⚠
Incident severity & notify classifier
SecurityClassify a security incident into a severity level and the notifications it triggers — GDPR 72h, customers, regulators — fast, at 3am.
The routing recipe
Job typeClassification / routingPolicyFastest that passes
routes to
DiffusionGemma 26B-A4B nvidia
nvidia:google/diffusiongemma-26b-a4b-it
93.4
match score
84.3
task fit
100
cost
97.7
speed
Why this model from benchmarks & capabilities
- "Classification / routing" leans hardest on instruction; DiffusionGemma 26B-A4B scores 76/100 there — #40 of 54.
- Strength on this class of work shows up on IFEval, MMLU (zero-shot), where Google models are competitive.
- Cost: runs on NVIDIA free credits / self-hostable NIM — effectively $0 marginal. Speed: ~1100 tok/s (estimated).
- Chosen over GPT-OSS 20B for cheaper & faster.
| Capability the job needs | Weight | DiffusionGemma 26B-A4B | Catalog rank |
|---|---|---|---|
| instruction | 0.90 | #40 of 54 | |
| speed | 0.90 | #1 of 54 | |
| knowledge | 0.30 | #39 of 54 |
Relevant benchmarksIFEval MMLU (zero-shot)
EconomicsCost: runs on NVIDIA free credits / self-hostable NIM — effectively $0 marginal.
Speed: ~1100 tok/s (estimated). vs runner-upChosen over GPT-OSS 20B for cheaper & faster. (GPT-OSS 20B, score 89.3)
Speed: ~1100 tok/s (estimated). vs runner-upChosen over GPT-OSS 20B for cheaper & faster. (GPT-OSS 20B, score 89.3)
Example result
Sample input ▾
An employee laptop with a synced copy of a customer export (names + emails, no financial data) was stolen; the disk was unencrypted.
Severity: SEV2 — confirmed loss of personal data, limited sensitivity, bounded scope.
Personal data: Yes (names + emails).
Notifications: GDPR — likely notifiable to the supervisory authority within 72h (assess risk to individuals; low-but-not-negligible); affected customers — prepare a notice, send if risk is more than minimal; law enforcement — file a theft report.
First action: remote-wipe the device, rotate the employee’s credentials/sessions, and start the 72h clock now. Escalate to DPO/legal.
Test it on your own data