Glossary

Torno keeps a small, fixed vocabulary and uses it identically across the app, the API, the CLI, and MCP. These are the nouns you will meet in exports, webhooks, and audit records, defined once here.

Structure

Term	Meaning
Workspace	The organization-level tenant: members, Principals, policy, a home region, and one or more Properties.
Property	A thing being grown: an app, a marketing site, or a pre-launch idea. Data, Environments, the Brand Kit, Loops, and Metrics are Property-scoped.
Environment	One of the fixed product-data contexts `dev`, `staging`, and `prod`, each with its own write keys.
Brand Kit	One per Property: versioned voice, positioning, visual identity, proof points, banned claims, and compliance notes that ground generated content.

The Loop and its objects

Term	Meaning
Loop	The durable container and state machine for a growth objective, from listening through Decision and back, with one target Metric and append-only memory.
Turn	One pass through the Loop lifecycle, keeping its transition, evidence, proposal, and Decision history.
Signal	An immutable, timestamped observation — `voice`, `behavior`, `outcome`, or `metric_delta` — with source, consent, provenance, and trust attached. Corrections are new Signals.
Analyzer	A versioned function over Signals or aggregates, with declared schemas and determinism. LLM use records model and prompt versions.
Insight	A claim with confidence, supporting Signal references, freshness and decay, and inherited trust labels.
Hypothesis	The structured bridge from evidence to action: proposed change, audience, Metric, expected effect, and Insight references, scored for impact, confidence, and effort.
Change	A concrete intervention — page variant, ad creative, ad campaign, product change, or content — with a rollout plan and an explicit exposure decision.
Exposure	Evidence that a Subject or Cohort actually saw a Change. Its definition, tag slug, goal events, and assignment form the measurability contract.
Experiment	An optional pre-registered wrapper assigning an audience across Change variants, with layers, a stats plan, target and guardrail Metrics, and explicit evidence states.
Result / Readout	Computed evidence from an Analyzer or Experiment, including uncertainty and data-quality state. Not itself a Decision.
Evidence state	One of `insufficient_evidence`, `trending`, `significant_win`, `significant_loss`, or `guardrail_breach` — always shown with bounds and data basis, never as a naked p-value.
Decision	`ship`, `kill`, `iterate`, or `extend`, with rationale and the deciding Principal, appended to the Loop’s memory.

Measurement

Term	Meaning
Metric	A named, typed, versioned definition read through one semantic contract carrying value, unit, bounds, status, freshness, source basis, and definition version.
Needle	The Metric-centered view: a target Metric with trend, decomposition, leading-indicator pairing, and the live Changes aimed at it with their evidence states. A view, not a second metric engine.
Leading indicator pairing	A declared weekly-steering Metric paired with a lagging revenue target; required for revenue-class Loops.
Guardrail Metric	A harm-sensitive Metric monitored alongside a target. A breach is an explicit evidence state and can pause a rollout automatically.
Layer	A mutually exclusive Experiment traffic group used to prevent unacknowledged overlap.
Program holdout	An optional small cohort excluded from all Changes, used to measure the growth program as a whole.

People and identity

Term	Meaning
Subject	A pseudonymous key used only to join anonymous first touch, assignment, signup, and revenue. It carries no PII and is not a person profile.
Cohort	The reporting noun for a segment, behavioral group, acquisition group, or Experiment arm. Cohort reads are k-gated and subtraction-safe.
Customer	An identified post-signup person, created from `identify()` traits in profile mode. Individual reads require the `customer_read` scope; cohort-only mode has no Customers.
Account	An optional B2B rollup of Customers, with aggregate lifecycle and revenue context.

Governance and agents

Term	Meaning
Principal	An actor: human, agent, or API token. Principals carry scopes, and machine actions must include reasoning.
Policy	A declarative rule evaluated at the mutation choke point. Outcomes are allow, deny, or require approval, plus rate and spend enforcement.
Approval Request	A first-class snapshot of a governed action awaiting human judgment. Approval mints an execution grant rather than asking the caller to retry.
Loops-as-code	The `torno.config.ts` or YAML representation of Metrics, Loops, Changes, tags, and the event dictionary, with pull, diff, push, and code checkers.

Data and integrity

Term	Meaning
Write key	An Environment-specific collection credential. Client and server keys have different capabilities; rotation, grace, and revocation are explicit.
Event dictionary	The versioned tracking plan — optionally owned in code — defining event, goal, and trait names and their property schemas.
Tag slug	The stable Change exposure identifier used in code and data attributes, validated by `torno tag check`. Distinct from display and machine IDs.
Machine ID	A globally stable, prefixed public identifier. Internal database IDs are never exposed as product identity.
Display ID	A human-facing per-Property sequence such as `LP-1`, `INS-7`, or `CHG-14`. Ambiguous display IDs require Property context.
Provenance	Source and trust evidence attached to data: `owned`, `verified_partner`, `public_api`, or `unverified`.
Quarantine	The storage and rendering boundary for untrusted content and suspicious events. Quarantined text leaves only through escaped, fenced-data helpers.
`untrusted-derived`	The label inherited by any output based on untrusted inputs, even after summarization or analysis.

For how these objects behave in sequence, read The Loop and Loops in practice; for how agents address them, see The agent interface.