Week 13 · Analysis and communication
The aim is simple: complete this week's work, record what proves it is finished, and make the decisions required before moving on.
Next step
Evidence to produce: A reproducible conditional primary estimate/interval and positive, null or contrary conclusion.
Decision needed
The conditional primary interval and all prespecified descriptive outputs regenerate from the locked records.
Deliverable
Reproducible conditional interval, absolute accuracy/degradation curves, seed/cell/action sensitivities, resource tables and report/poster draft.
Execution
Open a task for its instructions and the evidence needed to count it as complete.
Run 10,000 within-action performance-cluster bootstrap replicates with fixed seeds/actions and query-weighted multiplicities. Report Delta and its 95% percentile interval; claim smaller degradation only if the upper bound is below zero. Use the pretested v3 report dispatch; an interval containing zero is inconclusive, not equivalence or absence of an effect.
A reproducible conditional primary estimate/interval and positive, null or contrary conclusion.
Show clean and absolute corrupted accuracy, Contextual-minus-Contrastive corrupted differences, every seed/cell effect, SupCon, exact-official results and 20 leave-one-action-out sensitivities with a fixed gallery. Show the clean/corrupted gap decomposition beside Delta in percentage points and per-1,000-query descriptive scale; no new practical-effect threshold. Keep residual participant dependence and fixed-seed/action/gallery conditioning beside the interval.
All prespecified tables and curves, including sign changes and conditional limits.
Explain the frozen encoder, fixed transfer recipes, nested synthetic corruptions, selection, test seal, resource costs and retained failures. State that margins/regularization are part of the recipe and that v3 does not complete the v2 benchmark.
A report/poster draft with claims traceable to locked evidence and honest scope limits.
Reproduce every table and figure from immutable ranks/manifests, resolve presentation/reproducibility defects and explicitly label any post-opening scientific deviation.
A reviewed reproduction log and permitted compact final evidence package.