SCFCE-R observed-cascade external stress test
================================================

Dataset:
- Twitter15 and Twitter16 public propagation benchmarks.
- Node tuples in the release are [uid, tweet ID, delay].
- The executed v4 benchmark projects propagation records to UID nodes because
  SCFCE-R intervenes on users.
- Repeated events from a user consolidate to one UID node; duplicate edges
  collapse in a directed simple graph; genuine UID self-loops are removed.
- This is a user-projected observed propagation graph, not the original event tree.

Frozen protocol:
- PAPER_MODE=True
- primary model seed 42
- weights 0.8 GAT / 0.1 betweenness / 0.1 degree
- risk gate 0.55
- harmful risk 0.85
- neutral reliability placeholder 0.50
- false-only primary labels
- primary budgets 5% and 10%
- 36 external paired hypotheses, globally Holm-corrected
- 10,000 bootstrap resamples
- 50,000 random-sign permutations

Primary conclusion:
The frozen SCFCE-R policy does NOT transfer successfully. It is worse than all
nine deployable baselines on both datasets at both primary budgets, including
random intervention. The negative result is retained intentionally.

Scientific boundary:
This benchmark is a retrospective structural counterfactual. It does not
identify the causal effect of a real platform moderation action and does not
include a calibrated upstream misinformation detector.
