Skip to content

Beta · turned on per workspace by cajeX

Proof, Read From the Systems You Run

A directive says what should be true. Live Evidence reads the repository, the tracker and the wiki to see whether it is, and records what it found for a person to decide on.

The Gap Between the Rule and the Running System

cajeX already knows what should be true about an application: the directives in its scope, the design that was selected, the decisions that explain it. Each of those is a claim. Until now, the only way to test a claim was for a person, or an AI review, to read documents.

Documents say what was intended. Whether the default branch is actually protected, whether the exception actually expired, whether the storage is actually encrypted, is written in the systems themselves. Live Evidence reads it there.

Four Rules It Is Built On

Evidence is only worth having if an auditor will accept it. These are the properties that requires.

Checks only read

A check can use only the operations cajeX has classified as read. cajeX asks each platform for read access only, and writes nothing with it.

The same steps, every run

A check's steps are approved once by a person. Every run makes exactly those calls. No model chooses what to look at when evidence is collected.

A person records

cajeX proposes a result with its reasons and the data it read. A fail always waits for a person. A pass records itself only where your team has allowed it.

Evidence expires

Each result has a validity window. When it closes, the check reads Stale, never Passing.

How It Works

Each application has an Evidence tab that walks through these five steps and shows what is waiting for a person.

1

Connect

An admin gives cajeX read access to a platform: the cajeX Evidence app on GitHub, or a service account for Jira and Confluence.

2

Say where it runs

Link each application to the repositories and projects it runs on. cajeX suggests matches from what the connection can see.

3

Approve the plan

cajeX works out which ready-made checks can prove which directives, shows a dry run that records nothing, and waits for one approval.

4

Collect

On demand, on a schedule, or when something changes in the repository.

5

Record

A person records each result, or approved checks record their own passes. A failing control raises one finding, listing every directive it affects.

Four Ways a Directive Can Be Shown to Hold

They are kept separate, because they are not the same strength of evidence and should not be added together.

Decided

An accepted decision record says the application complies.

Reviewed

Documents, the design, or a review of the code show it.

Proven

A check read a system, and its result was recorded.

Attested

A person confirmed it, and said what they looked at.

What It Reads Today

Three platforms, all in beta. Each check is a ready-made template tested every night against known passing and failing states on cajeX's own test accounts.

GitHub

Beta

Connected with

The cajeX Evidence app, or a fine-grained token where the app cannot go

Examples of what a check shows

Changes to the default branch need an approving review. Merges wait for required status checks. Actions are pinned to a commit. Critical alerts are fixed within their deadline. The latest release ships with a software bill of materials. Storage declared in Terraform is encrypted at rest.

Jira

Beta

Connected with

A service account with OAuth 2.0 client credentials

Examples of what a check shows

Exceptions have not expired, and each is linked to a risk. Corrective actions are closed on time. Changes are approved before they are released.

Confluence

Beta

Connected with

A service account with OAuth 2.0 client credentials

Examples of what a check shows

Decision records carry a status, and the status is current.

Evidence an Auditor Can Check Without Us

  • A record that is only added to. Every result, attestation and correction is a new entry. Nothing is edited in place.
  • Sealed each day. Each day's entries are sealed, and Verify trail names any entry that no longer matches its seal.
  • A package with its own verifier. Export a period's evidence for an application. The package carries a small program and its instructions, so an auditor can confirm it is intact without access to cajeX.
  • Attestations with files. Where only a person can confirm something, they record what they looked at and attach it. A second person can be required to confirm, and it cannot be the same person.

Beyond Settings

The same approach, pointed at the code, the pull request, the drawing and the ticket.

Was it built as decided?

Checks that read the code: the team's own enforcement, manifests, infrastructure as code, the folder structure against the design, and dependency rules. Then a judged review that cites the files it relied on. The results are named plainly: decided but not built, designed but drifted.

On a pull request

cajeX can post a check run on a pull request saying which approved rules the change would break, using a separate app that can write a check run and nothing else.

On the diagram

Pin where an application runs to a shape in its architecture diagram, and the shape shows the state of its checks. The markers sit on their own layer; hide it for the clean drawing.

A ticket for a finding

Raise a Jira issue from a finding, linked back to it. The write uses its own credential, kept apart from the one that reads.

Frequently Asked Questions

Is Live Evidence available now?
It is in beta. It is off for every workspace until cajeX turns it on for that workspace, and it needs Applications to be on. The GitHub, Jira and Confluence connectors and their check templates are beta too. If you want to try it, ask us.
Can a check change anything in our systems?
No. A check can use only operations classified as read, and cajeX asks each platform for read access only. The two things cajeX can write, a check run on a pull request and a Jira issue for a finding, are fixed actions with their own separate credentials, and each is switched on deliberately.
Does an AI agent decide what to inspect?
Not when evidence is collected. A model can help draft a check, and a person approves it. After that, every run replays exactly the approved steps. Where a result needs judgement, a model reads only the data the check quoted, with no tools, and its proposal still waits for a person unless your team has allowed approved checks to record their own passes.
What does an auditor receive?
An evidence package for a period: the results, what was read, and the files, with a record that is sealed each day. The package includes its own verifier, so an auditor can check that nothing was altered without access to cajeX.
How long is evidence kept?
The full outputs and files are kept 30 days on Free, 90 days on Basic, 180 days on Pro, 400 days on Team and 3 years on Enterprise. The record of what was checked and its hashes outlives that. A legal hold stops deletion for the evidence it covers.
Where are our credentials stored?
In a vault that belongs to your workspace, encrypted under a key your workspace holds. Deleting the workspace's key makes the stored credentials unreadable.
Which systems can it read?
GitHub, Jira and Confluence today, all in beta. More are planned; we would rather say which when they have passed the same tests these three did.

Want to Try It on One Application?

Live Evidence is in beta with design partners. Tell us what you run on and we will show you what it can prove.