Development ·
FlashProve releases code and proofs for checking model identity
A public research prototype links inference outputs to committed weights, with executable code, a Lean protocol and documented proof-generation measurements.
Why it matters
Providers and evaluators could check that a served computation uses the intended model while keeping weights private.
What to watch
Independent protocol audits, reproduced performance and implementations covering broader decoding and model configurations.
Still uncertain. The fast implementation is not formally guaranteed to match the proved protocol. Integer greedy decoding, high compute overhead and the unmaintained artifact limit practical adoption.
Impact assessment reviewed 2026-11-05.
- Who it helps
- Inference providers, security researchers and model evaluators.
- Availability
- November 2 research release with Apache-2.0 code and Lean proofs. The repository labels version 0.16.0 an unmaintained research artifact.
A check on model identity
The provider first commits to its weights, then supplies a proof for a computation. Anthropic reports roughly 110 proof-generation tokens per second on one H100 for its specified Llama 4 Maverick configuration, at 70–95 times the GPU time of ordinary vLLM inference.
The guarantee is narrower than answer quality
The proof covers integer arithmetic and greedy token selection. It does not establish that an answer is true. The fast executable implementation is not formally guaranteed to match the Lean protocol; sampled or floating-point outputs can differ.
FlashProve checks which committed model produced a reply ↗FlashProve research artifact ↗
What you can do
- Read the linked primary evidence and its access limits before planning a trial. Open reference ↗
Where the evidence stops
No build, proof or benchmark executed. The released setup uses a 1,024-token context and substantial compute; it is not a general-purpose verification service.
Sources read. Sources read November 5, 2026 UTC. No third-party project, model inference or complete product workflow was executed.
Sources & further reading
Web · paper
FlashProve checks which committed model produced a reply ↗
Primary page read November 5. Claims were checked against the page; the workflow was not executed.
Anthropic and research collaborators · page checked 2026-11-05
GitHub · implementation
FlashProve research artifact ↗
Primary page read November 5. Claims were checked against the page; the workflow was not executed.
Anthropic · page checked 2026-11-05
853 stars · observed 2026-11-05 13:19 UTC
More sources reviewed (1)
Web · analysis
FlashProve explanation of compute overhead ↗
Adds an accessible independent explanation while explicitly disclosing AI assistance and lack of reproduction.
Gabriel Anhaia · page checked 2026-11-05
Coverage note: Original YouTube demonstrations and independent cross-platform attention were not established. No end-to-end workflow was executed.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
2026-11-02: A public research prototype links inference outputs to committed weights, with executable code, a Lean protocol and documented proof-generation measurements.
Assessment: Providers and evaluators could check that a served computation uses the intended model while keeping weights private.