CORTEX v2 is an embedded AI orchestrator that reads your Guardian events, auto-tunes routing thresholds, learns your optimal model mix per task type, enforces sovereignty compliance - and lets you chat with it about your own data.
Guardian is reactive - it blocks problems as they happen. CORTEX is proactive - it reads the patterns and acts before something breaks.
Anomalies fire into ClickHouse. You'd need to query logs manually to see patterns. CORTEX reads them for you and classifies what matters.
Default heavy_threshold=3.0 was calibrated on average traffic. Your workload evolved. Nobody noticed. CORTEX recalibrates every week.
Cost spikes reach your Slack 6 hours after they started. CORTEX surfaces a Critique card within seconds of a Guardian event and sends it to your preferred channel.
Each phase runs on a different cadence and uses the cheapest model that can handle the job. Total cost: under $0.30 per workspace per month.
Reads the raw Guardian alert, classifies severity (info / warning / critical), and decides whether to notify immediately or batch. Filters noise - not every anomaly is worth a ping.
Audits the last 24h of routing decisions. Detects heavy-tier creep, provider cost drift, cache hit regression. Produces a daily digest card in the Inbox if anything is off.
Weekly strategic report: spending trends, routing efficiency score, comparison to anonymised peer percentiles, top 3 actionable recommendations. Delivered as a rich Inbox card you can act on with one click.
Takes the Strategic Advisor's recommendations, proposes concrete parameter changes (routing thresholds, Guardian sensitivity), creates a versioned entry with rollback support, and applies it - or asks for your approval first.
Detects patterns across all opted-in workspaces: emerging model quality issues, provider outage precursors, cost anomalies that correlate across users. Surfaces fleet-wide insights only visible at scale.
Analyses 30 days of usage_events per task type (code / reasoning / creative / tool_use / chat). 10% of responses are scored by a Haiku judge for real output quality - not just HTTP success. The optimizer combines error rate, retry rate, and judge score into a quality_proxy, then picks the best model per task. Your key, your cost, visible in CORTEX costs.
Sets a hard sovereignty_zone constraint (global / eu / fr / strict) on your workspace. Only models whose data residency matches your zone are eligible - the scorer never sees the others. GDPR, DORA, NIS2, HDS compliance baked in as a filter, not a suggestion.
Most routing proxies apply a single tier threshold to everything. CORTEX v2 goes further: it tracks performance per task type (code, reasoning, creative, tool_use, chat) across your actual usage, scores every model on quality + cost + latency, and writes a workspace-specific routing config. Each 'auto' request is then matched to the best model for that kind of work.
code / reasoning / creative / tool_use / chat - each task type gets its own best model. A cheap fast model for chat, a capable one for multi-step reasoning.
savings / balanced / quality_first / agent / batch - you choose the trade-off curve. The optimizer adjusts quality vs cost vs latency weights accordingly.
Set your zone (global / eu / fr / strict). Only models with matching data residency enter the scoring - the rest are invisible to the optimizer.
Scoring formula
score = w_quality × quality_proxy − w_cost × norm_cost − w_latency × norm_p95
quality_proxy = 1 − error_rate − 0.3 × retry_rate + 0.3 × judge_score. judge_score: Haiku rates 10% of your real responses 1-10 (your Anthropic key). Fallback to user_feedback signal when no judge data. Costs and latencies are min-max normalised within the candidate set per task type. Requires ≥ 100 requests in the last 30 days to activate.
Typed cards pushed to your dashboard, email, Telegram, and Slack. Every card has an action button. Critical alerts have rollback buttons.
Agent runaway - 47x normal spend
Workspace acme-prod hit $43 in 12 minutes. Guardian blocked after 6 loops. Root cause: recursive tool call on vision model.
Raise heavy_threshold 3.0 → 3.4
14% of heavy-routed requests in the last 7 days scored 3.0-3.4 but received adequate quality from Standard. Estimated monthly saving: $28.
Weekly digest - May 5-11
Routing efficiency 8.4/10 ↑0.3. Cost $312 vs $541 baseline (42% saved). Ultra-light routing captured 38% of traffic. 0 Guardian incidents.
Every parameter change is versioned + rollback in 1 click
AI speaks your language - FR, EN, ES auto-detected
All changes require your approval before applying
Guardian Triage fires in < 5s after an anomaly
Scale gets CORTEX Inbox alerts. Enterprise unlocks the full brain with Network Intelligence and custom tuning.
| Feature | Scale | Enterprise |
|---|---|---|
| Guardian Triage (real-time) | ||
| Workspace Intelligence (daily) | ||
| Strategic Advisor (weekly) | ||
| Self-Tuning Engine | ||
| Network Intelligence | ||
| CORTEX Inbox (alerts only) | ||
| Parameter versioning + rollback | ||
| Multilingual AI outputs | ||
| Mix Optimizer (per-task-type routing) | ||
| Sovereignty zone filter (EU/FR/strict) | ||
| CORTEX Chatbot (on your own data) | ||
| Market Intelligence (new model alerts) | ||
| Custom tuning parameters |
Scale gets CORTEX Inbox alerts only (Guardian Triage notifications, no auto-tuning).
Upgrade to Scale or Enterprise to activate CORTEX. First insights appear within 24 hours. ~$0.30 / workspace / month.