diff --git a/src/tracegrad/pipeline.py b/src/tracegrad/pipeline.py
index 816a856..b5d50e5 100644
--- a/src/tracegrad/pipeline.py
+++ b/src/tracegrad/pipeline.py
@@ -1,10 +1,10 @@
 """Run orchestration: the fixed order the stages execute in.
 
 The pipeline is the only place that knows the whole sequence, and it holds the
-run lock while it executes.  Each completed stage writes a resume checkpoint, so
-a killed run picks up where it stopped instead of re-spending the attribution
-budget — the attribution cache does the real work there, and the checkpoint just
-records what has already been paid for.
+run lock while it executes.  Each completed stage writes a resume checkpoint for
+inspection, but a re-run still executes every stage from the start — the
+attribution cache, not the checkpoint, is what avoids paying twice for the same
+instrument.
 
 The pipeline never writes the prompt.  It produces a proposal; ``apply`` is the
 only writer.