Breaking: We can now prove that AI is correct. DeepProve-1 is the first production-ready zkML system to cryptographically verify a full LLM inference. Lagrange has successfully proven the inference of OpenAI’s GPT-2, moving verifiable AI from theory to production: 🧵

Every month, we’ll be publishing detailed updates on Lagrange’s latest progress with research, development, and engineering at the blog below:. 👉

Join millions who have switched to Grok.

Introducing Lagrange's Engineering Updates!. Your go-to resource for the latest in Lagrange's research, development, and engineering 👨‍🔬. July marked a breakthrough for DeepProve, proving OpenAI’s GPT-2 as the first production-ready system to cryptographically prove an entire LLM:.

4/ With DeepProve, desks don’t blindly trust trade signals. They trust them first. This is how verifiability enters the financial stack. Learn more about DeepProve:

3/ With DeepProve, desks can prove every model ran correctly, on real-time data, with zero tampering. → Fast enough to operate in live trading environments (1000x faster than competitors).→ Private enough for sensitive quant strategies.→ Reliable enough for regulatory review.

2/ Here’s how provable trade signals work:. 1. Market data flows into an trading desk AI model.2. A trade recommendation is generated.3. DeepProve verifies the model, data, and output.4. The provable result routes to execution. DeepProve: helping traders make money, faster.

In finance, milliseconds mean millions. AI-driven models are being used to flag trades. But if the model is wrong—or worse, manipulated—the financial cost can be unrecoverable. DeepProve ensures that signals are accurate, authorized, and verifiable.

3/ Proofs of training will be essential for healthcare, finance, and government. Imagine multiple hospitals training a shared cancer detection model—each contributing private data, without ever revealing it. Or banks jointly building a fraud detection engine that regulators can.

2/ Proofs of training are one of four ZK proof types we’re researching to enable:. – Verifiable model training across parties. – Protection of sensitive datasets. – Trust in multi-institution AI collaboration. They prove that a model was trained correctly without revealing the.

The DeepProve roadmap charts a path toward cryptographically safe and transparent AI. One critical building block: enabling privacy-preserving, verifiable training across multiple stakeholders. Let’s dive into proofs of training. 🧵

2/ Learn more about DeepProve-1, our latest release, here 👇.

Would you trust AI with a scalpel? What about a bulldozer? Would you trust it with a $10B portfolio or the universe?. AI doesn’t need trust. It needs verification. AI needs DeepProve.

7/ DeepProve-1 marks an inflection point: Verifiable AI is no longer in development, it's ready for production. It is now possible to generate ZK proofs for modern LLMs, enabling AI systems that can be trusted through cryptographic verification.

6/ Next, we’re targeting LLAMA—one of the most widely used open-source LLMs. We’ll also be optimizing proof size and generation speed so verifiable AI can run at scale for high-throughput use-cases. Stay tuned….

5/ The core features of DeepProve-1 are immediately relevant for high-stakes domains:. → In finance, it can prove model-driven decisions follow fairness mandates.→ In healthcare, it allows audits without revealing patient data.→ In defense, it confirms mission models operate.

4/ This makes AI auditable, scalable, and trustworthy, enabling us to:. → Verify outputs without exposing inputs or model IP.→ Ensure AI systems meet regulatory and policy requirements.→ Confirm correct weights, approved models, and untampered data.

3/ The DeepProve-1 release introduces:. → Support for complex, graph-based model structures.→ Full transformer layer coverage (attention, layer norm, embeddings, softmax).→ GGUF format compatibility for direct use with community-adopted models.→ A dedicated LLM inference.

2/ This milestone extends verifiability from basic model types (MLPs, CNNs) to transformer architectures—the foundation of today’s leading LLMs, including LLAMA, Gemma, and Mistral. Proving GPT-2 brings us significantly closer to supporting widely adopted open-source LLMs.

Learn more about Dynamic zk-SNARKs:.

Previously, proof systems were rigid - one small change meant starting from scratch. The Lagrange research team developed Dynamic SNARKs to fix exactly that. Now, proofs can now evolve with new data, without unnecessary recomputation.