Read the latest research update → Short, dated explanations of published work. The study records below remain the source of truth.
14 of 14 studies · Most recently decided first
One multiplier, ×1.48, fitted on the earlier half lifts held-out gap-day coverage from 0.70 to 0.83. The under-coverage lives in non-earnings gaps and in up-gaps; earnings gaps were already covered by the earnings conditioner, so the new one belongs on gap days that are not earnings sessions.
The event-banded set is wider, not narrower, and no better at equal coverage. The finding underneath mattered more: the served analog set for an event day is the state's twins, not the event's, and the served band under-covers event days at 5 days. That under-coverage is what study 103 repaired.
At k = 200 the member distances span too narrow a range for a kernel to separate anything: the median change in width is 0.2%. Inverse-rank weighting narrows the band 11% but loses 5.2 points of coverage, failing the guard. The unweighted band stays.
Tighter neighborhoods are genuinely more similar in what followed: the dispersion ratio falls monotonically as k shrinks, at every horizon. The tails narrow; the interquartile range barely moves. The state definition carries real information, but about how much, not which way.
One arm done, the score still sealed. The arms were parked on 2026-09-03 along with the rest of the intraday program, evidence intact. The verdict will be written when the second arm runs.
Zero of 64 cell-horizons clear the gates. Pooled continuation sits at 0.48–0.50 everywhere, and the mean continuation excess is within 0.3 points of zero in discovery and mildly negative in the holdout. Direction is closed on large-move states; this adds to studies 89 and 93.
RPO remains inspectable state and is explicitly not promoted into ranking. The opened sample is spent. The source lane behind it (149 official SEC reports, hash-addressed) stays as provenance.
The event state changed the selected analog sets but not their forward coherence. Coarse fiscal-period, timeframe and reporting-lag similarity stays inspectable with zero ranking authority. The sample is spent and is not tuned.
The frozen discovery evaluation failed every gate. A deterministic semantic duplicate was detected and terminated: the lineage cannot be renamed and retried as a fresh idea.
The source census covered every reviewed buyback issuer and found no qualifying dates under the frozen definition. Stopped before an outcome task existed; no return, price-path, order or position field was accessed.
Neither the historical evidence nor the implementation costs supported the family, so it was retired as a family. Family-level retirement prevents cosmetic retries.
The attractive headline effect was rejected: underpowered, confidence bounds crossing zero, and a large pre-event placebo suggesting the move was already underway before the filing. The lineage was retired without opening 2026.
Falsified on its frozen population. No promotion, no shadow launch, no parameter change and no outcome rescue.
Falsified on its frozen 30-signal manifest. The priced-financing ordinals after it (V0.8, V0.9) were spent the same way and V0.10 is not authorized: the lineage is closed, not renamed.
The study met its stated criteria on the evaluated sample. That is not proof of trading profitability, universal validity, or future performance.
The hypothesis did not clear its decision criteria. The result remains visible, with the original question and the reasons it did not pass.
Stopped work may end at a source or feasibility check. Parked work is paused. Read each study's record before interpreting its status.
Ledger updated . This is a curated public collection, not a claim that every internal experiment is published. How the research is conducted →