PROOFLOOP GUIDE

Documentation

Understand how ProofLoop investigates software bugs, applies patches, executes tests, and records evidence to support its verification results.

01 · Introduction

What is ProofLoop?

An AI-powered bug repair workflow that connects proposed code changes to actual execution results.

ProofLoop investigates reported Python bugs, generates and applies patches, and executes tests to evaluate whether the proposed change resolves the reported issue.

It's central principle is evidence-backed verification: a proposed fix should not be presented as verified unless the recorded execution evidence supports that result.

Bug investigationPatch generationTest executionEvidence-backed verification

02 · Process

The repair workflow

Each repair moves through a sequence of investigation, modification, and execution stages.

01

Report the bug

Describe the issue and provide the relevant context. You can use the built-in demo fixture or upload a Python repository.

02

Reproduce the failure

ProofLoop executes the baseline tests to establish whether the reported failure can be reproduced.

03

Generate and apply a patch

The repair agent analyzes the issue, proposes a code change, and applies the generated patch to the working repository.

04

Run regression tests

The patched code is tested again. Execution results are recorded as evidence for the final verification decision.

Two ways to start

Use the built-in demo fixture to explore the workflow, or upload a Python repository with a bug description and expected failure information.

Start a repair

03 · Trust

How verification works

Verification is based on observed execution results, not simply on whether a patch was generated.

Verified

The expected baseline failure was reproduced, the source was changed, and the supplied regression tests passed after the patch.

Unverified

The available execution evidence does not establish that the repair satisfies the verification requirements.

Important distinction

A successful patch-generation step is not proof that a bug has been fixed. Likewise, passing a supplied regression suite does not prove that every possible bug in the repository has been eliminated.

04 · Traceability

Understanding execution evidence

Each repair report can contain structured records describing what happened during execution.

Execution details

Commands, executor information, exit codes, timeout status, standard output, and standard error where recorded.

Patch information

Changed files, patch diffs, and source-integrity hashes where available.

Test outcomes

Baseline failure reproduction and post-patch regression results.

Report metadata

Repair summary, verification status, timestamp, and verification scope.
Explore execution evidence

05 · Execution environment

Sandbox execution

Repository repair commands are executed in an isolated Docker environment.

Isolated execution

Docker-based repository execution

The repository repair workflow uses a Docker sandbox with network access disabled, a read-only filesystem configuration, a non-root execution user, and resource and timeout limits. Execution is not silently moved to the host Python environment if Docker is unavailable.

Network disabledRead-only filesystemNon-root userResource limitsExecution timeout

06 · Transparency

Current limitations

Understand what a successful repair report does—and does not—establish.

  • Verification is limited to the tests and execution evidence available to the repair workflow.
  • Passing the supplied regression tests does not guarantee that the repository is free of other bugs.
  • Generated patches should be reviewed before being adopted into a production codebase.
  • Execution evidence may be incomplete when a job fails, times out, or cannot produce a valid report.
  • Repository repair currently targets Python projects.