dsh-rerun-report

Everything between "a pile of attempt records" and "a decision a team can act on": descriptive aggregation, paired comparison with a measured noise floor, gates, trajectory explanation, Markdown and HTML rendering, and trend.

Three commitments shape the statistics:

• The resampling unit is the task, not the attempt. A run generalizes to "tasks like these", so the uncertainty that matters is between-task variance.
• Comparison is paired. Task difficulty dwarfs harness effects, so the statistic is the per-task difference, which cancels difficulty exactly.
• The bootstrap is seeded. A report that reshuffles its own confidence interval on re-render is not evidence.

UNDERPOWERED is a distinct verdict from NEUTRAL, because "no difference detected" and "this run could not have detected one" are different claims.

A library, not a Cordis plugin. Every entry point is a pure function of stored records, which is what lets the same code render a report at the end of a run, re-render one from the store months later, and compute a trend across many runs without any of them booting a tree.

Model Experience

None. This package never contributes to a model request: it runs in the orchestrator process, and the model calls it cares about happen inside a separate attempt subprocess running a different build of DSH entirely.

KV Cache effect

None; it neither assembles nor sends a provider request.

Known Limitations and Deferred Work

• The bootstrap resamples tasks only. Within-task repeat variance enters through the cell value rather than through a second resampling stage; a two-stage bootstrap is deferred until a suite is wide enough for it to matter.
• An unmeasured noise band defaults to zero, which flatters every difference. The renderer says so above the tables, but nothing refuses to render.
• Gates are fixed. Their thresholds are configurable; which gates exist is not.