Runtime behavior governance
Phalanx: runtime governance for AI assistants.
A deterministic control layer that helps deployed AI systems resist jailbreak pressure, enforce policy boundaries, and produce auditable action decisions.
What Phalanx is
Runtime governance, not a one-shot filter.
Phalanx evaluates each turn, tracks how risk evolves across a session, and decides what the application should do next under the deployment profile. It is designed for operational control: allow, clarify, constrain, refuse, route, lock, or escalate.
Public versions
Two versions for different governance boundaries.
Jailbreak governance layer
Phalanx: Guardian
A runtime governance layer focused on jailbreak pressure, prompt manipulation, and unsafe instruction patterns before they steer an assistant outside its intended boundary.
Jailbreak + policy governance layer
Phalanx: Locked Formation
Extends Guardian with deployment-policy control: constrained answers, refusal, safer routing, session locking, and human-review handoff under an auditable action policy.
Three-layer model
Turn risk, session posture, action policy.
TurnRiskState
Turn parser
Produces ALLOW, UNCERTAIN, or REFUSE plus risk families, dimensions, and evidence references.
SessionRiskState
Session governor
Tracks cumulative conversation posture across normal, monitored, soft_guarded, or locked states.
ActionPolicyResult
Action policy
Decides what the application should do now under the selected deployment profile.
Action decisions
Routing surface
Integration surfaces
Designed for product control planes.
Audits reveal brittle behavior before deployment. Phalanx is the runtime control layer that can help govern behavior after deployment.
- Local SDK
- HTTP service
- Container deployment