Developers · Glossary

AI agent glossary

The words we use for AI agents, governed decisions and audit records, in plain English. 42 terms, each with its own link.

Engineering terms from earlier versions of elan1.ai, such as KernelPort, write intent and Mode A/B, are marked as engineering terms, with their full specification to follow in the developer documentation at launch. Old glossary links land on the matching term here.

Last reviewed 2 November 2026 · v1

A

Agent

Software that uses an AI model to decide on steps and call tools to carry them out. On elan1, an agent can propose an action; it cannot commit one on its own.

Allow

One of the 3 outcomes elan1 Govern (agent governance and assurance) gives an action. A rule you wrote clearly permits it, so it runs without anyone waiting, and the outcome is recorded.

Answerable

A named person stands behind an action and can show the rule, the decision and the record.

Ask a person

Also: ask_person

The outcome for an action your rules say needs a human. It waits for the named approver, who answers in the console or with one tap in Slack or WhatsApp. It resolves to approved, refused or expired: a request nobody answers within 24 hours is refused, nothing changes, and the expiry is recorded.

Assurance level

How far a claim can be checked without trusting elan1: Stated (we say it), Verifiable (you can test it from the record) or Attested (an independent party has examined it). Every row in the claims register carries one.

Audit trail

A log of every edit to a company's books. In India, the Companies (Accounts) Rules 2014 have required one since 1 April 2023, kept for 8 years. elan1 Work's record is designed to support it.

C

Calibration

Also: Confidence

Checking that a stated confidence matches reality: answers given at 90% confidence should be right about 90 times in 100. Every elan1 Decide (decision intelligence) answer carries a confidence, and confidence and risk together choose its route. Decide joins real outcomes back to keep the confidence honest. No calibration, no automation.

Capability contract

A stable interface in elan1 Run (sovereign agent infrastructure), such as chat, search, models or memory, that your apps call instead of a specific tool. There are 10, each with at least 2 open-source implementations.

Claims register

The public list of every claim elan1 makes, each with its type, status and assurance level, signed off by the founder and legal before publication. A product fact describes how a product works as designed, not how well. A design target is a figure the design aims for, such as an answer in under 800 ms, 95 times in 100. A modelled estimate is calculated from stated assumptions. Neither a target nor an estimate is a measurement, and both are labelled wherever they appear.

Component swap

In elan1 Run, replacing a component that has failed, been archived or changed licence. The replacement passes the contract's conformance tests on your setup, shadows real traffic and is signed off by 2 named people. Your apps do not change.

Engineering term

Conformance

Proof that an implementation meets a contract. A component is offered for an elan1 Run capability contract only once it passes that contract's conformance tests, and a replacement must pass them again on your setup before a component swap. The tests will be published with each contract in the developer documentation at launch.

Back to A to Z, from C

D

Data principal

Under India's Digital Personal Data Protection Act 2023, the person the personal data is about. The organisation that decides why and how it is processed is the data fiduciary, comparable to a controller under the GDPR. A data principal can ask for access, correction and erasure, and can raise a grievance with the fiduciary's grievance officer.

Data residency

Where data is stored and processed. elan1 processes customer data in AWS Mumbai, with no US region. New regions open on a signed contract, never a forecast.

Security and data residency (related to Data residency)

Back to A to Z, from D

E

Emergency stop

Also: Kill switch

One action in elan1 Govern that pauses every agent at once and starts the incident report.

F

Engineering term

Fusion contract

An engineering term from earlier versions of this site for the versioned specifications that fix how one part of the elan1 platform hands work to another, so each part can change without breaking the others. They will be documented in the developer documentation at launch, alongside the elan1 Run capability contracts.

Back to A to Z, from F

G

Gate

Also: Rung

A check an action must pass before it commits. elan1 Work (agentic enterprise automation for ERP and CRM) runs 8 on your books, such as whether the trial balance still balances and the GST return still generates. Every action climbs the same checks in the same order; each step is a rung, numbered 0b to 9 in the engineering specification.

Governed action

An action proposed by an agent, decided by your rules or a named person, carried out only as decided, and recorded. Ask, check, approve, act, record. Every time.

Governed agentic platform

The category elan1 proposes to define and lead: agentic platforms in which every consequential action passes a governance layer that allows, asks a person or refuses, and leaves a record.

H

Hash chain

A sequence of records where each one includes a cryptographic hash of the one before it. Changing, removing or inserting any record breaks every link after it, so tampering shows. When every record is present, signed and linked, the offline verifier answers “Chain intact”.

Back to A to Z, from H

K

Engineering term

KernelPort

The engineering name for the single interface through which a write intent enters elan1 Govern. Every way in, whether the MCP proxy, the REST API or an elan1 application, meets the same port, so there is no side door to your systems of record. Product pages call it the governed path; the interface specification is published with the developer documentation at launch.

Back to A to Z, from K

L

Engineering term

L1, L2, L3

Short names for the three assurance levels: L1 Stated (we say it), L2 Verifiable (you can test it from the record) and L3 Attested (an independent party has examined it). No claim is at L3 today.

Layer model

Also: Control plane

How the elan1 platform fits together: applications on top, Govern in the middle, Run underneath. The applications layer, where work is proposed, holds elan1 Work and elan1 Decide. The governance layer holds elan1 Govern: every action passes through it, and only it can allow one. Earlier pages called it the control plane. The runtime layer holds elan1 Run, which carries, routes, meters and records, and never decides which actions happen.

Back to A to Z, from L

M

MCP proxy

Also: MCP server

A way to route an agent's tool calls through elan1 Govern using the Model Context Protocol, the standard many agents use to reach tools, without changing the agent's code.

Engineering term

Mode A and Mode B

The two deployment modes earlier pages paired with the assurance levels (L1 to L3) to describe how a claim could be checked in a given setup. The modes will be documented for each deployment in the developer documentation at launch. This site now states each claim's assurance level on its own, in the claims register.

Back to A to Z, from M

N

Named approver

The specific person your rules name to answer a request, such as the finance head for a tax mismatch. The record shows who approved and when.

O

Offline verifier

Also: govern verify

An open-source command-line tool that checks an elan1 Govern record on your own machine, with no account and no network call, so an auditor need not trust elan1.

Back to A to Z, from O

P

Engineering term

Pre-gate

Also: G1–G9

One of the nine checks, numbered G1 to G9 in the engineering specification, that a write intent passes before any rule is applied or any person is asked, such as the agent's verified identity and the shape of its parameters. An intent that fails a pre-gate is refused and recorded. The checks and their order will be published with the developer documentation at launch.

Proof record

Also: Action record, 16-field schema

The signed record every governed action leaves, whatever the outcome: 16 fields covering who proposed what, under which rule, the outcome and who decided.

Back to A to Z, from P

R

Refuse

One of the 3 outcomes elan1 Govern gives an action. It does not run, and the attempt is recorded. An error refuses too: any doubt or fault falls toward a human, never toward an action. A refusal costs 40% of one governed action, because it takes the same checks.

Route

Where elan1 Decide sends an answer: act (it goes forward, through elan1 Govern like any other action), review (a person checks it), escalate (an uncertain or high-risk answer goes to a more senior reviewer or a reasoning model) or hold (the case waits, unchanged, until more information arrives or a person picks it up). Confidence and risk choose the route.

Rule

A statement you write, and version, that tells elan1 Govern which actions to allow and which to ask a person about. Anything no rule clearly allows is refused.

S

Engineering term

Seal

Also: Seal v6

What closes every proof record: a signature over all its fields, and the hash of the record before it. Together they make each record tamper-evident and let the offline verifier check it without trusting elan1. The seal format is versioned; earlier pages referred to Seal v6. The current specification will be published with the developer documentation at launch.

Shadow mode

Running alongside your current process, on your own data, without acting. elan1 Decide must match that process before any automation is switched on.

Specimen

A sample proof record with illustrative values, published so you can read the format before you see a real one.

Back to A to Z, from S

T

The answerable enterprise

An enterprise where the routine runs itself, exceptions reach the right person, and nothing happens that no one can explain.

Tool call

An agent's request to use a tool, such as posting a bill in an ERP. Routed through elan1 Govern, each call gets one of 3 answers.

Trial copy

A copy of your system where elan1 Work builds and tests a change first, so the checks run before anything touches live data.

Typed answer

An answer in one of 3 fixed forms from elan1 Decide: yes or no (Boolean), a choice from a list (Choice) or a score (Score), each with a confidence.

U

Undo window

The 30 days in which any elan1 Work change can be reversed. The reversal is checked like any other change.

Unmatched call

A tool call no rule covers. elan1 Govern refuses it. Nothing runs by default.

W

Engineering term

Write intent

An agent's proposed change to a system of record, stated in full before anything is written: the tool, the action, the target record and the exact parameters. elan1 Govern decides on the intent, not on the change after the fact, and the proof record keeps a hash of the parameters so they cannot be swapped later. Product pages call it a proposed action.

Missing a term?

If a word on elan1.ai needs a plain definition and is not here, tell us and we will add it. Write to elango@elan1.ai. Deeper technical terms will be covered in the developer documentation, which opens at launch.

Version history

Versions of this page
Version	Date	Change
v1	2 November 2026	First published at /developers/glossary with 42 terms, including the engineering terms from the old site. Replaces /resources/glossary; old term links redirect to their anchors here.

Intelligence you can answer for.

Start with one process. Add the next when it holds.