Executive summary
DeploySutra AI answers one question before a release: "What exactly changed, how risky is it, and can we get back if it goes wrong?" It does that offline, from files you already have, and it never deploys anything.
Trace the Change. Prove the Risk. Deploy by Human Decision.
The problem
Release reviews rely on people reading YAML diffs at the end of a long day. Risky changes (a Service flipping to a public load balancer, a container turning privileged, a password pasted into an env var, replicas dropping to zero, a rollback plan nobody tested) hide in noise. Tools that "approve" releases create false confidence, and tools that need cluster credentials are hard to adopt.
What it does
- Compares two snapshots of Kubernetes manifests (multi-document YAML or JSON) and lists every change with its source line.
- Applies 19 versioned rules (stable IDs DS-K8S-001 to 019) and three profiles (development, staging, production) that decide what blocks.
- Builds a blast-radius map from the manifests you supplied and marks everything else
UNKNOWN. - Checks rollback evidence and refuses to treat a declaration as proof.
- Summarises an exported OpenTofu/Terraform plan without keeping any state, value or address.
- Writes every result as JSON, Markdown and HTML, sealed with a SHA-256 digest.
- Runs an independent self-audit that measures its own safety claims on your machine.
What makes it different
| Principle | How it shows up |
|---|---|
| Never approves | The gate has two states only: BLOCKED and HUMAN_REVIEW_REQUIRED. There is no READY. The audit proves no input reaches any other state. |
| No credentials, no network | No cluster access, no telemetry, no API key, zero dependencies. |
| Honest about limits | Every report lists what was NOT_RUN or NOT_TESTED and why. |
| Proof over promises | 312 automated tests, a 50-case rule corpus and a self-audit are shipped with recorded results. |
Recorded results (version 0.2.0)
- Test suite: 308 passed, 0 failed, 4 skipped of 312 on win32 10.0.26300, Node v24.15.0 (2026-11-17T17:09:01.008Z).
- Rule corpus: 50 of 50 cases exact (31 positive, 19 negative twins). It is author-labelled, not independent.
- Self-audit: 42 passed, 0 failed, 7 explicitly not run of 49 checks. Overall: PASS_WITH_NOT_RUN.
Who should use it
Platform and DevOps engineers preparing changes, release managers who need a defensible gate, SREs reviewing rollbacks, security reviewers checking exposure and privilege, and CI maintainers who want a fail-closed check with no credentials.
What it is not
Not a deployment tool, not an admission controller, not a vulnerability scanner, not a compliance certification, and not proof of live state. See the assurance document for the full list.
Decision guide
Adopt it as a pre-merge or pre-release gate input. Keep your existing CD system, approvals and runtime policy engines. Treat HUMAN_REVIEW_REQUIRED as "a person must now decide", never as "safe".