Inference proof planner
What would it take to verify model inference?
A cryptographic proof can tie a checked token to committed model weights and its input. This worksheet estimates the scale of proving selected tokens and shows the other records needed to evaluate a consequential AI action.
Based on Anthropic's November 2, 2026 FlashProve research prototype. This page performs calculations in your browser; it does not accept or verify proof files.
The claim it can support
If a compatible proof passes a trusted verifier against the intended commitment, prompt and token, that check supports an exact computation claim for that token. It does not establish who was authorized to request it, what tools ran, or whether the answer was correct or safe.
Planning worksheet
Size a proof sampling plan
A FlashProve proof checks one token against a committed model. Enter a workload and choose how many responses and tokens you would prove. Figures below use Anthropic's Llama 4 Maverick measurements, not observed costs for your system.
Selecting a response does not prove its entire output. This worksheet counts only the tokens you choose to prove. Selected request counts round up to whole requests.
Estimated footprint
- Selected requests / day
- 100
- Token proofs / day
- 100
- Share of output tokens proved
- 0.1%
- Proof storage / day
- 1.88 GB
- Proof storage retained
- 56.34 GB
- Serial verification / day
- 0.47 hours
Estimate updated: 100 token proofs per day, 1.88 gigabytes of proof storage per day.
Verification time sums 17 seconds per proof on one verifier using 24 CPU threads. It is not response latency or a staffing or cloud bill estimate. GB uses decimal units.
The published prototype proves individual token computations with a context of up to 1,024 tokens, for an integer model using greedy decoding. Eligibility of actual traffic, storage overhead beyond proof files, verification scheduling, and provider support require separate testing.
Proof record design
A proof needs a binding record.
For a meaningful audit, keep the verified computation and the surrounding action context linked by stable identifiers. These are design requirements to test, not capabilities supplied by this page.
- 01
Inference claim
Record the trusted model commitment, verifier version and result, exact input window, token ID, proof digest, and when the check ran.
- 02
Request authority
Bind the request to a workload identity, principal, authorization decision, and policy version. A model proof does not supply these.
- 03
Observed actions
Link independently observed tool calls and their outcomes to the same request. A proved token does not show that every downstream operation was seen.
- 04
Review boundary
State which tokens were proved, which were not, the retention period, and any verifier or provider compatibility limits.
Source and current limits
Anthropic reports 18.78 MB per proof and 17 seconds to verify using 24 CPU threads for its Llama 4 Maverick benchmark. Its fast C++ verifier is not formally verified; the slower Lean verifier carries the stated formal relation. The repository describes the release as a research artifact that is not maintained. These are vendor-reported results, not an independent TKOResearch reproduction.