L1
L1 Analyst
SOC · New York

Platform Health

Operations evidence behind the AI-SOC layer — scale triggers (§8.1), model & prompt registry, redaction precision/recall and hallucination findings (NIST AI RMF Measure).

Scale triggers
1
watch / trigger
Production models
6
8 total registered
Canary trafficShare of traffic routed to the canary advisor model. Increase only after eval set passes.
10%
altisrc-soc-mistral-7b-v3.3
Override rate (7d)
4.2%
analysts override AI verdict
Grounded citations
96.4%
AI citations that resolve
Hallucinations open
1
awaiting review

Scale triggers — Phase 0 → Phase 1 baselineOne L4 GPU + Redis Streams + vLLM is the baseline. Each metric is tied to a documented trigger from §8.1. Multi-GPU / Kafka introduced only when validated by these signals.

Auto-refresh 30s
MetricCurrentSLO / capTrend (8h)Trigger conditionStatus
Redis queue depth12 items100 itemsSustained > 100 → add AI worker
Healthy
Case → brief p95 latency6.4 s8 sp95 > 8s while GPU > 80% → add 2nd GPU
Watch
L4 GPU utilisation64 %80 %Sustained > 80% with rising queue
Healthy
vLLM tokens / sec1,820 tok/s1,200 tok/sFloor 1200 — alert if drops below
Healthy
Worker count (active)3 pods6 podsAuto-scale to 6 on queue pressure
Healthy
DLQ depth (24h)1 msgs10 msgs> 10 in 24h → ops page
Healthy
Kafka adoption trigger0 %100 %Replay / multi-consumer demand observed
Healthy

Model registryEvery component that produces AI output is versioned, owned, and promotable. Canary runs in parallel with production for evaluation before cutover.

Prod Canary Shadow Retired
ComponentVersionStatusPromotedOwner
AI Triage Advisor (LLM)altisrc-soc-mistral-7b-v3.2 Production2026-05-22Nexus AI
AI Triage Advisor (LLM)altisrc-soc-mistral-7b-v3.3 Canary2026-06-04Nexus AI
Risk Scorerrisk-xgb-v1.7 Production2026-04-18AI-SOC Eng
Correlation Agent (rules)corr-rules-v2.4 Production2026-05-30AI-SOC Eng
MITRE Mappermitre-bert-v0.9 Production2026-05-12Nexus AI
Redaction Engineredact-regex-v4.1+presidio Production2026-06-01Privacy Eng
RAG Indexrag-index@2026-06-04 Production2026-06-04AI-SOC Eng
AI Triage Advisor (LLM)altisrc-soc-mistral-7b-v3.1 Retired2026-04-22Nexus AI

Prompt registryVersioned system / user prompts. Hash + author let any AI output be traced back to the exact prompt that produced it.

  • case-brief.systemv12
    sha256:7a3f…b2
    2026-06-02 · K. Yamamoto
  • kill-chain.userv8
    sha256:2dc1…81
    2026-05-28 · K. Yamamoto
  • fp-likelihood.userv5
    sha256:9b04…ee
    2026-05-19 · M. Okafor
  • handoff-draft.userv4
    sha256:0f7a…3c
    2026-05-10 · M. Okafor

Redaction engine — precision 97.6% · recall 94.8%Privacy guardrail evaluated against a labelled PII test set. Recall measures missed PII (false negatives = privacy leak risk). Precision measures over-redaction.

Eval set: 1,240 · last run 2026-06-07
SSN
Precision
98.4%
Recall
96.9%
Loan #
Precision
97.8%
Recall
94.2%
Account #
Precision
97.1%
Recall
93.6%
Email
Precision
96.5%
Recall
92.8%
Name
Precision
94%
Recall
89.4%
Recall < 95% on any PII class blocks promotion of a new redaction engine version.

Hallucination findings (1 open)Cases where AI output diverged from evidence. Detected by citation grounding, analyst override, or eval-set replay.

  • AI cited SOP-DLP-004 §2 — section does not exist
    2026-06-08 08:41 IST · Citation checkOpen
  • MITRE T1572 mapped where evidence points to T1567.002
    2026-06-07 17:12 IST · Analyst overrideAcknowledged
  • Confidence 92% but only 1 weak signal — overconfident
    2026-06-06 20:38 IST · Eval setFixed
  • Suggested closure as FP while DLP marked QUARANTINED
    2026-06-05 13:24 IST · GuardrailFixed
Scale-out is trigger-based — Kafka and multi-GPU are introduced only when these metrics prove the need. Every AI output is traceable to a model version + prompt version + RAG index snapshot.