diff options
Diffstat (limited to 'protocol/protocol.py')
| -rw-r--r-- | protocol/protocol.py | 10 |
1 files changed, 7 insertions, 3 deletions
diff --git a/protocol/protocol.py b/protocol/protocol.py index 4298756..185c4f2 100644 --- a/protocol/protocol.py +++ b/protocol/protocol.py @@ -166,9 +166,13 @@ def cross_batch_direction_stability( the same regardless of which samples are in the batch. That is the fingerprint of a **sample-invariant global drift** — i.e. the reference vector against which Γ is computed is NOT per-sample credit, it is a - constant artifact of the trained network's geometry. On healthy - BP-trained or EP-trained networks this value is small (~0.05-0.18); on - DFA/SB/CB pre-LN ResMLPs we see ~0.5-0.99. + constant artifact of the trained network's geometry. In our 5-method + CIFAR-10 audit at K=10 batches of 128 samples, healthy BP/EP networks + cluster near zero with all six values in [-0.04, +0.12]; drift-dominated + DFA/SB/CB pre-LN ResMLPs reach high tails up to +0.99 (5/9 of the nine + degen values exceed the 0.30 default cutoff). See + results/protocol_audit/audit_table_s42_s123_s456.json for the + per-method per-seed values. Args: model: same contract as the other diagnostics. |