sum

SUM is an open-source project that provides a trust layer for AI-transformed text. It helps users reshape text, inspect changes against the original source, and share checkable reviews. The project's long-term aim is to become a knowledge distiller with useful levels of detail and perspective over retained sources. The core functionality includes: - Source/rewrite review with receipt-aware export in a browser workbench - CLI and MCP tools for model-based research measurements and offline verification - Meaning-diff analysis that shows which source claims survived a rewrite and which vanished - Cryptographic receipts using Ed25519 signatures over JCS-canonical bytes, verifiable across Python, Node, and browser runtimes Key features: 1. Meaning-loss receipts: Signed, replayable certificates over named meaning-loss proxies, with dependency-light verification (no numpy/scipy/torch/GPU/network required) 2. Render receipts: Attest that a transformation happened (issuer, inputs, slider position, model, time) 3. Certified chains: Bind ordered hop receipts into integer-exact Bonferroni budgets with joint confidence 4. Compliance validators: Support for EU AI Act, GDPR, HIPAA, ISO/IEC 27001, SOC 2, and PCI DSS 5. MCP server: Integration for agent swarms with verification and meaning-diff tools 6. Transform substrate: Registry of transforms with receipts via Ed25519/JCS/detached JWS The project emphasizes transparency about what is proven versus measured. The cryptographic side is mechanically proven with three independent verifier implementations agreeing byte-for-byte. The semantic side (extraction quality, fact preservation) is empirically measured with explicit per-corpus numbers and boundaries. Installation is via pip with optional extras: 'sum-engine[sieve]' for basic functionality, 'sum-engine[verify]' for lightweight verification, 'sum-engine[research,judge]' for meaning-diff analysis, and 'sum-engine[research,receipt-verify,judge]' for minting receipts. The project is Apache-2.0 licensed and aligned with open standards (C2PA, W3C VC 2.0, JOSE/JWS/JWKS).