The result of checking supported command and path references in selected files against one repository commit. Kontrua reports how many checked references hold or broke and what it could not check; this is not a verdict on every sentence. The check is deterministic and uses no model. Verdicts are free and unlimited on connection, default-branch pushes, and pull requests.

One checkable statement in a document: a command or a file path written inside a code span, or in a fenced shell block. Prose is deliberately not scanned. Every claim comes back verified (it resolves against the code at that commit), broken (it resolves to nothing), or unverifiable.

A claim Kontrua could not check mechanically. It is reported and counted separately, never as an error: a Makefile with pattern rules, a command from a runner Kontrua cannot enumerate, a path with no extension. Absence of evidence is not evidence of falsity, so an unverifiable claim is never called broken.

A broken claim a person with write access told Kontrua to stop reporting, from the repository page or with @kontrua ignore in a pull request comment. It is reported as unverifiable from then on, never as verified, keyed on file and text so it follows the claim when the paragraph above it moves. Every dismissal records who made it and can be restored.

A document Kontrua declines to judge, and says so. Planning and dated records - anything under plans/, rfcs/, adr/, roadmap/, proposals/, drafts/, journals/, postmortems/ or retros/, files with a dated name, basenames carrying words like PLAN, ROADMAP, RFC, SUMMARY, REPORT or SESSION, and any document whose header says Status: DRAFT - describe what should exist later or did exist once. Auditing them as current truth is the largest single source of false alarms, so they are listed as out of scope instead of counted.

The metered lane: a run can regenerate enabled managed outputs and propose targeted edits for supported broken references in other files. Proposed replacements must pass verification or remain unchanged with a reason. Review the full diff: managed documents may change beyond one line. Deleting a line is off by default.

The agent-facing artifacts whose claims Kontrua keeps true against the code: AGENTS.md and conventions, plus an optional codemap and architecture facts. It is what AI coding agents read before they plan.

One file in the context layer. Kontrua maintains four: AGENTS.md, the codemap (docs/CODEMAP.md), conventions (docs/CONVENTIONS.md), and architecture facts (docs/architecture.facts.json). Each is configured under the context section of the repo config.

The agent orientation file at the repository root, following the emerging AGENTS.md convention: what the project is, how to build, test, and navigate it. Enabled by default.

A map of the repository written for agents: modules, entry points, and dependencies. Default output: docs/CODEMAP.md. Opt-in - off by default.

Code patterns observed in the repository, each backed by cited examples from the code. Default output: docs/CONVENTIONS.md. Enabled by default.

Machine-readable JSON facts about the codebase, extracted deterministically - no LLM involved. Default output: docs/architecture.facts.json. Opt-in.

How far the context layer trails the code: the number of commits since the context last matched HEAD. Fresh context means agents plan against the repository as it is, not as it was.

Documents intended for people: README, architecture, and changelog. An existing README is verify-only in the managed-output plan, though the separate correction lane can propose targeted edits. Enabled architecture and changelog outputs can be regenerated. An existing API document is audited; API generation is retired.

One correction pass, and the only thing your plan meters. Created when a trigger fires, it regenerates the artifacts Kontrua owns, corrects the false lines in the files you wrote, and tracks progress through pending, running, success, failed and queued states. Verdicts are not runs: they cost nothing and no allowance limits them.

The proof attached to a claim: a path, a symbol, a command or a value that resolves at the commit the claim was written for. A claim without an anchor that resolves is withheld and listed in the pull request instead of written.

The branch Kontrua gates correction runs on. A pull request only starts a run when it targets this branch and is approved. Defaults to the repository's default branch. Verdicts are not restricted to it: a pull request against any branch gets one.

The event that starts a run: pr_approved, scheduled, ci or manual. A push does not start a run; it starts a verdict, which is free.

The internal name for a run record. Stores the commit SHAs, trigger type, status, progress log, and output artifacts.

A single generated file produced by a run - for example the updated AGENTS.md or README.md. Each run can produce one output per enabled artifact.

A run that could not be executed immediately because the monthly quota was exceeded. It is stored and retried automatically when capacity is available.

The CLI configuration file, written by kontrua init. Controls which context artifacts and human documents are enabled, their output paths, and the output mode. The dashboard keeps its own per-repository configuration in Kontrua and does not read this file.