Start with the current audit.
Coverage is the fraction of evaluated outcomes that landed inside a published range. The observed fraction can differ from the intended fraction. A failed or unavailable check needs attention even when the overall percentage looks close to its target.
The record above audits the named cohort-band method. It does not automatically certify the simple stock search's raw historical percentiles, and it does not measure trading profitability.
Earlier experiments answer narrower questions.
One published comparison evaluated 300 setups from February through June 2026. The reported five-day ranges from Chart Library and an ungrounded language model each covered 82.7% of the observed outcomes. Their average widths differed: 10.3 versus 18.5 percentage points.
That is a result from a fixed sample and comparison protocol. It is not today's service status, a guarantee of better forecasts, or evidence that every stock comparison is useful.
Read the archived evaluation and its original figures. The archive also contains separate model-judged reasoning experiments. A judge's preference for an explanation is a different measure from the accuracy of its financial conclusions.
Read the failures alongside the successes.
The research record preserves the question, the criteria set before the test, and the result. Studies that fail, stop, or remain unfinished stay visible.
For example, Study 102 found that matching event types did not improve ranges at equal coverage. Study 103 then tested a specific widening rule for non-earnings gap days. The result belongs to that tested setting.