How the honesty works
The marketing page says Cortex remembers, correlates, and cites its work. This page is for the engineer who wants to know how — the actual algorithms, the actual output, and where the numbers come from. No hand-waving.
Institutional memory is a retrieval problem
Structural recall, not keyword search
Every resolution is a structured artifact: symptoms, root cause, affected systems, the fix, the author, the timestamp. When a new incident opens, recall runs Jaccard similarity on symptom tokens plus semantic embedding against that corpus — a structural match on what actually happened, not a keyword grep and not an LLM guessing from vibes. Matches are scored at citation time (strong / moderate / weak), and the score travels with the citation forever.
Topology is a graph database, not a diagram
The dependency map is built from how your infrastructure actually fails — alert co-occurrence, metric correlation, declared relationships — and stored as typed edges in the knowledge graph. On PostgreSQL, Cypher queries via Apache AGE traverse "disk full on storage-array-7" to "$15K/hr revenue exposure" in milliseconds, and every hop in the answer cites the edge it crossed. No AGE available? The same queries fall back to recursive CTEs. Same answers, same citations.
Anomaly detection costs zero tokens
Baselines use Welford's online algorithm — mean and variance updated in constant time with every sample, from the very first data point. No training window, no nightly batch job, no GPU, and no LLM call to score a metric. Sigma thresholds (2σ warn, 3σ critical) are deterministic and explainable to an auditor. AI is reserved for the problems where it genuinely outperforms — diagnosis narrative, log extraction, risk scoring — never for arithmetic.
Confidence is earned, not asserted
Every remediation pattern carries a confidence score in [0.1, 1.0]. Success moves it up; failure moves it down; operator feedback (approve, reject, annotate, suppress) is first-class input. Automation is gated on that score per system — above your threshold it fires and records the outcome, below it suggests and waits. Every action writes a tamper-evident audit entry with a reversible diff. Progressive trust is a mechanism here, not a slogan.
The agent looks before it thinks
Before our diagnostic agent touches an alert, it reads a deterministic briefing: every release executing in the blast radius, every change touching the affected stack in the last two hours, every active engineering session, every neighbor incident on the topology, every active maintenance window. The agent doesn't have to remember to check — the briefing always runs first, costs zero LLM calls to assemble, and every finding cites its source record.
=== INVESTIGATION BRIEFING ===
System: svc-online-banking | Alert: high-latency | Window: 2h
ACTIVE RELEASES IN BLAST RADIUS (1):
- REL-00004: 2026-Q2 core-banking batch — currently executing
(bundles 3 changes) [LIKELY CAUSE]
RECENT CHANGES TOUCHING THIS STACK (2):
- CHG-00123: Patch svc-branch-print-01 (implementing,
risk=moderate) [LIKELY CAUSE]
- CHG-00124: Rotate session-cache Redis creds (scheduled, risk=low)
ACTIVE ENGINEERING WORK (1):
- alice: Refactor payment pipeline (in-progress)
ACTIVE MAINTENANCE / BLACKOUT WINDOWS (1):
- Month-End Close (4 systems) — ends 2026-04-30 [LIKELY CAUSE]
CORRELATED INCIDENTS ON NEIGHBORS (1):
- INC-001: DB timeout on svc-core-db (open) [LIKELY CAUSE]Every finding above is a real KG record — a release ID, a change ID, a session slug, a blackout window, an incident on a topology neighbor. Cross-domain causality (a database slowdown caused by the release that touched an upstream app tier twenty minutes ago) becomes obvious, not a second-pass discovery.
What a defensible AI diagnosis looks like
Verbatim Atlas output from a real Cascadia P1 — confidence 0.78, 11 tool calls, two cited prior records. Note what it doesn't do: pick the first plausible match.
Hypothesized chain: Evening peak member traffic surge → login
endpoint resource exhaustion (connection pool saturation, worker
thread limits, or upstream auth/session dependency failure) →
HTTP 503 responses. Confirmed in 1 prior incident at this
organization (inc-p1-online-banking-evening-000, 2026-01-31,
resolved in ~60 min). This is the first confirmed recurrence.
Patterns considered but rejected:
• resolution-incident-cortex-pg-1777163105 — matched
on symptoms (service degradation) but applies to cortex-pg
with stale-table-statistics root cause — different CI class.
• resolution-incident-cortex-app-1777141464 — matched
on worker exhaustion pattern, but applies to cortex-app's
Gunicorn/SSE stack — no evidence svc-online-banking uses it.
• pattern-cascadia-threshold-blip — rejected
because the alert hasn't self-resolved within 5 min and
member complaints are rising — sustained outage, not blip.It considered three other resolutions and explicitly explained why each was rejected. That — not the recommendation itself — is what makes the diagnosis defensible to the CIO who'll be asked, "did your AI guess?" A second fresh-context verifier pass then tries to refute the diagnosis against its own cited evidence before any auto-remediation is allowed; a challenged diagnosis routes to human approval automatically.
Safety engineering
Generated connectors run in a cage
The AI connector generator (paste an API doc URL, get a validated collector)
ships with a three-layer sandbox: AST static analysis that blocks
subprocess, eval, exec, and
pickle; an AI code review; and process-level isolation that
strips secrets from untrusted collector code. Nothing generated touches
production without explicit human approval.
Agents are governed, not trusted
Every AI agent has a class, an allowed-systems list, an autonomy tier (suggest → propose → approve-gated → autonomous), rate limits, a trust score that auto-pauses on degradation, and a global kill switch. Credential access is broker-mediated: agents reference secrets by name, values are injected only at execution time under a purpose-scoped grant, and every use and every denial is audited. Material never transits an LLM context.
Kick the tires yourself
Our Cascadia demo is a real, writable Cortex tenant. On a walk-through we break something live and you watch the machinery above run.