Skip to main content

Glossary

Concise definitions of the terms used across the Splyntra documentation.

Run​

The first-class unit in Splyntra: one end-to-end execution of an agent. A run fans out into planning steps, tool calls, retrievals, and sub-agent handoffs, and carries both performance and security signal. See Traces & risk.

Trace​

The full, hierarchical timeline of a run — a tree of spans capturing prompts, model calls, tool invocations, retrievals, and handoffs, with latency, token counts, and cost on each. See Traces.

Span​

A single operation within a trace (an LLM call, a tool call, a retrieval). Spans carry OpenTelemetry attributes and an OK/ERROR status; a thrown error is recorded on the span with the exception attached.

Span kind​

The category of a span — for example llm_call, agent, step, tool_call, retrieval, or vector_search. Instrumentors map framework operations onto these kinds (e.g. LangGraph nodes → step, MCP tools/call → tool_call).

Risk score​

A single 0–100 signal per run combining secret/PII detection, prompt-injection detection, content moderation, and policy violations, mapped down onto the exact spans where the risk appeared. See The risk score.

Detector​

A component of the security pipeline that inspects span content for a class of risk — secrets, PII, unsafe tool calls, moderation, or prompt injection — and contributes to the risk score. See Detection & redaction.

Guard​

An inline enforcement control that runs a high-precision prompt-injection pre-flight before a call reaches the model. Modes: off, monitor (flag only), and block (raise SplyntraBlocked). Fail-open by default. See Guardrails.

Redaction​

Client-side stripping of high-confidence secrets from spans before they leave your process. On by default (redact_by_default / redactByDefault). Risk is still scored even when raw values never leave the machine.

Scorer​

A function that grades an evaluation result against its expected output. Built-in scorers ship in the core; llm_as_judge is a commercial scorer registered by ee/scorers-pro. See Scorers.

Dataset​

A named set of evaluation items, each with an input, an expected_output, and optional context. Push once, reference by id. See Evaluation.

Baseline​

A stored reference run that later evaluation runs are compared against to detect regression. Set it once from a known-good run. See CI regression gates.

Gate​

An evaluation option (gate=True) that fails a run — and, in CI, the build — when scores fall below the baseline. Drives the splyntra eval run --gate CLI. See Gate CI on evals.

Ledger (activity / audit ledger)​

An append-only, per-org record with a monotonic sequence and a prev_hash → hash chain, so tampering is detectable ("chain verified / broken @ seq N"). Every governance decision and identity event is written here. Team+ feature. See Activity ledger.

Policy​

A rule in the policy engine (RBAC/ABAC/ReBAC) with a subject, resource, action, effect, priority, and conditions. Deny wins; the engine fails closed on load error. Enterprise feature. See Policy engine.

Delegation​

Controls over what an agent may do: spend limits (daily/monthly USD caps, Team+) and agent permissions / approvals / approval rules (Enterprise). See Delegation & approvals.

Authorize​

The decision an agent requests before a consequential action, via authorize() or POST /v1/authorize. Returns allow, deny, or needs_approval after evaluating explicit denies, spend limits, policy (deny-wins), and approval rules — and writes the outcome to the ledger. See Governance overview.

Agent identity​

A registered agent (id/name/owner/environment) in the identity registry, the basis for scoped credentials and trust. Enterprise feature. See Agent identity.

Credential​

A scoped secret issued to a registered agent — shown once, stored as a SHA-256 hash with a prefix, with scopes[], rotation, revocation, and expiry.

Trust policy​

A cross-agent allow/deny rule matched by agent-id or group:*. Deny wins; the default is deny. See Trust policies.

Federation​

Trusting an external OIDC issuer (Entra/Okta/Auth0) so agents authenticate with a federated token or exchange it (RFC 8693-style JIT) for a short-lived scoped Splyntra credential. See Federation.

OTLP​

The OpenTelemetry Protocol. Splyntra's ingest is OTLP/HTTP, with paths /v1/traces and /v1/logs. See Ingest & OTLP.

GenAI semantic conventions (gen_ai.*)​

The OpenTelemetry attribute conventions for generative-AI spans — model name, token counts, and request/response metadata. Splyntra reads them (alongside traceloop.entity.* and splyntra.*) so third-party OTel spans ingest natively.

Ingest key​

The API key that authenticates telemetry, sent as Authorization: Bearer <key>. Stored only as a SHA-256 hash; shown once at creation. splyntra_dev_key is a dev-only fallback rejected in production. See API overview.

Project​

A logical grouping of agent telemetry within an organization. Plan limits cap the number of projects. See Dashboard overview.

Environment​

A label separating telemetry by deployment stage (e.g. development, staging, production). Set on the SDK; used for filtering. Plan limits cap the number of environments.

Org (organization)​

The top-level tenant that owns projects, members, and a plan. Cloud is multi-tenant with path-based routing (/{orgSlug}/…); Community runs a single implicit org. See Organizations & teams.

Plan / feature​

A Cloud subscription tier (free, pro, team, enterprise) that gates features and limits. Enforcement is server-side, not just navigation hiding. See Plans & pricing.