--- id: glossary title: Glossary module: substrate cluster: Substrate type: glossary version: v0.2.1 status: Draft tier: free contract_role: "" canonical_url: "https://grow.goodcombinator.ai/library/registry/glossary" download_url: "https://grow.goodcombinator.ai/library/registry/glossary.md" license: CC-BY-4.0 (proposed — owner confirmation required) source: GROW by Good Combinator retrieved_at: 2026-05-29 --- # Glossary Locked one-sentence definitions used across the GROW library. Every artifact in Systems, Execution, Trust, and Coordination uses these terms in this sense. Synonyms are not permitted in artifact text; if a concept is missing, propose an addition rather than coining a local term. --- **Failure mode** — A named, reproducible way the agent can produce a wrong, harmful, or non-compliant outcome, characterized by a detectable signal and an expected correct behavior. **Escalation** — Routing a decision out of the agent's authority to a higher-authority human or system because policy, confidence, or impact thresholds have been exceeded. **Fallback** — A pre-declared safe alternative behavior the agent executes when its primary path fails a precondition, exceeds a budget, or trips a guardrail. **Override** — A human-issued instruction that replaces the agent's chosen action with a different action, recorded as an event with rationale. **Evidence** — Any retained artifact (log slice, tool trace, document, retrieval chunk, signed attestation) referenced by a provenance record to support a claim or action. **Decision origin** — The source of a single decision in an agent's trace. **Base enum** (v0.1.0, locked by C2): `agent` (autonomous), `human-override` (human replaced the agent), `fallback` (declared safe alternative fired), `escalation` (decision routed out to higher authority). **Extended values** (v0.2.0, additive backward-compatible superset introduced for the Execution/Trust/Coordination clusters): `workflow-step` (decision taken inside a modeled workflow step — C5), `sensor-triggered` (decision initiated by a fused telemetry alert — C6), `compliance-determination` (a rules-based compliance ruling — C9), `permission-decision` (a zero-trust permission grant or deny — C10), `governance-vote` (an executed DAO decision — C11). Base-enum consumers built before v0.2.0 remain valid: any record carrying a base value is unchanged, and a base-only consumer that encounters an extended value treats it as `agent` for control purposes until it declares extension-awareness. **Confidence band** — The qualitative trust level assigned to a source or retrieval, drawn from the enum `high`, `medium`, `low`, `unknown`, used by eval rubrics to weight scoring and by reliability policies to gate action. **Regression** — A previously-passing eval case that fails after a change to the agent, its tools, prompts, models, or data; tracked per failure mode. **Drift** — A gradual change in agent behavior, input distribution, or source content that degrades performance without a single discrete cause; detected by trend, not by a single failed case. **Severity** — The graded impact class of a failure mode, drawn from the enum `critical` (safety, legal, or irreversible-impact), `high` (material harm, recoverable), `medium` (degraded experience), `low` (minor defect), `info` (observability-only). **HITL gate** — A required human-in-the-loop checkpoint that blocks the agent from proceeding until a human approves, rejects, or modifies the proposed action. **Irreversible-impact boundary** — The line in a workflow beyond which an agent action cannot be undone (money moved, message sent, record filed, physical actuation); HITL gates are mandatory before crossing it. **Evaluation gate** — A versioned checklist whose pass is a precondition to advancing an artifact, module, or release to the next status (`reviewed` or `shipped`). **Provenance record** — A timestamped, append-only entry in the provenance store capturing inputs, outputs, decision origin, evidence pointers, and rationale for a single event. **Lineage** — The directed chain of provenance records and source references connecting an output back to its originating inputs, retrievals, and decisions. **Source confidence** — The confidence band assigned to a specific data source in the provenance source register, used as the C4 contract payload into eval rubrics. --- ## Cluster terms (added v0.2.0 for Execution, Trust, Coordination) **Workflow step** — A single unit of a modeled workflow with declared inputs, outputs, owner (agent or human), and branch outcome; the granularity at which Course 4 emits C5 provenance records. **Handoff** — The transfer of a task or decision between two owners (agent→human, human→agent, or agent→agent) at a defined gate, carrying a declared payload and acceptance condition. **Process map** — The directed model of a workflow's steps, decisions, branches, and handoffs; the Course 4 analogue of the S3 lineage map. **Telemetry source** — A sensor, device, API, log, stream, or external feed that emits signals; classified in Course 5 by latency, reliability, frequency, and a `confidence_band` (reusing the C4 enum). **Normalization** — The Course 5 transformation that aligns formats, units, and timestamps and resolves missing, duplicate, or conflicting signals before fusion. **Fusion** — The combination of multiple normalized telemetry signals into one operational view, with source conflicts resolved by trust, freshness, and relevance weighting. **Alert** — A Course 5 signal-threshold crossing classified `warning` or `critical`, linked to an action or fallback; the C6 payload into S1 reliability. **Model tier** — A named cost/quality class of model or deterministic logic (`premium`, `standard`, `cheap`, `deterministic`) that Course 6 routes among; the high-risk-step protection rule forbids demoting tiers on `critical`/`high` failure-mode steps without a waiver. **Routing** — The Course 6 decision selecting a model tier, tool, cache hit, batch path, or human-review path for a given task, logged via C7. **Latency budget** — The Course 6 ceiling on acceptable end-to-end response time for a task class, traded against cost and reliability. **Unit economics** — The Course 6 comparison of fully-loaded operating cost per task/user/workflow against the value or revenue it produces, expressed as a gross margin. **Jurisdiction rule** — A Course 7 governing requirement (statute, ordinance, policy, form, deadline, or approval flow) attributed to a named agency and classified universal or local. **Expert-review boundary** — The Course 7 line at which a determination must be routed to a qualified human (attorney, engineer, compliance officer); crossing it autonomously is gate-failing. Subsumes the "not legal advice" boundary. **Zero-trust** — The Course 8 default-deny posture: users, tools, documents, connectors, and model outputs are treated as potentially unsafe and must be verified before being granted action authority. **Least privilege** — The Course 8 rule that every principal (agent, tool, role) holds the minimum permissions required for its declared task and nothing more. **Action class** — The Course 8 grouping of an operation by blast radius (`read`, `external-write`, `financial`, `comms`, `admin`, `destructive`), each mapped to a control in the tool-use control list. **Tool allow-list** — The Course 8 enumerated set of tools an agent may invoke, with per-tool controls (`auto`, `confirm`, `human-approval`, `blocked`); anything not listed is denied. **Decision rights** — The Course 9 explicit assignment, per decision class, of who proposes, who votes, who executes, who may veto, and who is accountable. **Quorum** — The Course 9 minimum participation threshold for a vote to be valid, distinct from the approval threshold required to pass. **Treasury** — The Course 9 pool of funds whose allocation is governed by decision rights, budgets, milestone gates, and retroactive-reward rules. **Proposal** — The Course 9 unit of governance action: a structured request carrying evidence, a review period, and conflict-of-interest disclosures, resolved by a vote and executed under accountability tracking. **ICP (Ideal Customer Profile)** — The Course 10 definition of the target user, economic buyer, budget owner, urgency, alternatives, and switching friction for a commercialized capability. **Productization** — The Course 10 conversion of a one-off or consulting-dependent capability into a repeatable offer with standard workflows, packages, and interfaces. **Moat / defensibility** — The Course 10 durable advantage protecting a commercialized offer: proprietary workflow, data advantage, distribution, integrations, compliance moat, or operational expertise. **Proof-of-value (PoV)** — The Course 10 measurable outcome a pilot must demonstrate to justify purchase, expressed in metrics drawn from the S2 scoring system. --- ## Change log - **0.2.1 — 2026-05-29.** Resolved the mutual-downstream contradiction with `shared-artifact-schema` directionally: glossary consumes the schema, so moved `shared-artifact-schema` from `downstream_consumers` to `upstream_deps`. The edge is now schema→glossary (symmetric, acyclic). - **0.2.0 — 2026-05-29.** Extended `decision_origin` with the five additive cluster values (`workflow-step`, `sensor-triggered`, `compliance-determination`, `permission-decision`, `governance-vote`) as a backward-compatible superset. Added 23 cluster terms for Execution (Courses 4–6), Trust (Courses 7–8), and Coordination (Courses 9–10). No base-enum semantics changed. - **0.1.0 — 2026-05-28.** Initial Systems-cluster locked definitions.