README: link 2026-09-13 single-weight corruption result
Browse files
README.md
CHANGED
|
@@ -134,6 +134,19 @@ and get a different digest. If the disagreement is because the spec text permits
|
|
| 134 |
two readings, that is the finding most worth paying for: it means the document
|
| 135 |
is not yet sufficient, which is the entire thing CIS-2 claims to be.
|
| 136 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 137 |
---
|
| 138 |
|
| 139 |
Aefinity AI Inc. · Justin Brian Thompson
|
|
|
|
| 134 |
two readings, that is the finding most worth paying for: it means the document
|
| 135 |
is not yet sufficient, which is the entire thing CIS-2 claims to be.
|
| 136 |
|
| 137 |
+
## Results: single-weight corruption detection (2026-09-13)
|
| 138 |
+
|
| 139 |
+
Does a single flipped weight bit in a deployed model get noticed by text
|
| 140 |
+
output, a cheaper top-k logit check, or the CIS witness digest? Of 247
|
| 141 |
+
single-weight bitflips on a pruned BitNet-2B that changed the model's
|
| 142 |
+
output, text/argmax replay missed 182 (74%, CI 68-79%), a top-k logit-set
|
| 143 |
+
check missed 82 (33%, CI 28-39%), the CIS witness digest missed 0 (CI
|
| 144 |
+
0-1.5%). 43 other flips produced a bit-identical digest to baseline (no
|
| 145 |
+
real effect); the digest correctly passed all of them. Limits: single
|
| 146 |
+
prompt/model, no adversarial tampering, most embedding-row hits are
|
| 147 |
+
integrity-only. Full report:
|
| 148 |
+
https://github.com/Aefinity-AI/cis2-spec/blob/main/docs/results/2026-09-13-BITFLIP-HEADTOHEAD.md
|
| 149 |
+
|
| 150 |
---
|
| 151 |
|
| 152 |
Aefinity AI Inc. · Justin Brian Thompson
|