The Protocol

Janus Protocol v3.6

Adaptive dual-process governance for AI agent systems. Builder generates. Watcher critiques. Trust mediates.

Read the Research Get in Touch

Architecture

Two processes. One protocol.

The Builder

Optimized for velocity and creativity. Operates with minimal governance overhead. Context window maximally available for the primary task. Permitted to be exploratory, intuitive, fast.

Mode: Generate. Ship. Iterate.

The Watcher

Optimized for security and logic. Does not generate solutions — it critiques them. Silent by default. Consumes zero context tokens during normal operation. Activates only when deviation is detected.

Mode: Observe. Detect. Intervene.

The separation is not a preference — it's a necessity. Zhang et al. (2024) demonstrated that generators cannot reliably evaluate their own output. Tsui (2025) measured a 64.5% blind spot rate. The verifier must be architecturally separate from the generator.

Governance Mechanism

Trust Elasticity + Multi-Signal Escalation

Governance intensity scales inversely with demonstrated competence. The primary metric is iterations to convergence, augmented by semantic and confidence signals.

N=1

PASS

First attempt. Trust is high. Agent operates freely. Governance is silent.

N≥2

WARN

Second attempt, same problem. Environment stability check. Are we building on solid ground?

N≥3

HALT

Third attempt, same pattern. Something is structural. Repeating is looping, not iterating.

Multi-Signal Escalation (v3.6)

Semantic Similarity

Sentence embeddings compare the current approach against the previous three. Similarity > 0.85 triggers HALT regardless of declared iteration count.

Catches "same approach, different words"

Confidence Inference

Infers confidence from natural language patterns. Hedging reduces score. Three consecutive declines or a single drop > 0.3 triggers escalation.

Zero self-reporting overhead

Iteration Count

The original N-Pattern. Backward-compatible fallback for environments without semantic or confidence signals.

Minimum viable governance

Recovery Architecture

When HALT triggers

A governance system is only as good as its recovery mechanism. Five contextual recovery paths, recommended based on the triggering signal.

Trigger Recovery Rationale
Semantic repetition Alternative approach Same approach is failing. Need fundamentally different strategy.
Confidence decay Request human input Agent uncertainty indicates information gap human can fill.
Iteration fallback Decompose task Problem may be too large. Break into smaller subproblems.
General Reset context Context window may be polluted. Fresh start with clean state.
General Pause and reflect Metacognitive intervention. Step back before stepping forward.