This directory freezes the v0.6 confirmatory study before new observations are collected. The machine-readable protocol binds the study dimensions, 33 repetition seeds, primary comparison family, missing-run policy, and analysis.
The repetition count comes from the checked-in v0.5 observations:
The power calculation is a conservative normal approximation used only for planning. The frozen primary test is the two-sided exact repetition-cluster sign-flip test with Holm correction. Five pilot repetitions estimate variance imprecisely, so the final report must include achieved power and cannot generalize beyond the declared synthetic tasks.
The 33 scheduled repetitions are not replaced or selectively rerun after a cell-level infrastructure failure. Up to four missing repetitions are covered by the planning reserve; fewer than 29 analyzable repetitions makes the affected comparison inconclusive. Each task has one bounded attempt, and the result records that attempt count.
Verify the frozen files:
python examples/agentcert-bench-24-task-study/verify_preregistration.py \
--protocol benchmarks/preregistrations/agentcert-bench-v0.6/protocol.json \
--digest benchmarks/preregistrations/agentcert-bench-v0.6/protocol.sha256 \
--power-plan benchmarks/preregistrations/agentcert-bench-v0.6/power-analysis.json \
--source-analysis examples/agentcert-bench-24-task-study/results-v0.5/analysis.json
The protocol digest is evidence that a later result used the frozen design; it is not proof that the study operator followed it. Signed evaluator receipts, run manifests, and the deviation log provide the complementary execution evidence.