Anthropic's FlashProve tries to prove which AI model answered. Estimators and GCs can ask vendors that question today
Anthropic published an open-source prototype for proving an AI response came from the model a vendor claims. It's not usable yet, but the problem it targets, silent model swaps, is one a contractor can address in a contract.
A contractor has no way to prove which AI model produced a takeoff summary or a bid-leveling sheet. Anthropic published a prototype this week that aims to fix that, but it's years from a jobsite tool. The useful move now is contractual: ask your AI vendors to name the model, pin the version, and give notice before it changes.
What did Anthropic release?
On Nov. 2, Anthropic published FlashProve, an open-source research prototype for "provable inference." It uses a zero-knowledge proof so a provider can show that a response came from a specific committed model, without revealing the model's weights. The stated goal is to let users and auditors check that an answer came from the claimed model "rather than a cheaper substitute."
TechCrunch reports the researchers describe the post as intermediate progress, and there are no plans to deploy it at Anthropic or any other lab yet.
How far is it from being usable?
Far. Per Anthropic's own write-up:
| Property | Reported figure |
|---|---|
| Proof generation speed | About 110 tokens per second on one H100 |
| Proof size | About 19 MB per 1,024-token request |
| GPU cost vs. normal serving | Roughly 70–95 times the GPU time |
| Context limit | 1,024 tokens |
| Output type | Greedy decoding only, not sampled output |
A 1,024-token limit is a few pages of text. A spec section or a bid package is many times that. So this is a proof of concept, not something a sub can switch on.
Why should an estimator or GC care?
The problem it targets is real for anyone buying AI software. If a vendor's estimating, takeoff-assist or RFI tool sits on top of a hosted model, the vendor can change which model runs underneath. Cheaper models can be fine. They can also quietly change how the tool reads a scope exclusion or an alternate.
Here's what that looks like on a job:
- An estimator builds trust in a bid-leveling assistant over three projects, then the vendor changes the underlying model and the output style shifts without notice.
- A GC's PM relies on AI-drafted RFI responses, and nobody can say which model version produced the one that went to the architect.
- A sub's accuracy check from the pilot no longer describes the product it's paying for.
None of these needs cryptography to manage. It needs paper.
What should you ask vendors now?
- Which model and version runs our workflows? Get it in writing, not on a marketing page.
- Can we pin a version? Or at least get advance notice before it changes. Set the notice window in the agreement.
- Will you re-run our acceptance set after a change? Keep 10 to 20 past bids or RFIs with known right answers and ask for the results.
- What happens if quality drops? Termination or credit rights, stated plainly.
What's still unknown?
Whether labs and vendors will ever publish the model "fingerprints" a system like this needs. TechCrunch notes that it only works if they cooperate. Even then, a proof says which model answered, not whether the answer was right. An estimator still has to check the quantities.
Takeaway
Build a small test set from past work this week, and add model-name and change-notice language to your next AI vendor renewal. That gets you most of the protection FlashProve is chasing, with none of the 70x GPU bill.
Forward this to the person who signs your software renewals. Construction AI Brief publishes three times a week at constructionaibrief.com.
- Can a contractor verify which AI model an estimating tool is really using?
- Not technically, not yet. Anthropic's FlashProve is a research prototype for exactly that, with no deployment planned. Today the practical check is contractual: require the vendor to name the model, notify you of changes, and let you re-run a test set.
- What is FlashProve?
- FlashProve is an open-source research prototype from Anthropic that uses a zero-knowledge proof to show a response came from a specific committed model, without revealing the model's weights. Anthropic describes it as a research prototype, and its limits include 1,024-token contexts.
- Why would an AI vendor swap in a different model?
- Cost is the obvious reason. TechCrunch's coverage notes that verification could expose labs that quietly substitute a cheaper model. A construction software vendor reselling model access has the same incentive, which is why model-change notice belongs in the contract.
- What should a GC ask an AI vendor about model changes?
- Ask which model and version runs your workflows, how much notice you get before it changes, whether you can pin a version, and whether the vendor will re-run your acceptance test set after a change.