Technical Glossary · QA & AI Terminology Explained

A program that drives a large language model to carry out a sequence of actions against a real system, rather than merely producing text. In SAPFX, five agents plan, generate, repair, verify and document tests by driving a real SAP system over MCP, always under human review.

Optimizing content so it is understood and cited by AI assistants such as ChatGPT, Claude, Perplexity or Gemini, not just by classic search engines. In practice: structured data, self-contained answers and verifiable facts.

Retraining an existing model on a specific corpus in order to specialize it. CabrIA plans to apply this locally, on the customer's real test executions, so the model speaks the vocabulary of their business without the data ever leaving their infrastructure. The first fine-tuning run remains pending: on 9 October 2026, the training set holds 1,892 admitted examples.

A model trained on large volumes of text, able to understand an instruction in natural language and respond to it. In QA it is used to generate test cases, explain a failure, or turn a specification into an executable scenario.

An open protocol that exposes tools and data sources to an AI model in a standardized way. SAPFX relies on the rf-mcp server to let an agent drive Robot Framework and a real SAP system rather than hallucinating screen states.

The set of techniques that let a machine analyze human language. In QA, NLP is used to match functional requirements written in prose to the test cases that cover them.

Deployment on the customer's own infrastructure rather than in a public cloud. The golden rule at AI Cabra is 100% on-premise or sovereign cloud (France, EU): test data, source code and logs never leave your perimeter.

A technique that combines document retrieval with generation: the model answers from documents retrieved at question time, not from training memory alone. CabrIA uses it in its qa-brain server, in service (4,119 passages indexed on 9 October 2026, dense retrieval with BM25 as a complement): every answer cites its passages, and an answer without a citation is rejected and replaced by the raw passages.

Automatic analysis of images. In software testing it is used to locate a UI element or detect a visual regression when no reliable technical identifier is available in the DOM.

A neural network architecture where only a subset of specialized expert subnetworks is activated for each generated token. In CabrIA, a 35-billion parameter MoE model activates only about 3 billion parameters per token, which preserves broad reasoning capacity with a measured throughput of about 100 tok/s on a local workstation, depending on conditions.

A cognitive architecture separating persistent know-how (conventions, syntax rules, self-healing reflexes), meant for the model weights through LoRA SFT/DPO fine-tuning still to be done, from volatile facts (screen revisions, recent traps, tickets) kept in a local hybrid RAG, the FastMCP qa-brain server, in service: 4,119 passages indexed on 9 October 2026. Every RAG answer cites its passages; an answer without a citation is rejected.

A high-performance asynchronous Python implementation of the Model Context Protocol. It serves as the interconnection engine used by CabrIA for its qa-brain memory server.

The practice of using an independent third-party AI model (a 30-billion parameter model in CabrIA's testbench) to grade and validate generated test cases and assertions, so that the main model is not the sole judge of its own output: the judge confirms and traces the scores, without claiming to remove all bias.

A development and inference hardware environment built on two 128 GB unified memory workstations, without a datacenter cluster: local model serving in quantized GGUF format, with a historical throughput of about 100 tok/s depending on conditions, and a first LoRA training run in bf16 planned. It aims at sovereign on-premise test AI with no per-token cloud fees; no throughput is guaranteed.

AI Cabra's formal specification defining the operational boundaries, input and output contracts and compartmentalized responsibilities of the five testing agents of rf-test-agents (rf-planner, rf-generator, rf-healer, rf-verifier, rf-istqb). No agent judges its own work, and four mechanical guards, run in continuous integration on every push and by a Claude Code edit hook, check what the agents write.

The standard MCP server for Robot Framework allowing AI agents to execute keywords and inspect system state in live sessions. It powers the perceive → act → perceive loop, ensuring every test step is proven against the running application prior to code generation.

The discipline of designing systems where AI agents do not just answer but act: it separates reasoning, authorization, execution and verification, so that a result rests on observable facts rather than on the agent's own verdict. A method that uses agents, such as STDD, is not an agentic architecture for that reason. It is the scope of CabraMesh, our V2 work still under research.

The approach of choosing a model for a specific job rather than a single model for everything: its capabilities are qualified on representative cases, then compared for quality, latency and resources, on local infrastructure. Changing the model must not change the business rules. It is a research track of CabraMesh; adding more models is not in itself proof of progress.

A representation of a user interface as exposed by the browser and assistive technologies: roles, names and states. A test can resolve its targets through it with getByRole, rather than relying on fragile CSS or XPath selectors; it is the approach planned for CabraFlow, in design.

The ability of an automated test to recover independently when an element's locator changes. In SAPFX, every repair is logged and never silent: an invisible fix would turn a test into a false witness.

The description of a scenario to verify: initial state, action steps, and expected result. It is the core atomic unit of QA, and what our AI tooling helps generate and maintain automatically.

A view of the screen designed for an AI agent: each actionable target gets a number, with its human label, its ID and its value (Get Screen Map), and the annotated screenshot draws these numbers on the image (Get Annotated Screenshot). The agent names a number, the library acts on a verified ID; reserved for interactive driving, never in a shipped suite.

Continuous Integration and Continuous Delivery: the automated pipeline that builds, tests and deploys on every code commit. Automated tests deliver real value only when embedded into CI/CD.

The share of codebase or business journeys verified by tests. SAPFX measures 95% code coverage in continuous integration across its 3,959 unit tests replayed without SAP (version 0.8.5), with a strict 85% floor enforced in CI; its 462 scenarios passed live on real SAP systems cover business journeys.

Objective, tool-measured tracking of executed code lines during automated testing. In SAPFX, coverage is measured with pytest-cov in continuous integration: 95% at version 0.8.5, with a strict 85% floor enforced in CI. It tells which share of the code the tests execute, not that this code is free of business defects.

A passive collector capturing test execution events (Robot Framework listeners, Playwright reporters, MCP hooks) coupled to a real-time automated scrubber that masks passwords, tokens, and personal data on the fly before archival. On 9 October 2026, CabrIA's collection holds 4,146,190 consolidated records; a record is neither an admitted example nor a validated scenario.

A strict quality policy forbidding the admission of any AI-synthesized test scenario or code into a training dataset unless it has successfully passed dry-run and full real execution against offline fixtures. This systematic validation of generated code remains to be completed before training. Not to be confused with the tests of the CabrIA software itself: 1,070 of 1,071 passed on 9 October 2026 (the last one queries the lab's real services and timed out), and they do not measure model quality.

The International Software Testing Qualifications Board, whose standards and terminology serve as the worldwide benchmark for testing methodology. The sap-istqb and rf-istqb agents structure test plans and test cases along this framework; CabraFlow, in design, plans a documentation export in this format.

An approach where tests are authored with reusable business keywords while technical automation mechanics stay in a separate layer. Functional analysts can read and maintain tests without writing code.

An identifier targeting a specific UI element. In SAPFX, locators reside in a dedicated resources layer rather than inside test scripts, allowing UI updates without modifying test logic.

An open-source browser automation library by Microsoft with automatic waits and accessible role-based locators. SapFioriLibrary builds on it through the Browser library; CabraFlow, in design, plans to build on it too.

A mature open-source, keyword-driven test automation framework where test scenarios read like natural phrases. It is the foundation of SAPFX, ODOOFX and rf-test-agents, and one of the three export targets planned for CabraFlow, in design.

Testing that exercises a complete business process across multiple applications, exactly as a real user would. On SAP, a single process may cross SAP GUI, Fiori and OData APIs seamlessly.

A test that crosses several channels of the same system: it prepares through the API, acts on the screen and verifies through an independent channel, in the same Robot Framework report. Example validated live: the same product price read through OData, the WebGUI and RFC, the three values compared with each other. See nine hybrid tests.

Measuring application behavior under heavy load: latency, scalability thresholds, and architectural bottlenecks. AI Cabra delivers performance testing with AI-assisted diagnostics.

A deterministic script run in continuous integration on every push, and by a Claude Code edit hook, that replaces natural-language instructions given to the AI with mechanical checks (SHA256 test plan provenance, rejection of raw CSS/XPath selectors, ban on Sleep). rf-test-agents has four: a violation fails continuous integration, and check_conventions also blocks the edit itself.

The check that rejects a truncated table reading: a grid or a table often renders only part of its rows, all of them clean, with nothing to signal what is missing. SAPFX checks the number of rows read against the total the table declares (Table Extract Should Be Complete), on SAP GUI, the WebGUI, SAPUI5 and RFC.

The transparent log (recorded in docs/heal-journal.md) documenting every automated locator adjustment executed by rf-healer: timestamp, deprecated locator, validated replacement locator, and technical rationale. It guarantees zero silent healing and provides complete auditability.

Checking, before any read or write, that the test is talking to the intended system, with a criterion that really tells two systems apart: the release, the kernel, the client served. Two systems can announce the same identifier and the same host name; System Identity Should Be then rejects the target instead of reading the wrong system.

Knowing which tests to replay when a transport lands on an SAP system. SAPFX reads the transport over RFC, read only, follows each object through one level of use, crosses the names with the Robot Framework suites by what each test really executes, then a human decides, suite by suite, to replay or to skip with a reason; the decision is recorded and bound to the impact's fingerprint. Published in version 0.8.5 of the libraries (13 October 2026). See the section.

A rule that decides, before any write gesture, whether SAPFX may write to the SAP system: read only, explore (one approval per gesture) or full, with an audit trail of the decisions. In human mode, the approval comes from a person, through the workbench MCP tool sapfx_approve_write or an input box where they type a confirmation word, and covers one target only; an agent cannot approve itself. An operational guardrail, not a security control: the barrier remains the SAP authorization of the test user. Published in version 0.8.5 of the libraries; the MCP tool belongs to the workbench, not to the PyPI package. See the section.

A delivery decision computed on the results file of a Robot Framework run: tests executed, failures, skips without a reason, error screenshots in a green test, healed locators, plans changed since generation. Three profiles (strict, standard, ramp-up), explicit criteria, and a criterion that cannot be read counted as not measured, never as zero. Published in version 0.8.5 of the package. See the section.

The mindset shared by SAPFX, ODOOFX and rf-test-agents. The specification is a broad vision in human language, structured or not (business goal, user story, BDD scenario, existing test case); the planner confronts it with the real system and writes its observed version under specs/, reviewed by a human, which becomes the source of truth; the test suite is derived from it (every step executed live before it is written, the plan's SHA256 checked by a guard); and what a test needs that the library cannot do yet becomes a library keyword, never a workaround in a suite. BDD is one form of specification, not a competitor; what sets STDD apart is the observed version that counts and also drives the automation code. The same logic as TDD, one brick earlier: the spec, green by default in theory, is confronted with the real system before any test is written. STDD is a method that uses agents, not an agentic architecture. rf-test-agents V1.0.0 is the open-source tooling under Apache 2.0 that applies it. CabraMesh names our V2 research in agentic engineering: bounded roles, task-appropriate local models, governed actions and verifiable evidence. CabraMesh remains under research; no deployment or measured gain is claimed. Explore CabraMesh: V2 agentic engineering.

See also: the STDD film (8 min 34), which shows it on a real SAP case.

SAP's proprietary programming language powering standard transactions and custom extensions. SAPFX live scenarios run against real ABAP systems (ABAP Platform 1909 and 2023), not against a simulation.

The two primary ERP generations from SAP: legacy ECC and modern S/4HANA. SapEccLibrary drives both via SAP GUI Scripting, protecting test assets during multi-year transformation phases.

A logical partition of an SAP system: one system hosts several clients (000 for administration, a working client such as 001), each with its own business data, users and configuration. A test that does not check its client can read the right tables in the wrong place: SAPFX reads back the client actually served, channel by channel, before any write.

The short code that opens an SAP GUI screen: SE16 to display a table, SM36 to schedule a job, BP for business partners. Run Transaction in SapEccLibrary opens it and judges the result on the type of the status message, never on its translated text, so it works whatever the logon language.

The programmatic automation API built into the desktop SAP GUI client. It is the mechanism leveraged by SapEccLibrary, enabled on both server and client machines.

A dynpro is the classic screen of an SAP GUI transaction, where every field and button carries a technical ID; a table control is the input table it contains, such as the field list of a table in SE11. SapEccLibrary reads a table control by column title (Read Table Control), window by window, and returns only the rows actually filled.

The standard transaction that displays the content of an SAP table. SAPFX uses it to count and read tables through the screen, for instance to check the number of records seen in SAP GUI against the number returned by an OData service or the RFC channel on the same system. A structure, which is not a table, is rejected there with a type E message.

SAP's standard list grid (ALV, ABAP List Viewer), found in most reports. It loads its rows only as you scroll and renders the others as empty cells: a reading can have the right number of rows and still be hollow. Read Full Grid scrolls the grid to read everything and Count Blank Grid Rows rejects a hollow reading; columns are read by technical ID, independent of the language.

The message shown at the bottom of an SAP screen after an action, which carries a type (S success, E error, W warning, A abort, I information), a class and a number. SAPFX judges a result on the type and identifier of the message (for example MO/E/402), never on its text, which is translated into the user's language.

The browser version of SAP GUI, served by the ABAP server: the same transactions, rendered in HTML. SapFioriLibrary drives it by the SID of each element, the same identifier as desktop scripting, and reads its grids by checking the rows received against the total the grid declares: the WebGUI sends only one page of rows, renumbered from 1.

The modern web design system and UX framework for SAP, replacing classic desktop screens. Fiori applications are technically SAPUI5 implementations.

The JavaScript UI framework powering SAP Fiori applications. Element IDs change dynamically between runs: SapFioriLibrary is proven compatible from UI5 1.71 to 2.0.

The entry point to SAP Fiori apps: tiles grouped by catalog, some of which display a counter. SapFioriLibrary opens an app by its semantic intent (Open Fiori App), the stable route that survives theme and language changes, and waits until a tile counter is actually loaded before reading it (Wait For Flp Tile Counter).

SAP Fiori app templates generated from the annotations of an OData service: a list (List Report) and a detail page (Object Page). Their control IDs are stable, which makes them good test anchors; a table there renders only its instantiated rows, hence the check against the declared total before any extraction.

A standardized REST API protocol used by SAP to expose backend data to Fiori apps and external systems. SapApiLibrary natively handles OData v2 and v4.

The ABAP server component that publishes OData services. SapApiLibrary reads its service catalog and the $metadata contract with its labels, handles the CSRF token for writes and groups several requests into a single $batch. Its preflight check names the cause of an inactive service instead of letting a test fail on a silent error.

A data view defined in ABAP Core Data Services, on which SAP Fiori apps and the modern OData services of SAP S/4HANA are built. A SAPFX test can count the rows of a CDS view over RFC and compare that number with the counter a launchpad tile displays, with no rounding and no tolerance.

SAP's current programming model for transactional applications, drafts included. The ABAP server generates on the fly an SAP Fiori Elements preview of any published RAP service: Build Fe Preview Path builds its address, which makes it possible to test the user interface with no app deployed.

SAP's native protocol for remote function calls and BAPIs. In SapApiLibrary, the optional RFC channel goes through pyrfc: table reads (Read Rfc Table), BAPIs judged by message type (Call Bapi, Commit/Rollback Bapi Transaction), background jobs and spool (Wait For Background Job, Read Job Spool), IDocs (Create Outbound Idoc, Wait For Idoc Status), and read-only ABAP security configuration and surface.

A standard SAP business programming interface, called over RFC to read or change an object (a partner, an order). Call Bapi in SapApiLibrary judges its result on the type of the messages in the RETURN table (E, A or X fail the call), never on their text, and Commit Bapi Transaction commits the transaction.

SAP's intermediate document for exchanging data between systems (ALE, EDI), tracked through a sequence of numeric statuses. SapApiLibrary creates an outbound IDoc, waits for its status by code (Wait For Idoc Status; 03: passed to port) and finds the linked inbound IDoc by its tRFC transaction ID, never by timestamp.

A job is a program scheduled in SAP (transaction SM36, monitored in SM37); the spool keeps the list it prints (SP01). SapApiLibrary waits for a job by bounded polling of the job table (Wait For Background Job), with no fixed pause, and reads its spool through two independent routes. A step that prints nothing has no spool, which SAPFX tells apart from an empty spool.

SAP BTP is SAP's cloud platform; SAP Build Work Zone publishes sites and launchpads on it, protected by an identity provider. SapFioriLibrary logs in through the identity provider's form (Log In Via Identity Provider) and enters the app opened in its iframe; SapApiLibrary opens the site's API channel over OAuth2 with a service key.

The legacy testing tool built into SAP Solution Manager 7.2, whose standard maintenance ends on December 31, 2027 (SAP Note 3585220). SAP steers customers toward SAP Cloud ALM, which manages and tracks tests; their automated execution goes through tools that connect to it, and existing CBTA scenarios must be taken over in another tool. SAPFX is an open path, with no license fees, to rewrite them in Robot Framework.

The automated test tool built into the ABAP server (extended Computer Aided Test Tool), which records and replays transactions. It is still available; its scenarios can be rewritten in Robot Framework with SAPFX when a team wants a single toolset for SAP GUI, SAP Fiori and APIs.

SAP Solution Manager 7.2, which carries CBTA, reaches the end of its mainstream maintenance on December 31, 2027 (SAP Note 3585220). SAP points customers to SAP Cloud ALM, which manages and tracks tests; their automated execution goes through tools that connect to it.

The standard SAP transaction that analyzes system workload: frequency, volume and response times of the transactions and programs actually run. It helps choose which transactions to test first, the ones the business really uses.

Our open-source SAP test automation ecosystem for Robot Framework, published under Apache 2.0 and distributed on PyPI as robotframework-sapfx (version 0.8.5 of October 13, 2026). Three libraries cover SAP GUI for Windows, SAP Fiori, SAPUI5 and SAP GUI for HTML, OData and RFC/BAPI, and a single test can cross several channels, with no license fees and no vendor lock-in. Designed for testing, it can also write, delete and cross-check its own writes through a second channel, which opens up, as a direction, the generation of test data sets, the population of databases for performance testing and the loading of configuration or migration data. Explore SAPFX.

See also: the “SAPFX in pictures” page, with screenshots taken live and excerpts from validated suites, and the film “The SAPFX mindset” (11 min 11), which sums up its mindset in seven principles.

The three core libraries of SAPFX, 442 keywords documented at version 0.8.5: SapEccLibrary (195) for SAP GUI for Windows (ECC and S/4HANA), SapFioriLibrary (98) for SAP Fiori, SAPUI5 and SAP GUI for HTML on top of Playwright, and SapApiLibrary (149) for OData v2/v4 and RFC/BAPI APIs, the optional RFC channel going through pyrfc. Keyword reference.

The SAPFX plugins for the rf-mcp MCP server (SAP GUI, SAP Fiori, API) give an AI agent perception of the SAP screen and step-by-step keyword execution; the sapfx-mcp overlay adds session state, screenshots, hot reload and, since 9 October 2026, approval by a human of a write to SAP (sapfx_approve_write), without modifying rf-mcp. They are part of the SAPFX workbench, not of the PyPI package, which carries the libraries. The MCP bridge among the workbench tools.

Two recorders turn a real journey into Robot Framework keywords: a browser extension for SAP Fiori and the WebGUI, and a desktop recorder for SAP GUI. They export a complete suite, a resource and suite pair with no raw ID in the test, a test plan and an HTML report; the recorded flow remains a draft to review. See them at work; their options and limits.

Five specialized agents: sap-planner checks the request against the real system and writes an observed plan, sap-generator runs each step before writing it, sap-healer repairs the resources layer without touching assertions, sap-verifier reviews in read-only mode and sap-istqb documents and reviews the test design following ISTQB. They apply the STDD method, under human control, in the SAPFX workbench. The agents and their framework.

The SAPFX workbench tool that scores a test request, a prompt handed to the AI agents, before the agents work on it: its form and template share are checked without any model call, eight criteria (intent, facts versus assumptions, level of evidence, assertion, target, repeatability, capabilities, posture) are judged by a language model, local by default, that scores zero when it does not answer, and a read-only RFC probe checks that the named transactions, tables and function modules exist on the target system. The score measures how the request is written, not its cost; "playable" does not replace exploring the system. Added on 10 October 2026. Prompt Eval among the workbench tools.

Our open-source QA engineering agent suite for Robot Framework under the Apache 2.0 license, distributed on GitHub (CyrilM29/rf-test-agents, version 1.0.0). It implements the Agent Contract v1 architecture (5 specialized agents), the standard rf-mcp (RobotMCP) protocol, and 4 deterministic mechanical guards, run in continuous integration on every push and by a Claude Code edit hook; 310 unit tests passed in continuous integration on 5 October 2026. Explore rf-test-agents. STDD is a method that uses agents, not an agentic architecture. rf-test-agents V1.0.0 is the open-source tooling under Apache 2.0 that applies it. CabraMesh names our V2 research in agentic engineering: bounded roles, task-appropriate local models, governed actions and verifiable evidence. CabraMesh remains under research; no deployment or measured gain is claimed. Explore CabraMesh: V2 agentic engineering.

The name of our V2 research in agentic engineering: a local target architecture where agents with bounded roles act within an approved scope, with task-qualified models, deterministic controls before any action and verifiable evidence, under human oversight. Under research: no deployment or measured gain is claimed, and there is no public repository. Not to be confused with rf-test-agents, the open-source V1 (version 1.0.0, Apache 2.0). Explore CabraMesh.

Our open-source Odoo test automation vertical for Robot Framework, under the Apache 2.0 license, published on PyPI (the robotframework-odoofx package, version 0.2.0 of 13 October 2026, alpha phase). It targets Odoo Community and Enterprise through three channels: the RPC API (XML-RPC and JSON-2) and the web UI (login, navigation, reading of views, form entry), both working and proven on Odoo 16 and 19, and ORM introspection, with five agents applying Spec-Test-Driven Development. Explore ODOOFX.

An AI Cabra R&D program, with no published version: an open AI model to be specialized in Quality Engineering through local fine-tuning, planned on real test executions, paired with a RAG knowledge base where every answer cites its passages. On 9 October 2026, the qa-brain RAG is in service (4,119 passages indexed), 1,070 of 1,071 automated tests passed and the training set holds 1,892 admitted examples (as of 5 October); the first fine-tuning run remains pending. 100% on-premise or sovereign cloud (France, EU), test data does not leave that perimeter. Explore CabrIA.

CabrIA's QA memory MCP server, in service: a local RAG combining dense retrieval, BM25 as a complement and reranking over keyword documentation, test plans, guides and lessons drawn from real incidents, with 4,119 passages indexed on 9 October 2026. Test agents query it before their decisions; every answer cites its passages, and an answer without a citation is rejected and replaced by the raw passages. Discover CabrIA.

An AI Cabra R&D program, in design, for a keyword-driven test orchestrator on top of Playwright. The goal: you would demonstrate business journeys while narrating out loud, and the test would write itself in business micro-phrases and replay on three targets (native runner, Playwright export, Robot Framework export), locally. On 6 October 2026, six specifications are written (12 August 2026), with no code, prototype or demonstration yet; the voice mode awaits a quantified criterion (over 85% correct steps, under 2 s latency) never measured. Explore CabraFlow.