mirror of
https://github.com/luxfi/aml.git
synced 2026-09-24 00:25:20 +00:00
anomaly: attribution on the alert, and a mutation harness
The scorer's output reaches the surface with its causes, so an alert can say which features drove it rather than only that a model raised it. That is the property the whole design was chosen for: a supervisor asks why this alert fired, and a path-length model can answer where a neural one cannot. tools/mutate.py is the harness these packages were checked with. It reports ANCHOR MISS and COMPILE FAIL as distinct from KILLED, because a mutation that never applied and one that would not compile are not evidence that a test works — counting them as kills is how a suite comes to look stronger than it is. Co-authored-by: Hanzo Dev <[email protected]>
This commit is contained in:
Reference in New Issue
Block a user