Glossary
The vocabulary the rest of the docs assume. Terms are grouped, but each entry can be read on its own.
Identity & credentials
Section titled “Identity & credentials”Mechanism, not category. DID, VC, and VP are credential formats — implementation mechanisms beneath the PaloNexus pillars, not the product category. Externally, PaloNexus speaks of agent identity, signed agent credentials, delegated authority, proof of authorization, and short-lived runtime credentials; DID/VC is one supported way to implement them.
DID — Decentralized Identifier. A self-describing identifier of the form did:<method>:…
that resolves to a public key. PaloNexus uses two methods:
did:key— a self-certifying DID where the identifier is the public key. Agents aredid:key, minted at startup, resolvable 100% offline (no registry, no network).did:web— a DID anchored in DNS + web PKI, resolved over HTTPS from/.well-known/did.json. The issuer / trust root isdid:web(rotatable by publishing a new key). No blockchain is involved.
VC — Verifiable Credential. A signed, tamper-evident claim about a subject. PaloNexus issues VCs as JWT-VC (a VC encoded as an EdDSA-signed JWT — JSON Web Token). Three kinds appear in these docs:
- Membership VC — issued by the trust root to an agent, proving it is an enrolled member
of the org. Held by the agent after
agent.provision(). - Delegation VC — issued when a human approves a task-scoped delegation; it encodes
who delegated what action on what resource, time-boxed. Present on a
Delegationonce itsstatusisapproved. - Capability VC — a credential granting a specific capability to an agent.
- Compliance Credential — a named-standard attestation (GDPR, HIPAA, SOC2-TypeII,
EU-AI-Act-Art50) about an agent, issued by a
compliance_auditor-role human as a real, signeddid:webJWT-VC. Self-attested — PaloNexus verifies the signature and revocation status, not the underlying audit claim. See Governance credentials. - Provenance Credential — a self-declared attestation of what produced an agent’s
outputs (base model, training-data lineage, declared owner), issued by a
provenance_attestor-role human. A routine model update supersedes the prior credential (no cascade); an explicit revoke does cascade. See Governance credentials.
VP — Verifiable Presentation. A holder-signed wrapper an agent builds fresh on each
egress call: the agent signs (with its did:key) over an audience + nonce, wrapping its
Membership VC. It proves who is calling, right now without forwarding a raw bearer token.
Carried in the X-Palonexus-Agent-VP header (or Proxy-Authorization at the egress proxy).
JWT-VC. A Verifiable Credential serialized as a JSON Web Token, signed with EdDSA (Ed25519). The wire format for every VC above.
StatusList / StatusList2021. A compact, verifiable revocation list. A credential carries
a position in the list; the /authz decision point checks it so a revoked credential is
rejected immediately — this is what makes live revocation mid-run possible.
Subject (stable subject). The durable enterprise identifier for a human, from the
directory sync — not their email. The SDK’s HumanOwner.subject is this value; emails
can change, the subject does not.
did:key / did:web — see DID above.
Authorization model
Section titled “Authorization model”/authz. The single authorization decision point. Identity + registry + policy converge
to answer one question — may this caller reach this target? — for both ingress and agent
egress. Returns 200 (allow), 401 + needs-approval (delegation required), or 403 (deny).
ext_authz. Envoy’s external-authorization filter — the mechanism that routes every
request to one decision point, /authz, before it reaches an upstream. The decision it
asks is the agent-egress question — may this agent make this call, on behalf of this
human, for this task, right now? — and the same /authz answers it for north–south
ingress too (the foundation egress governance builds on). Either way there’s no
per-service auth code, just one place to reason about access.
Deny-by-default. The platform’s default answer is no. Unknown agent, invalid or missing credential, expired delegation, target not allowlisted, or an unreachable decision point all deny. Access is granted only by an explicit, current allow.
Fail-closed. When the decision point itself is unreachable, the system denies rather
than assuming allow. In the SDK this surfaces as a raised ControlPlaneUnavailable — never a
silent success.
TBAC — Task-Based Access Control. Authorization scoped to a task (an incident, thread,
or run id, e.g. INC-4821) rather than a standing role. A delegation is granted for one task
and times out — the agent does not keep the privilege afterward. TBAC is the mechanism
beneath the just-in-time access pillar — how task-scoped authority is implemented, not
a product category.
One authorization contract. The design principle that every action — an inbound request
or an agent’s outbound call — converges on the same single /authz question and answer
shape. Previously described in these docs as the “one-decision model”.
Just-in-time access. One of the five pillars: agents hold no standing enterprise credentials; access is issued for one task, target, action set, and short time window, then expires or is revoked automatically.
On-behalf-of. An agent never acts as itself for regulated work; it acts on behalf of a
human subject. The control plane records both the actor (agent) and the subject
(human). Carried in the X-Palonexus-On-Behalf-Of header.
Delegation. A task-scoped, human-approved, time-boxed grant letting an agent perform a
specific action on a specific resource. Modeled by Delegation; states are
pending → approved | denied | expired | revoked.
Membership VC / Delegation VC / Capability VC — see VC above.
OPA — Open Policy Agent. The policy engine holding the org-wide Rego bundle. Policy is deny-overrides: an inline allow plus an OPA deny equals deny (OPA can veto, never rubber-stamp).
Rego. OPA’s policy language; the org policy bundle (authz.rego) is written in it.
Registry. The catalog of services/targets an agent may reach, each with an owner, an
upstream, a dataClass, and a verbatim requireScope. /authz resolves the target here.
requireScope. The exact scope string a registry target demands. Must match the SDK’s
CONTROL_PLANE_SERVICES mapping verbatim (guarded by a parity test) so the model and the
enforced system never drift.
dataClass. A target’s sensitivity: public | internal | regulated. Regulated
targets drive the needs-approval path.
Governance & authority
Section titled “Governance & authority”Authority graph. The complete chain a PaloNexus decision resolves before it allows an action: human identity → organizational role → ownership of resource/service → authority to delegate → agent identity → task mandate → permitted operation → target resource → time/risk/budget constraints → issued runtime credential → recorded action. It is what makes the decision about authority, not merely interception. See Core concepts.
Authority trail. The record set proving the authority graph for every action: which agent acted, under whose authority, who approved the delegation, whether the approver was entitled to approve it, what task justified it, what credential was issued, and whether the authority was later revoked. Implemented as the hash-chained, tamper-evident audit chain (see Audit hash chain).
Authority-bound agent. An agent that can act only with authority a real person or
service owner was entitled to delegate — and only for the approved task, resource, and
time. The positioning term for what the SDK and code surfaces call a governed agent
(code symbols such as governed_node keep their code-accurate names).
Owner / Sponsor / Approver / Operator / Auditor. The five human roles around a governed agent. Every agent must have an owner (accountable) and a sponsor (business backing) — the no-orphaned-agents rule. An approver authorizes delegations, an operator runs the agent, an auditor reviews its trail.
compliance_auditor / provenance_attestor. Two narrower, credential-issuing roles,
distinct from the general auditor above: compliance_auditor may issue/revoke Compliance
Credentials, provenance_attestor may issue/revoke Provenance Credentials. Deliberately
separate roles — the person who knows an agent’s model lineage is often not its compliance
auditor or its business owner. See
Governance credentials.
org:agents:*. The org-scope authority strings carried on org roles:
org:agents:own, :sponsor, :approve, :operate, :audit. HumanOwner.agent_authority
is derived from these.
Governance (no-orphaned-agents). The rule that registration requires a mandatory owner
and sponsor. Violations raise GovernanceError client-side before any network call, and are
re-validated server-side.
Risk tier. An agent’s governance risk classification: low | medium | high | critical.
Runtime. The approved execution environment recorded for an agent, e.g. doks_prod.
Scenario. A seeded end-to-end story (e.g. devops-incident) tying an agent to its owner,
sponsor, approver, operator, auditor, permissions, and a negative persona.
Negative persona. The seeded user who must be hard-denied for a scenario — the red test case.
Runtime & frameworks
Section titled “Runtime & frameworks”SDK. The palonexus Python package — one typed front door over the control plane and
agent-idp. Lean core; framework bindings are opt-in extras.
Control plane. The Go decision engine hosting /authz, the registry, and the audit
chain. The SDK talks to it (and to agent-idp) over HTTP.
agent-idp. The identity provider service: agent governance, provisioning, delegations,
revocation, and directory sync. The SDK’s pn.agents / pn.revoke / delegation calls hit
it.
Egress. An agent’s outbound action — a model call, tool call, or agent-to-agent hop.
The hard problem PaloNexus governs: every egress passes the same /authz decision carrying
agent + on-behalf-of identity.
Ingress. A north–south request into the cluster (client → gateway → upstream), the
classic API-gateway direction. It is the foundational decision the platform started
from — the same /authz spine that now also governs agent egress.
A2A — Agent-to-Agent. One agent calling another. The hop is itself gated and carries the original on-behalf-of subject.
HITL — Human-in-the-Loop. A run that pauses for a human decision (here, approving a
delegation) before continuing. In LangGraph this is interrupt() + Command(resume=...).
interrupt / Command(resume=…). LangGraph’s pause/resume primitives. The governance
adapter calls interrupt() on a needs-approval decision; the graph resumes after a human
approves.
Checkpointer. LangGraph’s durable state store (MemorySaver in dev, AsyncPostgresSaver
in production). Required for HITL — without it an interrupted run cannot resume.
thread_id. The run identifier in a LangGraph config ({"configurable": {"thread_id": …}}) under which the checkpointer persists state.
ToolMessage. A LangChain message returned in place of a tool result. The LangChain gate
substitutes a deny ToolMessage on a hard deny, so the model sees the denial instead of
the tool running.
Middleware / governed_node. The framework adapters. middleware(pn) gates every
declared tool call in create_agent; governed_node(pn, …) gates a LangGraph node.
Operations & observability
Section titled “Operations & observability”OTel — OpenTelemetry. The tracing/metrics standard. Governed calls carry a trace id
(PolicyDecision.trace_id) so a run is reconstructable in Tempo.
Audit hash chain. The tamper-evident decision log: each record’s prev_hash equals the
previous record’s hash. Editing or deleting any record breaks the chain;
pn.audit.verify_chain() checks it.
M2M — Machine-to-Machine. Non-interactive service credentials (e.g. the Logto seeder’s client id/secret) used by backend services rather than a logged-in human.
IdP — Identity Provider. A service that issues and verifies identities. Here, agent-idp (for agents) and the workforce IdP (for humans; Logto).
IdP / Workforce IdP. The enterprise identity provider that owns human identity — the system of record for employees, groups, org roles, and lifecycle status. PaloNexus is IdP-neutral: it integrates any OIDC/SCIM workforce IdP (Okta, Microsoft Entra ID, Auth0, Ping, Google Workspace, Amazon Cognito, Keycloak, Logto, …) via standards (OIDC, SAML, SCIM / directory sync, API sync, webhooks) and never owns human identity itself — it owns AI agents, delegation, task authorization, temporary elevation, and audit. See IdP Support Model.
OIDC — OpenID Connect. The OAuth 2.0-based authentication standard PaloNexus uses to integrate a workforce IdP for human sign-in and to verify tokens against the IdP’s JWKS (JSON Web Key Set — the IdP’s published signing keys). One of the standards that makes PaloNexus IdP-neutral. See IdP Support Model.
SAML — Security Assertion Markup Language. The XML-based SSO/federation standard some enterprise IdPs use for workforce sign-in; supported alongside OIDC for connecting a workforce IdP. See IdP Support Model.
Directory sync. Keeping the workforce directory in step with the enterprise IdP — employees, groups, org roles, and lifecycle status — via SCIM, API sync, or webhooks. A snapshot reconcile keyed by the stable subject, idempotent and tenant-isolated. See IdP Support Model.
Lifecycle status (joiner / mover / leaver). The employee’s state in the workforce IdP, propagated by directory sync: a joiner is provisioned, a mover changes groups/roles/org, and a leaver is deprovisioned (cascading revocation of any agent authority they held). See IdP Support Model.
Logto. The supported workforce IdP — the OIDC/SCIM identity provider PaloNexus
integrates with for human identity. The core is IdP-neutral as an architecture fact, so other
OIDC/SCIM IdPs (Okta, Entra ID, Auth0, …) integrate on the same surfaces. The seed-logto
tool seeds the sample organization into it for evaluation and testing. See
IdP Support Model.
SCIM — System for Cross-domain Identity Management. The standard (SCIM 2.0 User/Group) PaloNexus uses to sync the workforce directory from the enterprise IdP (Okta / Entra / Workspace). Sync is a snapshot reconcile keyed by the stable subject — joiner / mover / leaver / rehire — and is idempotent and tenant-isolated. See Enterprise IAM API.
STS — Security Token Service. The agent-idp service that exchanges an agent’s proof plus
a delegation’s evidence into a short-lived, audience-bound JWT (sub=agent, act=human,
cnf proof-of-possession, tight TTL), signed with the issuer Ed25519 key. Refused from a
revoked or expired delegation; logged metadata-only. See
Enterprise IAM API.
Northstar. The seeded sample organization (org_id 7gdgqfu5j0oo) that the
fixtures — agent names, example emails, and scenario ids in code samples — belong to.
Envoy. The data-plane proxy at the gateway that runs ext_authz against /authz.
NetworkPolicy. The Kubernetes object that confines an agent pod to the egress proxy, so egress enforcement holds for any framework, not just cooperating SDK code.
DOKS — DigitalOcean Kubernetes Service. A managed Kubernetes target for self-hosting.
SDK types referenced in these docs
Section titled “SDK types referenced in these docs”PolicyDecision. The typed result of task.check(): allow, needs_approval, reason,
subject, upstream, trace_id.
AgentIdentity / HumanOwner / Delegation / TaskSession / Credential / AuditEvent / Resource / AssetType. The core typed models — see the Quickstart and SDK overview for how they map to platform surfaces.
X-Palonexus-* headers. The request/response headers carrying egress identity and the
decision — -Actor, -On-Behalf-Of, -Task, -Action, -Resource, -Agent-VP,
-Subject, -Needs-Approval, -Deny-Reason. See the Headers reference.