Current programme snapshot: Trial 63 completed 49 evaluations, stopped the 47-row selection globally at S2, and selected no paper candidates.
Status date: 2026-09-08. This current snapshot supersedes the 5 September in-progress record. Zero candidates qualified for paper trading.
Current state
Trial 63 completed the 49-strategy daily evaluation phase. Across the wider set kept for further work, the checked inputs also accounted for all 84 ideas: 66 have implementations and 18 need specific backtester support.
The implemented set is now partitioned by actual disposition:
- 47 completed the primary common-window comparison.
- 2 completed separate coverage diagnostics outside that comparison.
- 17 large stock-universe ideas were not scored because their computation is unfinished.
- 18 ideas remain blocked by specific backtester gaps.
Thus 49 evaluations are complete and no rerun is queued for this 49-strategy research batch. That does not mean all Lab engineering work is complete. See Strategy tests and results — 8 September 2026 for the full 188-idea census and searchable status of each idea.
Selection status
The 47 primary rows all passed S1 integrity. In 200 of 252 comparisons, the strategy that looked best on one part of the history finished in the lower half on the other part. That is about 79%, against our 20% limit. It measures unreliable selection across the group, not the chance each strategy loses money. See the authors' PBO method paper.
S3 Holm, S4 clustered DSR, S5 cost stress, S6 recency, S7 one-per-family, and S8 cap were not reached. Their empty lists follow from that group stop and are not independent strategy failures. The 20% PBO limit is conservative, but ignoring it would not solve the evidence problem: the stored Holm and DSR diagnostics qualify zero of 47 strategies.
The PBO comparison used the 102 months shared by all 47 strategies, even though some individual histories are longer. The DSR diagnostic still allows for the original 84 attempts across 14 related groups. Any market-condition rule fitted to history already seen here would be exploration and would need later untouched observations.
Coverage boundary
W9-066 observed 47 months from 2018-02 through 2021-12. W9-084 observed 71 months from 2016-02 through 2021-12. Both are completed short-history checks, but neither entered the 47-strategy comparison because they could not be compared fairly over the full shared period.
Source-idea census
The programme began with 188 ideas carrying distinct catalogue IDs. Initial review kept 84 for further work and set aside 104. Of those 104, 64 need unavailable data, 7 need better data, 2 were handled separately, and 31 needed clearer rules.
Being set aside is not a zero return. An idea that is not tested yet or is blocked by the backtester has no performance result. The completed evidence does not say every source idea loses money.
Engineering status is not research evidence
The foundation for daily signals with minute-level execution is merged. Cost handling, execution resolution, the full position lifecycle, and scenario-replay proof are still unfinished. Reviewed exposure and daemon fixes passed 92 combined software tests on a separate integration branch, but they are not deployed. Those are software checks, not 92 profitable strategies.
The [ML page](/ml) separates the shared model implementation from the synthetic results attached to it. Real-data feature integration and verified model results are still missing.
At approximately 14:06 UTC on 8 September 2026, a read-only service check found both the paper runtime and the market-data ingestor stopped. No new fills were verified. The existing 16-member internal paper-family design is separate from Wave 9's zero selected candidates and does not prove that paper trading is running. Before any restoration, verify the authenticated feed, family state, and ledger continuity. This status page does not start either service.
Proposed next research sequence
None of this work has run or been approved.
- Diagnose the economics. Preserve the original result, then report actual strategy return, benchmark return, drawdown, exposure, trade count, turnover, and execution timing. The absolute results are missing today.
- Test one fixed daily condition rule. Compare one unchanged ETF strategy with the same opportunities under a lagged trend-and-volatility filter chosen before evaluation. Eight rows were positive relative to their benchmarks in 2017–2021, which motivates a stability diagnosis but does not prove those periods are predictable. Volatility-managed factor research motivates reducing risk in high volatility as a test, not as an established result here.
- Use the existing ML implementation. Compare the unchanged strategy and fixed filter with the existing regularized L2 logistic-regression trade/skip or exposure decision. Only then compare the existing shallow boosted tree on the same features and chronological partitions. Real-data feature integration and verified model results remain missing; the full portfolio must be replayed with cash, turnover, costs, and benchmark. See [ML evidence](/ml). Nonlinear asset-pricing research motivates the challenger test but does not establish an edge here.
- Test market time separately. Original 60:57 Final-Half-Hour Market Momentum needs a fixed morning signal, 15:30 decision, execution delay, and close exit. Original 128:54 Opening Range Breakout is separate. Both require minute-data feature and execution proof. Once the engine is ready, daily-signal/minute-execution wrappers W9-008, W9-018, W9-023, W9-039, W9-043, and W9-048 can compare a small declared set of delays.
- Reserve untouched observations for the decision. Freeze each change and charge every added test. Seen history may diagnose the failure; later forward evidence must determine whether an improvement persists.
The 6 September daily batch split remains historical provenance for the compute partition. The earlier 5 September programme snapshot remains intact as an archived in-progress record.
Evidence
- Manifest SHA-256:
70d41dada2b1061e8b4fe76fd2962c85e563d18525af6f5d66d87c60a9515174 - Primary batch receipt hash:
ee9800b50139aaf820d767bc1e5e610865316b495384847e340f9c60f98c0467 - Selection SHA-256:
053d8927e479329dd8aa83aa91efc8041f00323068ccbaeb030f0a5b4329855d - Final receipt SHA-256:
6ddcfaea5b83dc4f0571e6e833f0ad8ed013cd301e115f08c514cf5d4434d729