INVAR: Open-Source Tool Gives Every Local LLM Inference a SHA-256 Receipt
revuprender · reddit · 2026-08-21
A developer released INVAR (free, Apache-2.0), an open-source tool that adds verifiable provenance to local LLM inference.
How it works: INVAR wraps llama.cpp and gives each inference a 'worldline' — a SHA-256 certificate over the runtime binary + model weights + prompt + sampling params + output, hash-chained into an append-only log. invar verify re-runs any entry and compares digests; editing one byte causes rejection.
Honest scope (stated upfront):
- The default profile proves deployment-pinned reproducibility (same box + binary + weights + temp=0 + fixed seed → same output), not cross-machine bit-exactness.
- Receipts prove what ran, not that the answer is good — provenance, not a quality judge.
- Root on the box can fabricate history; host compromise is out of scope, mitigated by an off-box Ledger collector. Full threat model is in the repo.
The receipt format is an open spec (Computation Receipts) with published conformance vectors, so anyone can implement verification. Install via a one-line curl script; needs Python 3.10+ and llama.cpp on PATH; exposes an OpenAI-compatible endpoint.
Repo: github.com/anomly-labs/invar
More from coding & agent
- Fixing Inconsistent Layouts: A Multi-Agent Approach to Furniture Verification — andersonbcdefg · 2026-08-21
- User reports OpenAI Codex consuming 80% of weekly quota unusually fast — draginol · 2026-08-21
- Rewriting Infrastructure Primitives for Agents: Software Architecture for 2026 — vasuman · 2026-08-21
- Conductor maker pitches cloud agents as Git worktree parallelism hits scaling limits — charlieholtz · 2026-08-21
- hcloud-mcp: Manage Hetzner Cloud via natural language using MCP — modelcontextprotocol · 2026-08-21
- DaedalMap releases MCP connector for global tsunami data — modelcontextprotocol · 2026-08-21