The Three-Layer Architecture

SPECTRA is reasoned about — and built — through three layers. LLM reasoning is used for synthesis and routing; deterministic code is used for repeatability, release safety and regression control.

Layer 1 — Directive

Markdown SOPs, agent prompts, workflows and docs. They define goals, inputs, outputs, safety boundaries and edge cases. This is where intent lives.

Layer 2 — Orchestration

The AI agent routes tasks, reads context, chooses tools, handles errors and updates docs. The AI does not replace deterministic validation or test scripts.

Layer 3 — Execution

Deterministic scripts, validators, CLIs, tests, schemas and adapters. Repeatable logic belongs here — engagement gating, scope enforcement, evidence chain, attack-path generation, the tool adapter, the benchmark harness.

Why it works

Separating reasoning from repeatable execution is what makes SPECTRA testable and releasable: 1,800+ validator checks, deterministic state machines, and an evidence chain that anchors every claim. The agents are creative; the guardrails are not negotiable.

The adversarial core

SPECTRA’s flagship is the War Room / Duel Mode: Red and Blue clash on the same target and an impartial Referee scores them on evidence — detection latency, severity coverage, technique misses — never on prior knowledge of the plan. The expected result is not “invisible Red”; it is an honest measurement of which signals were produced, seen, and missed.