Skip to content

ce-retune: distinguish one streak's probability from evidence after candidate selection #1698

Description

@22nsuk

skills/ce-retune/references/noise-floor.md describes p^N as an exact test after estimating the baseline from A/A and archive runs. It also permits another streak after changing a failed candidate. The user guide repeats the one-armed-test claim.

For a fixed known p, independent runs and one pre-specified attempt, the calculation is valid. It is not automatically a significance level for a selected winner after repeated attempts, and an independently estimated baseline still has uncertainty.

A numerical counterexample (not measured CE performance): with p=0.58 and N=8, one planned all-success attempt has probability 0.0128063. Across 20 independent attempts under that same null, the probability of at least one success is 1-(1-p^N)^20 = 0.227235. Changing the candidate does not erase earlier selection opportunities.

Keep the operational streak bar, A/A measurement, fixed-build confirmation and stop-on-failure behavior. Qualify the statistical interpretation, retain all attempts, and distinguish candidate-selection evidence from a final check whose results were not used to select the candidate. When the assumptions or overall error control are unsupported, report the observed confirmation without an exact-significance claim.

Scope: the streak guidance and its user-facing explanation only; no new statistical engine, dependencies, fixed sample minimum, or changes to ce-optimize.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions