start of a research decision. In this archived engineering case, every walk-forward test window lost money and the final report returned needs_work.
This is real archived engine output on a synthetic verification fixture. It is not a customer strategy, a performance claim or a claim about market coverage.
Download the evidence extractSuccessful execution did not turn into an approval.
All three later test periods were unprofitable.
The model was not bound to the selected validation configuration.
Follow the evidence
Start with the verification fixture
The engineering workflow used a synthetic strategy to exercise backtests, studies and report publication. The selected configuration used a predeclared entry cadence. This was not an attempt to demonstrate a tradable edge.
Read what selection actually did
Each walk-forward window recorded one training trial and a
frozen_fallbackselection. That is a significant limitation: these receipts do not establish that a broad parameter search found a stable region.Inspect the later test windows
The three windows contain 380 trades in total. None was profitable. The selected public extract omits prices and monetary amounts; the outcome remains visible.
Keep the unfavourable result
The final report recorded
needs_work, with no failed or missing studies. That distinction matters: a study can finish correctly and still give you a reason to send a strategy back to research.
The walk-forward record
| Window | Test trades | Profitable | Selection |
|---|---|---|---|
| 1 | 133 | No | frozen_fallback |
| 2 | 114 | No | frozen_fallback |
| 3 | 133 | No | frozen_fallback |
The method version is validation-walk-forward-v1. The three fallback selections and three total training trials are included in the download.
What the ML result means
A separate ML comparison on the same engineering fixture returned no_useful_uplift. It was not bound to this selected validation configuration. The evidence does not justify claiming that a model improved this candidate.
That result belongs beside the other findings. It should not disappear simply because a different number looks better.
Provenance and limits
Evrant engineering recorded this case in the engine M1 verification archive on . This page was prepared on from the retained report verification and walk-forward receipts.
The archive records a match between the report’s published hash and an independent hash:
1cb40e76c32468675abd94fbb2d664bd0310e7b61c1e958da8fbd85e63f90d99The download contains selected fields and the SHA-256 digests of the source archive files. It is not the full report or a rerun package: you cannot recompute the original report hash from this extract.
Raw market data, licensed inputs, strategy source, binaries, account identifiers and private invocation settings are excluded. No customer research was used. These results describe a dated verification case, not the latest state of the service or a future trading outcome.