ProofKit catches fake-success in AI agents by requiring live evidence before any task is marked done.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install proofkit
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install proofkit using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is ProofKit?
ProofKit is the adversarial verification layer for autonomous agents: it treats every success claim as untrusted until a live artifact proves the work actually happened. It pairs a static fake-success scanner with a verify-real workflow that looks for concrete side effects like a non-zero message_id, a newly created file, a real HTTP 200, or a database row that changed.
In Openclaw Skills, ProofKit helps you stop dry-run theater, stub code, swallowed errors, and fabricated metrics before they reach CI or production. The result is a practical shift from trusting logs and return values to shipping only evidence-backed outcomes.
ProofKit Use Cases
- Autonomous posting, trading, or outreach agents where you must prove the post, order, or email really happened
- Multi-agent fleets where one agent's fake success would corrupt downstream inputs
- CI for AI-generated code to block stubs, demo data, and fabricated results from merging
- Any situation where an agent says it worked and you cannot afford to trust that claim blindly
How ProofKit Works
- Run the static fake-success scanner against one or more Python files with
python3 proofkit/verify_real_scan.py <file> [<file> ...]. - Let the scanner flag suspicious patterns such as hardcoded success returns,
message_id: 0,random-based simulators, emptypassbodies, swallowed exceptions, and dry-run defaults. - Move to the verify-real step: demand a live artifact that proves execution, such as real stdout numbers, a file that now exists with bytes, an HTTP 200, or a database row that changed.
- Add a tripwire so the downstream step cannot run if the artifact is missing, which turns false passes into hard failures instead of silent shipping.
- Use the verdict to classify the run as
VERIFIED LIVE,BUILT-BUT-UNVERIFIED, orFAKE-BROKEN, and optionally scale the same logic with/redteam,gatecheck, CI hooks, or Fleet mode.
ProofKit Setup
- Ensure Python 3 is available.
python3 --version
- Copy
verify_real_scan.pyinto your project. For Openclaw Skills, the free tier is just the scanner and the method.
mkdir -p tools
cp verify_real_scan.py tools/
- Run the scan on any agent or helper file you want to validate.
python3 tools/verify_real_scan.py path/to/agent.py path/to/helper.py
- Wire the same command into CI or a release wrapper so merges fail when fake-success patterns appear or when the live artifact is missing.
python3 tools/verify_real_scan.py "$TARGET_FILE"
- If you need the enforcement suite, install the full skill from ClawHub and keep a local model available for the adversarial passes. The toolset uses pure Python stdlib plus your existing local model.
ProofKit Data Schema & Taxonomy
Input and output model
| Object | Shape | Source | Purpose |
|---|---|---|---|
| Scan targets | One or more Python files passed on the CLI | User input | Code to inspect |
| Static findings | Stdout report and warnings | verify_real_scan.py |
Fake-success indicators |
| Live artifacts | Numbers, files, HTTP responses, DB deltas | Runtime checks | Proof that the task really executed |
| Verdicts | VERIFIED LIVE, BUILT-BUT-UNVERIFIED, FAKE-BROKEN |
Verify-real workflow | Final status |
| Enforcement results | PASS/DRIFT tables, majority-refute outcomes, CI failures | Premium suite | Policy gating |
Finding taxonomy
- Hardcoded success:
return True,success=True,message_id: 0 - Fabricated metrics:
random.*used as a stand-in for a real measurement - Incomplete code:
TODO,NotImplementedError, or emptypassbodies - Production contamination: demo, sample, or fixture data in live paths
- Error swallowing:
except: passfollowed by a success report - Dry-run theater: defaults that claim live work without actually running it
Data principles
- Precision-first, with word-boundary matching so unrelated names such as
exist_ok=Truedo not trip the scanner. - Evidence-first, because a clean static scan is necessary but not sufficient.
- No persistent database required; the workflow is file input plus artifact proof plus human-readable verdicts.
ProofKit Advanced Features
/redteamruns N independent skeptics per claim, asks each to refute, and kills on majority so the first optimistic pass does not win by defaultgatecheckgenerates 6+ guard-invariant proofs, such as kill-switch reachability, blocked trading, injection-defense integrity, and fail-closed quality gates, surfaced as a PASS/DRIFT table- Tripwire harness generation wraps a build step so a false pass cannot continue unless the real artifact exists
- CI hook support fails the pipeline on any fake-success flag or missing live artifact, making proof-of-work required to merge
- Fleet mode scales the suite across N agents and ranks which agents are most likely shipping fake success
- The suite can follow the data one layer down to expose wrappers where the real API is hidden behind a fake interface
SKILL.md
Loading
Related Openclaw Skills
Aomi Build
Aomi Build turns OpenAPI specs, SDK docs, and API references into production-ready Rust Aomi crates with typed, agent-callable tools for Openclaw Skills workflows.
Aomi Transact
Aomi Transact turns plain-English crypto and DeFi prompts into fork-simulated, wallet-signed on-chain transactions from the terminal.
CI Pipeline Debugger
CI Pipeline Debugger gives Openclaw Skills users a fast way to parse CI/CD logs, isolate failing steps, and turn pipeline errors into actionable fixes.
Merged Social and Poster Generator Skill
An automation skill that merges the visual design systems of guizang-social-card and magazine-poster into a unified programmatic design asset generator.
AgentPMT Audit Logs
AgentPMT Audit Logs gives you read-only visibility into agent chats, tool calls, workflow runs, and schedules so you can audit what happened and why.
Terraform Drift Detector
Openclaw Skills for Terraform Drift Detector compares Terraform state to live cloud infrastructure in real time and flags drift with actionable reports.
Featured*
