Arc Codex › DLB
We deliver deliberation.

About Deliberation Stack

A methodology demonstration.

Generative AI produces text. Deliberation Stack produces deliberated text — analysis that has already been argued over before it reaches you.

The system ingests operational signal from a private infrastructure: web traffic patterns, authentication logs, file-integrity changes, and ad-hoc operator events. Each event is analyzed by four simulated analyst perspectives, each tuned to a specific role:

  • Blue Team — Operational summary. What happened, what's normal, what's not.
  • Red Team — Facts only. No interpretation, no recommendations. The raw record.
  • Purple Team — Pattern analysis. Looks for what Blue Team's summary missed.
  • Counter-Analyst — Dissent. Challenges the consensus the other three produced, including the system's own heuristics.

The deliverable is a deliberated artifact: the kind of write-up a security team might produce after meeting about an incident, except it runs continuously on signal that's usually too low-volume to staff against.

Most AI-assisted tooling collapses multiple viewpoints into a single authoritative-sounding voice. Disagreement is the part that gets engineered out.

This system treats disagreement as the actual value. A CISO doesn't need the answer; they need the argument. The Counter-Analyst exists specifically to surface what the consensus missed, regardless of whether the consensus was "everything's fine" or "this is critical."

Two-tier. An autonomic reflex layer (Caddy, fail2ban, rate limits) handles ~99% of noise — bot scans, generic probes, background internet weather. The deliberation layer chews on the 1% that survives. This mirrors how human security teams actually work: the WAF blocks the obvious, the analyst reviews what the WAF didn't.

Severity is path-routed: trace events go through fast local models; critical events escalate to higher-quality models. The same article never costs the same to analyze twice.

Not a product. Not a SaaS offering. Not a SIEM replacement. A working argument that the meeting-room layer above a SIEM is where AI adds the most value, and that disagreement is the feature, not the bug.

← Back to feed

DLB is also the lab's summary of alerts and audits. Every monitoring source, from the security event log to the nightly backup check, reports here as OK or PROBLEM, and a problem stays visible until it is fixed. The rule behind it: a control counts only when an automatic check proves it, every hour.

The full security guide is for signed-in users: How the lab is secured.

Since November 2026 the lab also runs a small Kubernetes cluster (k3s) managed with GitOps: Argo CD keeps it matching Git and reverts manual changes, and Argo Workflows tests every relevant pull request, then builds and deploys a test copy of Arc's MCP server. The security choices are the ones you'd want at work: the CI's GitHub token can only read code and post commit statuses, secrets live only in Kubernetes Secrets, every image in our manifests is pinned by digest, pods run as unprivileged users with read-only root filesystems and no Linux capabilities, and the cluster sits in its own container with nothing new exposed on the network.

The full write-up: How the lab ships code.