runwardWho it is forThose who answer for it
Your teams ship with agents. You answer for it.
runward gives you the same truth as them: a verdict you recompute yourself, on your machine, without taking our word for it.
What you receiveWhat the check verifiesRegenerated reportsYour frameworksQuestions
runward-demo-renaming
What the check refuses
2 refusals of this kind: a cited piece of evidence is refused by the check.
Check detail, as written (the same for 2)
Where, refusal by refusal
- First-version measurementsrunward/floor.md, line 22, requirement frontier-deterministic-boundary
- Security assessmentrunward/governance/threat-model.md, line 50, requirement security-prompt-injection
What you receive
Three pieces, which you read without us.
They come out of the runward command, in the team’s delivery chain. None calls the network.
- The checkRefused by the checkthreat-model.mdline 5001
The verdict, line by line
At every version, the check says whether it refuses, and names the document, the line and the reason. Your tools read it as JSON.
- in-toto Statement v1check --strict --attestrunward verifysame verdict02
An attestation you recompute
In the standard in-toto format, unsigned: signing stays yours, with your own key.
runward verifyrecomputes it offline, with the repository and the same runward version, and refuses a changed field. - runward reportHTMLFor a readerwithout a terminal03
A report that reads on its own
A standalone HTML page, drawn from the same data as the check: the verdict, for whoever does not open a terminal.
What the check verifies
The presence, the pointers and the integrity of the evidence.
What it verifies
That the evidence the team requires exists, that it still points to the shipped code and, when the team sealed it, that it has not changed since. The check is not a model: same input, same verdict.
What stays your judgement
The quality of the code stays the judgement of your reviewers and your tools. When a line was endorsed, the check says under which declared name, and it flags those that carry no endorsement: a declared name, not a proof of identity.
Regenerated reports
The reports the agent does not write itself.
A report the agent committed itself proves only what it wrote. So the delivery chain can regenerate the reports before the check.
- 01The example shipped with runward
It cites four reports a tool writes, and the delivery chain shipped with runward regenerates them before the check. The script that regenerates them stays in the repository: review of those files is what protects it.
- 02runward-run
Attached to each release on GitHub since November 2026: the tests and the declared reports run out of the agent’s reach, and the check then reads them.
- 03A proposal by the agent
What it proposes stays refused until someone else has endorsed it.
Your frameworks
Pieces for your review, in open formats.
runward writes its verdict in open formats. They are supporting pieces for an assessment: compliance belongs to the assessed entity.
- A pack per framework
- The
runward compliancecommand assembles, from the team’s record, a reading pack framed by a framework, ISO/IEC 42001, NIST AI RMF or the EU AI Act, and exports it as OSCAL. It is a draft for your review. The evidence for a review - One provenance
- The
runward bundlecommand binds the pieces you give it (the attestation, the seal, the OSCAL export, your software bill of materials) into one in-toto statement, for the assessor. CLI commands - Assessing runward itself
- What runward touches and does not touch, for a regulated organisation that adopts it. Adopting runward in a regulated organisation
Questions
What the one who answers for it asks.
Do we have to take your word for it?
No. The check runs in your delivery chain and calls no network. The code is under the MIT licence, and the published package can be rebuilt with public commands. Verifying a release
Who endorsed what?
The check says what was endorsed, and under which declared name; it also says what was accepted on the team’s judgement alone. A declared name is not a proof of identity: identity is established by your access controls.
Can I show this verdict to an auditor?
As a supporting piece for their review, yes: they recompute it on their side, with the repository and the same runward version. The review itself stays theirs.
Does the check judge the quality of the code?
No: it verifies that the required evidence exists, points to the shipped code and, when the team sealed it, has not changed since. The check does not run your code: it reads files. Your tests run it, in your chain or with runward-run.
Try it on an example, in one command.
Try it on an example