F12

Explicit lever and recipe arms

The source summary is shown as authored. Follow its evidence links for study-specific status and cohort limits.

Source-reported narrativeValues in this source section that are not tied to a claim-ledger record remain source-reported and are not promoted to verified results.

Findings source section · FINDINGS.md

Hypothesis: H117–H139 ask whether specific flags, stage recipes, or bundles affect completion, cost, or time. A flag or recipe name alone does not establish that a distinct arm ran.

What ran: The lever register distinguishes bundled historical records, R51 planning arms, the R54 context handoff, R56 and R57 comparisons, cascade rounds, hill-climb iterations, R72, and withdrawn R74. L0 reconstructs R57's scoped 311-row pooled population and R72's P1 and partial P3 counts; R72's rows do not match the canonical public outcome export, and one planned R57 slot remains unidentified. See CL-WA-r57-source-reported-denominator-variants (INTERIM) and CL-WA-r72-lifecycle-and-outcome (INTERIM). The cohorts remain separate.

Result: Numerator/denominator: not recorded for a complete, attributable lever-by-lever comparison. Unit: official Build pass, cost, and wall time. Date: no common comparison date recorded. Status: many levers are not established as distinct runs; the linked rounds remain interim, descriptive, or withdrawn as noted on their pages.

What it does not show: A bundled or source-reported result does not establish a transferable isolated-factor effect for a named flag or recipe.

Evidence: F12 hypothesis and lever register, R51, R54, R56, R57, L0 reconciliation, R49, R49b, hc-1, R72, R74. Spec map: Shaping rules, ladder recipe, and prompt caching.