Your coding agent writes. Verify checks the evidence. You decide what ships.
NovarraAI Verify is an Autonomous Coding Verification Runtime: evidence-gated completion control for autonomous coding. Coding agents cannot certify their own work. Verify evaluates configured evidence in a separate reviewer context before any work is considered done.
Verify evaluates configured evidence in a separate reviewer context. It does not guarantee correctness, security, safety, bug-free code, or production readiness.
- Windows PowerShell 5.1 · PowerShell 7+
- Git required
- Claude Code CLI — only for Execute, only adapter in V1
- DryRun is the default
- Read-only Git wrapper — never stage, commit, or push
The problem
Coding agents cannot certify their own work.
An agent can describe its own output as done, but self-asserted completion is not evidence. Verify separates the author from the verdict.
Same author, same reviewer
When the author reviews its own work in the same context, the separation is weak.
Coverage is invisible
Unchecked paths, unwritten checks, and silent assumptions are not evidence.
Assertion is not proof
DONE must be an outcome of configured evidence and a separate review — not the writer's opinion.
Operating flow
From writer to verification to human decision.
Verify runs the agent inside a scoped workflow and keeps the decision about what ships with you.
01
Task
You define the assignment and the completion criteria it must satisfy.
02
Writer
The agent works inside an allow/deny scope. A writer cannot mark work DONE.
03
Git-observed scope
Repository changes are observed through a read-only Git chokepoint.
04
Deterministic validation
Configured machine checks must pass before completion can be accepted.
05
Fresh reviewer
A fresh reviewer context evaluates the work against the evidence.
06
Human gate
Protected paths require your applicable decision before work proceeds.
07
Reviewed work stays uncommitted
Nothing is staged, committed, or pushed automatically.
08
Human integration
You inspect the result and decide what ships.
Control architecture
Verified capability boundaries
These behaviors are part of the examined release candidate and are scoped to it.
What Verify does
10 verified controls — view details
- Local PowerShell-based orchestration on Windows: Windows PowerShell 5.1 or PowerShell 7+.
- Git is required for the repository being verified.
- Claude Code CLI is the only V1 execution adapter and is needed only for Execute mode.
- Dry run is the default mode and does not launch a worker or write runtime state.
- Persistent recovery keeps verification state across interruptions.
- Allow/deny scope controls define which paths and commands are permitted.
- Deterministic commands are checked before the agent may run them.
- A separate, fresh reviewer context reviews the writer's work before it is marked done; this is process separation, not third-party certification.
- Protected-path human gates require an applicable human decision before work proceeds.
- A read-only Git wrapper blocks stage, commit, push, merge, rebase, reset, clean, checkout, stash, tag, fetch, and pull.
- A writer cannot directly mark work DONE; DONE is a review outcome, not proof of correctness.
- Local PowerShell-based orchestration on Windows: Windows PowerShell 5.1 or PowerShell 7+.
- Git is required for the repository being verified.
- Claude Code CLI is the only V1 execution adapter and is needed only for Execute mode.
- Dry run is the default mode and does not launch a worker or write runtime state.
- Persistent recovery keeps verification state across interruptions.
- Allow/deny scope controls define which paths and commands are permitted.
- Deterministic commands are checked before the agent may run them.
- A separate, fresh reviewer context reviews the writer's work before it is marked done; this is process separation, not third-party certification.
- Protected-path human gates require an applicable human decision before work proceeds.
- A read-only Git wrapper blocks stage, commit, push, merge, rebase, reset, clean, checkout, stash, tag, fetch, and pull.
- A writer cannot directly mark work DONE; DONE is a review outcome, not proof of correctness.
V1 boundaries
4 material limitations — read before use
- V1 supports no Codex, OpenCode, Cursor, Linux, or macOS adapter or host.
- Claude Code CLI is required only for Execute mode; dry run needs no execution adapter.
- Verify cannot automatically stage, commit, push, merge, rebase, reset, clean, checkout, stash, tag, fetch, pull, install dependencies, run migrations, or deploy.
- DONE is a review outcome. It is not proof that code is correct, secure, safe, or production-ready.
- V1 supports no Codex, OpenCode, Cursor, Linux, or macOS adapter or host.
- Claude Code CLI is required only for Execute mode; dry run needs no execution adapter.
- Verify cannot automatically stage, commit, push, merge, rebase, reset, clean, checkout, stash, tag, fetch, pull, install dependencies, run migrations, or deploy.
- DONE is a review outcome. It is not proof that code is correct, secure, safe, or production-ready.
Product demonstration
A truthful look at the engine.
Selectable text from a real DryRun run of the examined release candidate — not a simulation.
[PASS] 01-sandbox-first-run :: init=0 dryrun=0 state-written=false[PASS] 02-web-node-app :: init=0 dryrun=0 state-written=false[PASS] 03-python-service :: init=0 dryrun=0 state-written=false[PASS] 04-hardened :: init=0 dryrun=0 state-written=falseEXAMPLE DRY RUNS: PASS
Sanitized output captured from real DryRuns of the examined release candidate. Text remains selectable. See verification method and limits.
Evidence
Tied to this exact release candidate.
Every number here names the release, the check, the environment, and what it does not prove.
Unit suite — 319 / 319
An independent local Windows PowerShell 5.1 run passed all 319 unit cases for Supervisor V1 v1.0.0 with exit 0. This does not prove commercial readiness or broad correctness.
Example DryRuns — 4 / 4
Four example DryRuns passed in an isolated local audit with no state.json writes. DryRuns do not demonstrate Execute-mode behavior.
Human authority
You are the only integrator.
Reviewed work stays uncommitted. Verify never stages, commits, merges, pushes, installs, migrates, or deploys.
Writers cannot mark DONE
Only the reviewer stage can reach DONE, and DONE is a review outcome, not proof of correctness.
Protected paths gate the flow
Protected-path human gates require your applicable decision before work proceeds.
Nothing ships itself
You inspect the resulting diff and decide what ships.
Compatibility
Transparent system requirements.
V1 is a Windows-hosted, PowerShell-based engine. Confirm the fit before you plan around it.
Required by V1
- Windows orchestration host
- Windows PowerShell 5.1 or PowerShell 7+ — The host PowerShell runtime used for orchestration.
- Git — Required for the repository being verified. The read-only Git wrapper never stages, commits, or pushes.
- Claude Code CLI — The only V1 execution adapter. Needed only for Execute mode.
- No package-manager install — No additional PowerShell module or package-manager installation; Git remains required, and Claude Code CLI is required for Execute mode.
Explicitly not supported in V1
- Linux or macOS host
- Codex, OpenCode, or Cursor adapter
- Automatic integration, deployment, or a hosted service
- A correctness or safety guarantee
Who it fits
Built for technical teams already running long autonomous work.
Prime candidates use Claude Code heavily and want evidence-gated completion instead of trust-based output.
Good fit
- Technical founders
- Senior developers
- Software agencies and AI development studios
- Fractional CTOs
- 2–20 person engineering teams already using Claude Code heavily
V1 is not yet a fit for
- Teams needing Linux or macOS orchestration
- Teams needing another coding-agent adapter
- Teams wanting automatic integration, deployment, or a hosted SaaS service
- Teams needing a correctness guarantee
Private preview
Founding Edition for hands-on technical teams
Apply for the first cohort if you run Claude Code on Windows and want evidence-gated completion with the final shipping decision kept human.
For the first 10 genuine external paid customers. Phase 1 is limited to non-EU business and professional customers, subject to law, final terms, and Merchant-of-Record support. No tax or VAT treatment is promised before checkout configuration is live.
Perpetual use of the purchased V1 major version. V2 and later major versions are not included.
Founding Edition launching soon. Checkout remains closed pending final approval.
What paid access depends on
- Final approved licence, refund, warranty, and privacy pack
- Merchant-of-Record eligibility, configuration, and tax handling
- A verified customer package and delivery process
- Merchant of Record acceptance and independent usability evidence
The published criteria can establish fit while these gates are completed. Paid public launch remains NO-GO. The release candidate archive is not offered for download.
FAQ
Straight answers.
What problem does Verify actually solve?
Coding agents can produce a lot of work and describe it as done. Verify runs the agent inside a scoped workflow, checks configured evidence in a separate reviewer context, and keeps applicable human decisions in the path before work is marked DONE.
Is Verify a hosted or SaaS service?
No. Verify is a release candidate that orchestrates entirely on your local Windows machine. It is human-readable PowerShell, requires Git for the repository being verified, and needs Claude Code CLI only for Execute mode.
Will Verify commit, push, or deploy my work?
No. A read-only Git wrapper blocks stage, commit, push, merge, rebase, reset, clean, checkout, stash, tag, fetch, pull, dependency installation, migrations, and deployment. Reviewed work stays uncommitted for a human to inspect and integrate.
What does DONE mean?
DONE is a review outcome, not proof of correctness. A writer cannot mark work done directly; a fresh reviewer evaluates it against the configured criteria, and the final decision about what ships stays with you.
What does the Founding Edition cost?
NovarraAI Verify — Founding Edition is planned at USD 297 one-time for the first 10 genuine external paid customers. It includes perpetual use of the purchased V1 major version; V2 and later major versions are not included. Checkout remains unavailable pending final approval.
What is the product called?
NovarraAI is the commercial brand and NovarraAI Verify is the product name. No trademark registration, professional trademark clearance, or non-infringement assurance is claimed.
What hardware and environment do I need?
A Windows machine with Windows PowerShell 5.1 or PowerShell 7+, Git for the repository being verified, and Claude Code CLI available if you want Execute mode. V1 supports no Linux or macOS host and no Codex, OpenCode, or Cursor adapter.
Can Verify guarantee my code is correct or secure?
No. Verify evaluates configured evidence in a separate reviewer context. It does not guarantee correctness, security, safety, bug-free code, or production readiness — that responsibility remains with the humans who review and integrate the work.
Where is Phase 1 available?
Phase 1 is limited to non-EU business and professional customers, subject to final Merchant-of-Record eligibility and the approved commercial terms. No tax or VAT treatment is promised on this prelaunch site.
Verify checks the evidence. You decide what ships.
Join the launch list or request an authentic product demo by email. Checkout and delivery remain unavailable.
Founding Edition launching soon. Checkout remains closed pending final approval.
This does not prove the product is commercially ready, broadly correct, secure, or production-ready for every repository.