idea34-openpassmarket-privacy-preserving/README.md

OpenPassMarket: Privacy-Preserving Federated Compiler Optimization Marketplace (Python prototype)

This repository contains a production-ready, Python-based core prototype for the OpenPassMarket MVP described in the concept. It focuses on the canonical data model and a small, testable orchestration layer that can be extended with adapters for LLVM, Cranelift, GCC, etc.

Key concepts implemented in this prototype:

• LocalProblem: a defined optimization task over a code region with tunable decisions and constraints.
• PerformanceSignal: anonymized metrics captured from evaluation (e.g., runtime, energy).
• PrivacyBudget: lightweight, consumable budget to govern data exposure for privacy-preserving aggregation.
• AuditLog: provenance trail for contracts, adapters, and signals.
• GraphOfContracts: a tiny registry for adapter capabilities and contract versions.
• Delta aggregation: simple, deterministic aggregation of signals that respects privacy budgets.

Public API additions (MVP):

• PassSpec: canonical specification for an optimization pass (pass_id, language, target_IR, transform_type).
• verify_invariants(local_problem, pass_spec, budget): lightweight pre-flight checker ensuring inputs are sane and within privacy budget before federation rounds.
• These APIs are exported from the package namespace for easy use in adapters and orchestration helpers.

How to run tests locally

• Ensure you have Python 3.8+ installed.
• Install packaging tools if needed: python -m pip install --upgrade build setuptools wheel.
• Run tests: bash test.sh.

Project structure

• pyproject.toml: packaging metadata (uses setuptools under PEP 621 style).
• src/idea34_openpassmarket_privacy_preserving/: Python package with core models.
• tests/: unit tests for core functionality.
• AGENTS.md: architecture and contribution guide for future agents.
• README.md: this file.

This is an initial seed that emphasizes correctness and testability. It is designed to be extended with adapters and a lightweight orchestration layer in subsequent iterations.