Course glossary

These definitions describe classroom usage. They do not claim product capabilities, industry metrics or universal standards. The terms match the preferred English column of the course bilingual glossary.

Term	Definition
agent	A program that observes its environment, chooses actions and uses feedback to continue working toward a given goal within defined boundaries.
coding agent	An agent whose main task is to modify and verify code.
agent loop	A control flow that repeatedly observes, decides, acts and processes feedback.
tool call	The process by which a model requests a program operation through an explicit interface.
tool interface	A calling contract that defines a tool's inputs, outputs, errors and permissions.
stopping condition	A rule that ends an agent loop and returns a result or evidence of failure.
context	The input information available to the current reasoning or generation step.
context window	The amount of context a model can process at one time.
context engineering	The engineering practice of selecting, organizing and maintaining information needed for a task.
context compression	Shortening context while preserving information relevant to the task.
context file	A file that provides constraints, operating instructions and verification entry points for repository tasks.
system prompt	A prompt that sets high-level behavior and constraints in a model interaction.
prompt	The text or structured input submitted to a model to guide the current operation.
token	A discrete text or data unit used to measure model input and output.
pretraining	The stage in which a model learns basic representations and predictive ability before adaptation to downstream tasks.
supervised fine-tuning	Adjusting model behavior using examples with target outputs.
reinforcement learning	Optimizing a policy using feedback on actions.
inference	Computing outputs using a trained model.
sampling	Selecting candidate model outputs according to a specified distribution or rule.
temperature	A parameter that adjusts how concentrated the output sampling distribution is.
subagent	An agent that takes on a bounded subtask for a main task.
plan mode	A working mode that develops and checks an action plan before making changes.
skill	A mechanism that packages instructions, resources and steps for a class of tasks into a reusable unit.
skills	A collection of skills that a task can select and combine.
hooks	Extension points that trigger checks or operations at specified events.
MCP	The name of an interface protocol for connecting model applications to external tools or resources.
worktree	A working directory associated with a version-control repository for independent checkout and modification of files.
permission	An authorization constraint on whether an actor may access a resource or perform an operation.
least privilege	The security principle of granting only the capabilities required for the current task.
sandbox	An execution environment that limits the resources a program can access and the effects it can produce.
safety launcher	An entry program that configures sandbox boundaries and resource limits before execution; retained course tools use it for teacher and CI checks.
gateway	A single entry point that model or tool calls pass through, where access, budgets and records are enforced.
trust boundary	A boundary at which inputs and outputs must be validated between domains with different trust levels or permissions.
deny by default	A policy that refuses any operation that has not been explicitly permitted.
fail closed	Blocking a protected operation when a check fails or evidence is insufficient.
timeout	A termination or error condition triggered when an operation exceeds its allowed duration.
resource budget	A predefined allowance for a task's time, number of calls or computing resources.
deterministic replay	Reproducing outputs for the same inputs using fixed records.
replay route	Running a task with a recorded model reply instead of a live API call, with evidence labelled as replay; used by course CI, graders and offline checks.
mock service	A service that responds according to predefined behavior and substitutes for another service to isolate a test.
user discovery	Using verifiable activities to understand users' problems, tasks and constraints.
jobs to be done	An analytical perspective that understands requirements through the tasks users want to accomplish.
product requirements document	A document that records target users, the problem, scope and acceptance requirements.
executable specification	A specification whose required behavior can be checked by a program.
acceptance criteria	A set of verifiable conditions for deciding whether a deliverable meets agreed requirements.
acceptance test	A test that executes and verifies agreed requirements and acceptance criteria.
specification-driven development	A development method that organizes implementation and verification around an explicit specification.
brownfield engineering	Engineering work that extends or transforms a system within existing code, data and constraints.
legacy code	Code that must be understood and maintained within existing behavior and historical constraints.
characterization test	A test that records a system's observable behavior to protect it during subsequent changes.
refactoring	A change that improves the internal structure of code while preserving its agreed external behavior.
dependency upgrade	Moving an external component to a new version and verifying compatibility.
lockfile	A file that fixes dependency-resolution results to support reproducible installation.
continuous integration	A process that automatically verifies whether integrated changes meet project quality requirements.
pull request	A collaboration record that submits a set of changes for review and merging.
code review	Examining code changes against requirements, correctness and risk.
model audit	A review of a frozen snapshot that the teaching team runs with a model; each finding is a claim the student confirms, refutes or marks cannot tell.
explain gate	A review checkpoint requiring the author to explain why a change works, identify risky inputs and provide evidence of verification they performed themselves.
commit trailer	Text in agreed fields at the end of a commit message that records provenance or assistance.
code ownership	The allocation of responsibility for maintaining and reviewing specific code.
unit test	A test that verifies a relatively small program unit in isolation.
integration test	A test that verifies how multiple components work together.
regression test	A test that checks whether an existing feature or a previously fixed problem has regressed.
property-based testing	A testing method that verifies general properties across sets of generated inputs.
mutation testing	A method that introduces controlled program changes to check whether tests detect faults.
mutant	A program version produced by a mutation operation for tests to detect.
mutation score	The proportion of mutants killed by tests, calculated using denominator rules declared in advance.
equivalent mutant	A mutant whose behavior is equivalent to that of the original program within the semantic scope under discussion.
fuzzing	A testing method that explores program failures using many or varied inputs.
test-driven development	A development method in which tests express behavioral requirements and guide implementation and refactoring.
test harness	A test program that organizes inputs, executes the subject under test and collects results.
test oracle	The basis or mechanism for deciding whether an execution result meets expectations.
static analysis	Examining a target program's structure or semantics without running it.
false positive	A check reports a problem when the reported case is not actually an instance of the target problem.
false negative	A check fails to report an instance of the target problem that is actually present.
prompt injection	An attack in which untrusted content attempts to change a model's intended task or constraints.
lethal trifecta	In this course's risk analysis, the combination of access to private data, exposure to untrusted content and external communication.
supply chain security	Practices for managing risks introduced through dependencies, builds and delivery chains.
sandbox escape	A program breaks through its intended execution boundary and gains the ability to produce additional effects.
attack surface	The set of entry points that an attacker can reach and attempt to exploit.
threat model	An analysis describing assets, adversaries, boundaries and potential failure paths.
inert payload	A payload that only records behavior in the course-designated local mock environment and produces no real external effects.
red team	A testing role that finds and verifies system weaknesses within an authorized scope.
hardening	Changes that reduce risk by restricting permissions, reducing entry points and adding verification.
evaluation	Collecting and interpreting system performance under explicit tasks, environments and decision rules.
evaluation quadruple	In this course, the decomposition of an evaluation into a request, environment, stopping condition and scorer.
benchmark	A set of tasks and evaluation rules used to compare or track performance.
data split	An arrangement that partitions data by purpose into sets such as training, calibration and testing.
data contamination	Evaluation information enters the tested system's training or development process and compromises independence.
cheating agent	An agent that gains scores by exploiting flaws in an evaluation instead of completing the intended task.
trajectory	Evidence recording an agent's inputs, actions, tool results and final state in chronological order.
scorer	A component that converts execution evidence into scores or decisions using fixed rules.
LLM-as-judge	An evaluation approach that uses a language model to assess outputs against given criteria.
judge calibration	Comparing judge outputs with trusted annotations and adjusting or assessing decision rules.
pass@k	A metric for at least one success among several candidates, with the sampling and estimation rules declared.
latency	The time from initiation of a request until completion of the specified response.
observability	The ability to understand a system's internal behavior through logs, metrics and trajectories.
regression gate	A checkpoint requiring specified evaluations to continue meeting their conditions after a change.
software factory	An engineering concept that organizes requirements, generation, verification and delivery into a repeatable production process.
runbook	A document listing diagnosis, operation and verification steps for known operational situations.
incident response	The process of identifying, mitigating and recovering from service problems and recording evidence.
postmortem	A retrospective analysis of an incident's sequence, causes, impact and verification of improvements.
idempotency	The property that repeating an operation, such as a retried request, produces its effect only once.
automated program repair	A process in which a program generates candidate repairs and verifies them against explicit conditions.
disclosure statement	A statement recording sources, AI assistance or the absence of AI use.
work record	The single record for an assignment: the artifact, verification results, a short explanation and one AI-use note.
AI-use note	One short note per work record saying what AI assistance was used, for what, and how its output was checked; later items refer to it by ID.
calibration	Comparing predictions made in advance with measured results to adjust judgment.
reflection	A record explaining changes in learning, mistakes and subsequent improvements based on one's own evidence.
reproducibility	The property that others can obtain comparable results again using fixed inputs and procedures.
confirmed	The verdict that a review point holds, shown by evidence someone else can recheck: an input with its actual result, a diff line or an evidence line.
falsifiable	Stated specifically enough that runnable evidence could overturn it, as with a diagnosis that names a checkable condition.
refuted	The verdict that a review point does not hold, shown by runnable counter-evidence or by a contract or issue showing that the requirement it asserts does not exist.
cannot tell	The verdict when the available inputs cannot decide a review point; it must state what is missing and why. Record files write it as undetermined.
independent review	An actual examination of evidence by a session or person who did not implement the content under review.
secure track	The course assessment track that verifies the student's own learning outcomes through controlled individual assessment.
open track	The course assessment track that evaluates outcomes through artifacts and process evidence while allowing AI collaboration.
equivalent alternative	An arrangement that changes the activity or interaction format while preserving learning outcomes and assessment requirements.
factsheet	A record of external facts with a date, sources and evidence for individual claims.