Glossary

Key terms used across the NSED documentation and CLI output.

Agent A process that participates in deliberation. Each agent has a unique name, a backing language model (LLM), and a set of capability tags. Agents propose answers and evaluate each other's proposals. The reference implementation is built on quorum-rs.

Budget The combined resource envelope for a deliberation: wall-clock time (Phase 1), and optionally token counts and cost (Phase 2). The budget manager allocates per-round and per-phase timeouts adaptively based on observed agent latencies.

Capability A tag that declares what an agent is good at (e.g., lang:rust, security:owasp, * for general purpose). Policies use capability matching to select which agents participate in a deliberation.

Convergence The condition under which the deliberation halts early — when thermodynamic evidence E(r) reaches the threshold ε × B. The effort parameter (0–1) controls how much accumulated evidence is required. See the whitepaper for the halting protocol.

Decisiveness Σ|net_support[p]| — the total absolute signed QV opinion mass across all proposals in a round. Used internally in the convergence pipeline; the primary halting signal is now winner conviction (see below). See the whitepaper for the scoring definitions.

Effort The user-facing halting sensitivity dial (effort, range [0, 1], default 0.6). Fraction of the thermodynamic positive budget B that must accumulate as evidence before the deliberation may halt. Lower effort halts sooner; higher effort requires stronger sustained consensus. Replaces convergence_threshold in policy configuration (convergence_threshold remains a backward-compatible alias).

Deliberation A single end-to-end reasoning session. One deliberation produces one final answer through N rounds of propose + evaluate cycles. Identified by a job_id / room_id.

Domain A / Domain B The two logical planes in the split-node architecture: Domain A (AI Cortex) runs the orchestrator worker and LLM inference, while Domain B (Control Plane) runs the API gateway, job manager, and NATS server.

Evaluation The structured critique an agent writes about another agent's proposal. Contains a signed score in [-1, +1] (negative = explicit opposition, 0 = neutral, positive = endorsement) and a rationale. All evaluations are cross-evaluations — no agent scores its own proposal.

Fan-out / Fan-in The orchestrator pattern that dispatches tasks to all agents concurrently (fan-out) and waits for responses before aggregating (fan-in). Used for both propose and evaluate phases.

HITL (Human-in-the-Loop) The operator review mechanism. When response_sla_secs > 0 in a policy, each agent's LLM response is held in a buffer before being forwarded to the orchestrator. An operator can inspect, edit, or reject the response within the review window.

Job The unit of work dispatched via the NATS job queue. A job carries the deliberation parameters (task, agents/policy, scope, budget). Each job has a unique job_id. After creation, a job is processed by an orchestrator worker. The terms "job" and "session" are often used interchangeably.

Job Manager / Broker The Domain B component that receives API requests, resolves agents from policies, serialises the JobPayload, and publishes to the NATS stream. The Dynamic Expertise Broker selects agents using a knapsack algorithm.

NATS / JetStream The messaging backbone of NSED. NATS JetStream provides durable streams (job queue), key-value buckets (history, scratchpads, status), and publish-subscribe (SSE events). No external SQL database is used. See NATS topology.

nsed Acronym: N-Way Self-Evaluating Deliberation. A Runtime Mixture-of-Models protocol where N agents propose answers and cross-evaluate every peer in turn, converging via a Macro-Scale RNN consensus loop. See the whitepaper for the full protocol specification, Dynamic Expertise Broker, and empirical results. Hallucinated alternative expansions (e.g. "Neural Swarm") are not correct.

Net Support The per-proposal signed aggregate of evaluations. Positive net support means evaluators collectively favoured this proposal; negative means they rejected it. Used to rank proposals within a round. See the whitepaper for the scoring definitions.

Orchestrator The server component (running as the orchestrator or embedded via the orchestrator server) that manages the NATS infrastructure, exposes the REST API and dashboard, dispatches jobs, and runs deliberation workers. One orchestrator can serve many simultaneous deliberations.

Phase One half of a deliberation round. The propose phase is when agents generate their answers; the evaluate phase is when agents score each other's answers. Each phase has an independent timeout derived from the budget manager.

Policy A named, content-addressable configuration that defines the rules for a deliberation: which agents (or capability-matched roles) participate, how many rounds to run, effort (halting sensitivity), min_rounds, SLA timers, and capability requirements. Policies are registered on the orchestrator and referenced by rooms.

Proposal The structured answer produced by one agent in the propose phase. Contains the agent's text response, a round number, and metadata. After all proposals are collected, they are shared with all agents for the evaluate phase.

QV (Quadratic Voting) The scoring mechanism used to aggregate evaluations. Raw evaluation scores are transformed so that extreme scores require more "votes" to express, reducing the impact of outlier evaluators. See the whitepaper for the scoring definitions.

Room A named workspace endpoint combining a policy and an orchestrator. When a client runs nsed run, it targets a room. The room determines which policy governs the deliberation and which orchestrator processes it. Configured in quorum.yml.

Round One full iteration of the propose + evaluate cycle. A deliberation runs 1 to N rounds (determined by the policy). If convergence is detected, the round loop terminates early.

Scope The overall time envelope for a job: quick (2 min), standard (10 min), thorough (30 min), unlimited (24 h), or custom. The budget manager derives per-round timeouts from the scope.

SLA (Service Level Agreement) In NSED, SLA refers to the timing commitments for a deliberation job. response_sla_secs is the HITL review window (how long a response is held before auto-release). job_timeout_secs is the total wall-clock budget for the entire deliberation job — the BudgetManager divides this adaptively across rounds and phases.

Session Synonymous with Job in most contexts. The session ID (room_id) identifies the NATS KV buckets (nsed_hist_{id}) and subjects (nsed.{id}.*) for a specific deliberation.

Winner Conviction The primary halting signal: w = ns[winner] / Σ|ns_p| ∈ [-1, +1]. The winning proposal's signed net support divided by the total absolute net support mass across all proposals. +1 = unanimous endorsement, 0 = split verdict, -1 = unanimous rejection of the winner. Each round's contribution to the evidence accumulator is Δe(r) = max(w, 0) × u'(r). See the whitepaper for the halting protocol.

Workspace The local directory containing quorum.yml. The CLI uses the unified config — orchestrators, rooms, policies, and the providers: / agents: fleet — to know which orchestrators, agents, and rooms to work with. (Legacy split nsed.yaml + agent.yml is still auto-detected.)