GHOST PROTOCOL

AI says Done. We demand proof.

Inspiration

AI coding agents can confidently announce that a task is complete even when important requirements remain untested. A green status or convincing explanation is not enough when software must preserve financial balances, enforce security rules, and handle adversarial inputs.

Ghost Protocol was built around one question: What if software had to prove its completion claim through executable evidence?

What It Does

GHOST PROTOCOL is an evidence-first software verification workbench that evaluates implementations against explicit requirements and executable tests. Rather than trusting natural-language completion claims, it examines observed execution outcomes and associates them with traceable evidence.

The project brings together:

  • A formal requirements catalog with acceptance criteria linked to test targets.
  • Adversarial testing for boundary conditions and security invariants.
  • Isolated test execution with recorded output and exit codes.
  • A four-state verification model: VERIFIED, VIOLATION, BLOCKED, and UNVERIFIED.
  • A SQLite evidence ledger with SHA-256 integrity verification.
  • A forensic dashboard for inspecting requirements, execution history, findings, and evidence.

The PAYMENT-004 Demonstration

The central demonstration tests an overdraft-prevention requirement.

A sender starts with ₹10,000 and attempts to transfer ₹15,000. The intentionally vulnerable fixture permits the transfer, leaving the sender at −₹5,000 and the recipient at ₹15,000. Ghost Protocol identifies this as a violation.

The corrected payment implementation rejects the transfer, preserving the sender's ₹10,000 balance and the recipient's ₹0 balance. The supplied execution results show the corrected engine as VERIFIED.

This comparison demonstrates why an expected failure in a deliberately vulnerable implementation must be distinguished from a successful verification of the corrected implementation.

How We Built It

The frontend uses React, TypeScript, Vite, and Tailwind CSS. The backend uses Python and FastAPI, with pytest-based verification targets executed through isolated subprocesses. SQLite stores evidence records, while SHA-256 digests support integrity checks.

The workflow connects requirements, executable tests, observed results, and evidence so reviewers can investigate how a verification decision was reached.

How Kiro Was Used

Kiro's agent-assisted workflow was used during the project's security-fix phase. The repository retains Kiro session and settings metadata. The development process focused on inspecting the implementation, addressing payment-safety issues, and checking changes against actual test and execution output.

A key lesson was that a deliberately failing vulnerability demonstration is not the same as a failure of the corrected implementation. Similarly, a timeout or blocked execution must not be reported as a successful verification.

Challenges and What We Learned

A central challenge was distinguishing an intentionally detected defect from an infrastructure or subprocess failure. Another was maintaining a clear chain between formal acceptance criteria, the test that executes them, the observed result, and the recorded evidence.

Ghost Protocol reinforces a simple principle: completion claims should be supported by reproducible results, and insufficient evidence should remain insufficient evidence.

AI says Done. We demand proof.

Built With

Share this project:

Updates

Submission history