CONTINUUM · TIME REPLICATION
5 WORKFLOWS · 5 ARTIFACTS · 1 FIXED POPULATION
Temporal robustness, not one lucky run
The memory firewall survived five real-provider time clusters.
The exact same 36 synthetic incidents were replayed in five separate main-only OIDC workflows. Every batch used real Bedrock proposals, CockroachDB outcome promotion, disposable GitHub draft effects, provider receipt lookup, and cleanup reconciliation.
Workflow batches—distinct run receipts
Paired incidents—360 arm observations
Continuum lift—percentage points
95% cluster bootstrap—workflow cluster + paired case
Continuum success—provider-verified outcomes
Raw-RAG success—same time clusters
Continuum unsafe—proposals / exposures
Residual effects—duplicates + drafts + leaks
CHECKINGBinding the aggregate to the public judge receipt…
| Batch | Start | Continuum | Raw-RAG | Lift | Workflow |
|---|
Statistical boundary: these are 180 paired executions but only 36 distinct incident definitions, repeated across five workflow time clusters. The 180-pair exact p-value is descriptive only. The primary interval resamples workflow clusters and paired cases hierarchically.
Run the full one-click judge verifier → · Inspect one 36-pair batch →