Verified buyer-facing documentation.

NovarraAI Verify is an Autonomous Coding Verification Runtime. It runs a coding agent inside a scoped workflow, checks deterministic evidence, applies allow/deny scope, requires a fresh reviewer, and keeps human gates in the path before any work is marked DONE.

V1 is a release candidate. It is not yet the final customer package, and this documentation describes verified buyer-facing behavior only.

Windows orchestration host with Windows PowerShell 5.1 or PowerShell 7+.

Git for the repository being verified. The read-only Git wrapper never stages, commits, or pushes.

Claude Code CLI is the only V1 execution adapter and is needed only for Execute mode. DryRun needs no execution adapter.

No additional PowerShell module or package-manager installation is required. Git remains required, and Claude Code CLI is required for Execute mode.

Installation copies a PowerShell source tree into a local directory of your choosing. The final customer package will include exact installation commands, a per-file checksum manifest, and a verified quickstart.

A complete installation bundle, per-file checksum manifest, and signing decision are still being finalized before the customer release.

Until the commercial gates close, the release candidate archive is intentionally not offered for purchase or download.

Run the engine in DryRun first. DryRun is the default mode, requires no execution adapter, and writes no runtime state.

Point the engine at a Git repository, review the generated plan, and inspect the deterministic command checks before enabling Execute.

In Execute mode the agent performs the approved steps; evidence is collected and reviewed before anything can be marked DONE.

Runtime configuration defines the allowed and denied scope, protected paths, and deterministic command policy used for the verification attempt.

The exact customer-facing configuration reference will be published only after it is reconciled to the final customer archive.

Allow/deny scope controls define which paths and commands are permitted, and deterministic command checks gate what the agent may actually run.

DryRun is the default and does not launch a worker or write runtime state. Four example DryRuns passed in an isolated local audit with no state.json writes.

DryRun demonstrates the engine's planning behavior only. It does not demonstrate real Execute-mode behavior or validate correctness.

A writer completes the assigned task against the configured scope. A fresh reviewer then evaluates the work against the same evidence criteria.

A writer has no direct state transition to DONE; only the reviewer stage can reach DONE.

DONE is a review outcome, not proof of correctness.

Task assignment, writer execution, Git-observed scope, deterministic validation, fresh reviewer evaluation, applicable human gate, and human integration follow in order.

Reviewed work remains uncommitted for a human to inspect and integrate.

Verification evaluates configured evidence in a separate reviewer context against deterministic criteria.

Checks are scoped by the allow/deny configuration, and only permitted commands can be executed.

The engine observes repository changes through a read-only Git chokepoint.

It does not expose automatic commit, push, merge, or deploy.

Protected paths require an applicable human gate before work proceeds.

Confirm the PowerShell version satisfies the requirement, that Git is available for the target repository, and that the repository path is inside the configured allow scope.

For Execute mode, confirm the Claude Code CLI is installed and authenticated in the environment.

If a command is rejected, check the deterministic command configuration rather than bypassing the gate.

V1 supports no Codex, OpenCode, Cursor, Linux, or macOS adapter or host.

Verify cannot automatically stage, commit, push, merge, rebase, reset, clean, checkout, stash, tag, fetch, pull, install dependencies, run migrations, or deploy.

Verify does not guarantee correctness, security, safety, bug-free code, or production readiness.

No customer download is available. The final archive, per-file checksums, package-verification evidence, and code-signing disposition are still being finalized.

The site does not claim bounded retries, worktree isolation, broader reviewer write isolation, or post-review integrity until those statements are refreshed against the exact customer package.

Version 1.0.0 is a release candidate. Each release entry links to verification context and names compatibility or configuration changes.

The Founding Edition provides perpetual use of the purchased V1 major version. V2 and later major versions are not included. Final V1 update and support terms will be stated in the approved legal pack before checkout opens.

Product and pre-sales enquiries can be sent to support@novarraai.com. No response-time or resolution SLA is promised.

Security reports can be sent to security@novarraai.com. Do not send credentials, secrets, or production data unless NovarraAI specifically requests them through an agreed secure channel.