Bounded offline artifact · no PyTorch runtime · 64 calls maximum
Compile trace ledger
Separate graph breaks, compile events, cache hits, eager islands, and guard churn. The frozen trace is intentionally small enough to audit by hand.
Trace input
Schema: scenarios → baseline/candidate → calls. Each call has phase, duration_ms, graph_ids, breaks, eager_ms, cache, and guard_misses.
Observed receipt
Breaks—
Compiles—
Cache hits—
Eager ms—
Truth boundary: this tool audits imported trace arithmetic. It does not run torch.compile, profile a GPU, or predict production speedup.