Capabilities tour

Each capability below starts from a problem, shows what GND does about it, and links to a page where you can run it. For prerequisites and limits, see supported capabilities.

Check what the agent produced

Pass, fail, warn, or cannot decide. An agent answered; was the answer acceptable? gnd check evaluates every constraint and returns one of five bands. Failures name the constraint and the offending facts. Unknowns are reported, never guessed. --format json gives a stable report with acceptance_ready. Example: a shipping request whose route exceeds its deadline fails constraints[1] (dispatch_deadline), and the repaired request passes. → Run a check in 60 seconds

Hard rules and advice. Not every rule should block. Alethic constraints decide acceptance; deontic ones produce warnings and the output can still be used. Example: recommending a book Ada already rated fails unread_only; recommending outside her preferred genre only warns. → BookRec, chapter 2

Thirteen constraint families. Uniqueness, mandatory, value, frequency, subset and equality, exclusion and exclusive-or, ring constraints (thirteen variants), comparison, cardinality, subtype partitions, guarded rules, source anchoring and textual requirements. → Constraint coverage

Batch and in-process. Check a directory of populations into one matrix with gnd check MODEL --populations DIR, or call the same checker from your application through the stable gnd.api. → Python API

Check plans, not only states

Plan mode. An agent proposes reimburse_expense(41) for an expense that is still a draft. gnd check --pre STATE --steps PLAN walks the plan, computes the result, and fails the step whose precondition does not hold (pre_submitted). The revised plan (approve, submit, reimburse) passes. → Check an agent's plan

The frame check. An approval also flipped a bystander expense to rejected. Every fact is individually legal, so no snapshot check can see it. Supplying the claimed --post makes GND compare declared and actual changes and fail the undeclared one. → A side effect nobody declared

Walk requirements. A session contract can require operations (a specific approval must happen) and forbid others (no payment without approval). The feedback reports required-op-missing or forbidden-op-present, without revealing private bindings. → Experiment sessions

Derive facts instead of trusting the agent's

Derived fact types. An agent should not guess which books are eligible; the model can compute it. Structured derivations support joins, stratified negation, recursion (warehouse A reaches C through B), exact integer arithmetic and aggregates (count, sum, min, max, avg). → Data and transitions

Claims checked against the derivation. If the application also claims derived facts, GND compares claims with the computed extension. A missing claim is reported as derived-extension-incomplete and a wrong one as derived-extension-unsound, each naming the tuple. → BookRec, chapter 6

Three-valued results. Opaque values and unknown inputs propagate as unknown, which is reported as band 4 (indeterminate). Unknown never becomes yes or no. → ORM: the executable ontology

Fixed reference data and calendar dates. Catalogs the agent must not rewrite are bound by path and SHA-256 as reference populations. Removing the catalog is an input refusal (exit 3), not an empty catalog. Calendar-day dates and durations support deadlines. → Data and transitions

Run agents in a bounded loop

Sessions. A contract fixes the inputs, the verifier, the acceptance mode (exact bytes, constraint satisfaction with coverage floors, sealed tests, or a transition walk) and the budget. gnd exp run loops attempt, check and closed feedback until one of six recorded dispositions. Example: attempt 1 is rejected at reimburse_expense, and attempt 2 is accepted at band 0. → Drive an agent against the verifier

Small models, any agent. Claude Code, Codex, any declared command or any OpenAI-compatible endpoint can be the agent. Every plan is checked the same way, so a cheaper model is safe to try: a wrong plan is rejected and retried, never trusted. Recorded: a 4B model that never succeeded on its first try was accepted in two of three bounded sessions. → Recorded agent sessions

Answer keys never touch the disk. gnd exp author-expectation hashes an exact expected answer and removes the source, so an agent cannot read the answer from the workspace. Saved external judgments can gate acceptance with confidence floors (gnd judgments check). → Experiment sessions

Author and evolve the ontology

Model by dialogue. gnd-model-from-text runs Halpin's design procedure with you: examples first, then facts, then constraints. Every constraint records its source and your words. → Model a domain by dialogue

Recover, adopt, import. Recover a model from an existing codebase with recorded evidence (gnd recovery), migrate a GSD project's requirements (gnd-adopt), or import an ORM model written in the RelationalAI ORM-skill dialect (gnd import rai-orm). → Evolution and recovery

Change rules with a blast radius. gnd evolve preview lists everything a change touches before it is published; apply publishes with compare-and-swap and rechecks linked projects. Example: the same frozen recommendations pass before and fail after an advisory rule becomes a hard one. → Change the rules safely

Compare versions. gnd diff reports bounded entailment between two models and how each judges given populations. → Python API: diff

Read it back. gnd verbalize population turns any population into sentences for a reviewer or the next agent turn. → Python API: verbalize

Build the application around the model

Code generation with clear ownership. gnd codegen writes the model-derived half of a Python project; your extension code is never overwritten (exit 11 refuses instead). → Generate a project

Conformance and architecture contracts. gnd conformance detects drift between project and model (exit 13), including declared architecture: the layout, entry points and smoke cases your extension promises. → BookRec, chapter 5

Publish only checked output. The generated application calls gnd.api.check before publishing, so the application checks itself at run time. → BookRec, chapter 4

Test the tester

Synthetic populations. gnd synth generates conforming, boundary and violating populations from the model, using 26 mutation operators, with recorded seeds. → Synthetic data

Measured verifier. gnd measure reports false-pass, false-fail and undecided rates per constraint, each with its own denominator. It never blends them into one accuracy score. → Read measurement per constraint

Documents

Extract facts, quote-not-obey. gnd-extract reads a document into a population with a digest-bound sidecar; instructions inside the document are treated as data. gnd verify checks the handoff, then the facts. → Documents in, documents out

Render test documents with distractors. gnd-render-doc writes a document from known facts and plants three kinds of distractors; gnd doccheck verifies the rendering, and gnd fidelity measures the round trip. → Documents in, documents out

Source anchoring. Bind values to spans in the source document and check that each has a witness. → Documents guide

Other engines and independent evidence

Soufflé. Export recursive derivations as standalone Datalog; trace which services a vulnerable dependency reaches. → Soufflé: dependency impact

Lean. Export finite definitions with explicitly unproved obligations. Prove that an approval policy step is redundant, and find a counterexample to a tempting converse. → Lean: policy proofs

PyRel. Export a standalone relational program that runs locally on DuckDB; Snowflake execution is not yet supported by gnd export. → PyRel: approval audits

Witnesses. gnd witness produces a finite derivation certificate and gnd replay re-checks it with a separate kernel. → Exports and witnesses

Review panel. Six reviewer seats across five CLIs with a quorum of three. The panel keeps every dissent and never applies a fix. → Exports and a second opinion

Assurance, honestly labelled

GND's checker kernel was machine-verified with the Nagini verifier in the development repository, and a result can carry a proof-backed assurance label when every applicable obligation is covered by current evidence. That evidence bundle is not shipped in this release, so public builds report assurance as unavailable. GND never upgrades an assurance label silently. Exports and replays are separate evidence and grant no assurance. → Formats and errors

See the release notes for the milestones in which these capabilities were built.