Proven
Guides

A tour of Proven, in pictures

What you see when you run the judges' lab on your own machine (make up, see LAB.md). Every picture was taken from a lab started in sample mode: recorded answers, no keys, nothing leaves the machine. Click a picture to open it at full size.

The start page

The lab ends make up by pointing you here. It holds the login, links to the pull requests, the dashboards and these documents, six scenarios you can try yourself, and the code and image security reviews.

The lab's start page: the pull requests, the dashboards and the documentation, then the scenarios to try

A review, as a page

Each demo pull request has its review as a page, set the way GitHub shows Proven's comment: the verdict on top, the findings table, and each finding's evidence folded under it.

A change with defects: blocked, with a test that fails on the change and passes on the base, and the advisory rows the dependency check matched.

The review of a change that applies a discount twice: blocked, with its findings and their evidence

A clean change: a short verdict that says what was checked. No false alarm.

The review of a clean change: a clean verdict and what was checked

The dashboards

Six Grafana dashboards, each opening with one sentence you can read in three seconds.

The overview: every pull request, how its review ended, and how well each kept finding is proved.

The overview dashboard: 5 of 7 pull requests blocked, 7 of 7 reviews finished, no false alarm on the clean changes

One review, step by step: each agent in order, how long it took, and how the two reviewers judged each finding.

The one-review dashboard: the agents of one review in order, and the two reviewers' verdicts

The evidence behind each finding: what the Investigator queried and what the Prover ran.

The evidence dashboard: findings kept and dropped, and the Investigator and Prover runs

Which models answered, and what the guardrail checked.

The models dashboard: calls by agent and reviewer, the guardrail's checks, and what Proven refused to keep

The four Meterless engines and what each did.

The engines dashboard: what H-MEM, Scout Intent, Markovian and the World Model did in each review

The model gateway, read from Nightshift's signed receipts. In a sample lab no receipt is loaded, and the panels say so.

The Nightshift dashboard: calls and receipts, here with none loaded yet

The image

The lab runs Proven's published container image. Its Docker Scout result, software bill of materials and build provenance are on Docker Hub.

The tag the judges pull, for both architectures: no critical, high, medium or low vulnerability, one left unrated.

Docker Hub: the judging tag of saphan/proven for amd64 and arm64, with no critical, high, medium or low vulnerability

On this page