# Glossary

## Agent

An LLM-driven system that can take actions such as editing files, running commands, reading diagnostics, and iterating on code.

## Boundary

A point where assumptions become externally visible or externally dependent, such as a public function, module export, file I/O operation, network call, database query, environment variable, or user input.

## Capability

An explicit permission or handle required to perform an effectful operation, such as filesystem access, network access, subprocess execution, environment reading, randomness, time access, or secret retrieval.

## Contract

A machine-checkable statement about code behavior. Contracts may include preconditions, postconditions, invariants, examples, schemas, or properties.

## Context economy

The degree to which a language and toolchain minimize the amount of prompt/retrieval context required for an LLM to generate correct code.

## Diagnostic

A compiler, linter, test, security, or runtime report describing a problem. In this project, diagnostics are expected to be structured enough for automated repair loops.

## Effect

An observable interaction beyond pure computation, such as reading a file, writing a file, accessing the network, using time, generating randomness, logging, spawning a process, or mutating external state.

## Escape hatch

A language feature or API that bypasses ordinary safety or checking, such as unchecked casts, unsafe memory access, dynamic evaluation, reflection, shell execution, or broad untyped values.

## First-pass validity

Whether generated code parses, formats, typechecks/builds, resolves dependencies, and has a runnable entrypoint before any repair.

## Hallucinated API

A function, type, module, package, method, flag, or configuration option generated by an LLM that does not exist in the selected language/tool/library version.

## Hidden test

A benchmark test withheld from the model and repair loop until scoring. Hidden tests reduce overfitting to visible examples.

## LLM-optimized language

A language/toolchain pair designed to improve the probability that LLMs generate correct, safe, maintainable code while reducing repair cost and context requirements.

## Machine-readable interface summary

A compact, structured description of a module or package containing exported symbols, types, effects, errors, contracts, examples, and version information.

## Maintainability

The ease with which humans or agents can understand, review, modify, and extend generated code without introducing defects.

## Project graph

The structured set of modules, dependencies, exports, imports, tests, contracts, capabilities, and build settings that make up a project.

## Property test

A test that checks a general invariant over many generated or sampled inputs rather than only fixed examples.

## Repair iteration

One cycle in which the model receives feedback, produces a revision, and the toolchain evaluates that revision.

## Semantic locality

The degree to which the information required to generate correct code is located close to the code being generated or available through compact interfaces.

## Structured concurrency

A concurrency model in which tasks have lexical or hierarchical lifetimes, cancellation is explicit, and spawned work cannot silently outlive its owner.

## Typed error

An error value whose possible variants are declared in the type/interface system and must be handled or propagated explicitly.

## Unknown symbol

A compiler/build failure caused by referring to a name that is not defined or imported in the current project and dependencies.

## Visible/public test

A test or example provided to the model as part of the task prompt or repair loop.