Formats and frameworks
Pieces your tools can read.
runward writes its verdict in open formats. They are supporting pieces for an assessment: compliance belongs to the assessed entity.
The formats
One command, one format, one reader.
These outputs come from the runward command, in your delivery chain. None calls the network.
runward check --strict --jsonThe verdict, the state of each document of the record, the gaps one by one.
Your tools and dashboards: a file to read, line by line.
runward check --strict --attestThe verdict, bound to the digest of the record and of the cited evidence.
The command runward verify, which replays it offline.
runward check --strict --sarifOne annotation per cause of refusal, on the line of the record concerned.
Validated against the OASIS SARIF 2.1.0 schema. You upload it to your forge yourself.
runward compliance eu-ai-actA component definition, in the NIST format: ten requirements, the OWASP categories ASI01 to ASI10, each with a status derived from the check.
compliance-trestle 4.2.0 loads it on every run of runward’s own chain.
runward compliance iso-42001One draft per framework: what runward provides, and what remains with the entity. The command says it itself: “a readiness draft, never a compliance claim”.
An assessor.
runward check --strict --vsa --resource-uri pkg:npm/request-triage@1.0.0A summary of the verdict for a policy engine. It asserts no SLSA level.
Documented recipe: an example Kyverno policy.
runward reportA standalone HTML page: the verdict, for a reader with no terminal and no account.
An assessor, alone.
runward check --freezeThe digest of each cited piece of evidence. A later change makes the check refuse.
The command runward check --strict.
- None of these outputs is signed by runward. Signing is your gesture, with your own key.
- runward also reads your reports as evidence: JUnit, SARIF, coverage, and a software bill of materials in CycloneDX format. It reads them; it does not produce them for your project.
The frameworks
A framework is a label and a draft.
The OSCAL file always keeps the same controls: the ten OWASP categories for agentic applications. The chosen framework changes the draft that comes with it, not those controls.
- iso-42001ISO/IEC 42001:2023Engineering evidence that an AI management programme can use as inputs.
- nist-ai-rmfNIST AI RMF 1.0A voluntary framework. The draft ties the written decisions, the verdict and the rules to its functions.
- eu-ai-actAI Act, Regulation (EU) 2024/1689The draft feeds the technical documentation of Annex IV, in particular its point 2: the design choices and their justification. It names what remains with the provider.
Concrete uses
Nine situations, each with its limit.
What you hold in each case, the command that produces it, and how far it goes.
- 01Change advisory board
A change produced by an agent reaches the board. You have a replayable verdict on the commit, with the gaps by cause.
runward report --format mdLimitIt is not an independent approval control: runward does not read the forge’s approvals.
- 02Architecture review
The review wants the structuring decisions and what would reopen them. You have the journal of written decisions, with their dates.
runward status --jsonLimitThe piece comes from the command.
- 03Internal audit
Months later, the audit asks whether the cited evidence is still the one that was checked. You have a seal, and an attestation that replays offline.
runward verify verdict.intoto.jsonLimitThe seal is not signed: it detects drift, not a rewrite by someone who can change the file. The replay compares the record to the digest; it verifies no signature.
- 04AI Act file
A provider builds the technical documentation of Annex IV. You have a draft that feeds its point 2 from the written decisions.
runward compliance eu-ai-actLimitA readiness draft, not an assessment. Execution logs come from the deployed system.
- 05Change control, in finance
You must show how each change was decided and verified. You have, per change, the decision, evidence that resolves and a replayable verdict.
runward check --strict --jsonLimitPieces to add to your own file. No reading specific to the DORA regulation is delivered.
- 06Supplier assessment
An integrator delivers code produced by agents. You require the commit, the version and the exact command, and you replay the check yourself.
npx runward@0.46.0 check --strict --jsonLimitOn a private repository, replaying needs read access.
- 07Takeover by another team
A team inherits a system. The check requires the recovery runbook and the handover note before it lets anything through.
runward checkLimitIt verifies that these documents exist and are filled. Their quality remains your judgement.
- 08Governance and risk tool
The compliance team wants a machine-readable piece, not a PDF. You have an OSCAL 1.2.3 file.
runward compliance iso-42001 -o preuvesLimitProven tool: compliance-trestle. The import into your own tool remains your verification step.
- 09Deployment admission
Production must accept only what passed the check. You have a verification summary that you sign with your key.
runward check --strict --vsa --resource-uri pkg:npm/request-triage@1.0.0LimitDocumented recipe, with an example Kyverno policy. It is not replayed in runward’s own chain.
Try it on an example, in one command.
Try it on an example